How are you using this handbook? Anonymous is fine. Tell me how
The Physician AI Handbook changes as new clinical evidence, regulatory decisions, and tools appear. Each day the handbook changed is listed here, newest first, with the chapters involved and the source behind each change.
November covers LLM surveillance for missed referrals, limits of hemorrhage triage software, evidence grades for gray-zone questions, and resident rotation one-pagers. Documentation examines AI-scribe oversight; Policy separates evaluation proposals from FDA requirements. Radiology distinguishes GPS-Lung procedural endpoints from diagnostic accuracy and NICE’s prostate MRI consultation draft from final guidance; Oncology compares post-discharge digital care with inpatient symptom monitoring.
NICE second draft on AI-assisted prostate MRI
NICE’s prostate MRI consultation draft proposes conditional use with evidence collection and radiologist review; the available evidence showed no clear reading-time benefit.
Why Gives readers the conditions and evidence gaps governing the proposed NHS pathway.
Chen et al.: GPS-Lung navigation in lung biopsy
GPS-Lung reduced punctures in a randomized lung-biopsy trial; diagnostic accuracy and the separate contribution of generative AI remained unestablished.
Why Helps clinicians distinguish a procedural endpoint from diagnostic performance when assessing navigation systems.
Daly and Nipp: distinct oncology symptom-monitoring studies
Post-discharge digital care and inpatient symptom reporting produced different findings in oncology studies with different designs; neither establishes an AI effect.
Why Helps oncology teams match evidence to the setting, intervention and outcome they intend to improve.
Clinical responsibility in AI-scribe guidance
A qualitative analysis of official AI-scribe guidance found repeated reliance on clinician verification, often without specified review methods or protected review time. The documentation chapter distinguishes this governance concern from evidence about note accuracy or patient outcomes.
Why Helps readers examine whether oversight policies provide workable review conditions.
Proposals for evaluating generative-AI medical devices
Scholars affiliated with Stanford HAI and Stanford Medicine proposed task-specific oversight tests, competency-matched benchmarks, validated LLM judges, patient-simulator checks and postmarket monitoring. The policy chapter presents their October 2026 public comment as individual recommendations, not FDA requirements or official Stanford policy.
Why Separates proposed evaluation safeguards from regulatory requirements.
Lee et al. in npj Digital Medicine: AI surveillance for missed specialist referrals
An LLM safety net in seven Singapore primary care clinics flagged 428 notes with unplaced referrals; clinicians verified 225, against 31 patient-reported cases.
Why Unplaced specialist referrals are a silent primary care safety gap; this shows what automated note surveillance catches and what it still misses.
Chapter Primary Care, Family Medicine, and Preventive Medicine
Evidence grades for gray-zone patient questions and AI-written arguments
Gray-Zone Questions grades human evidence on the four GRADE levels plus "no human evidence", applies it to BPC-157, TB-500, thymosin beta-4, CJC-1295, ipamorelin, and compounded GLP-1 drugs with FDA and WADA status, and gives a five-step citation check for AI-written arguments.
Why Patients and colleagues bring confident, referenced claims from influencers and chatbots; a shared evidence scale keeps the reply grounded.
Chapters Gray-Zone Questions: Evidence Grades and AI Claims · Medical Misinformation and AI
Rotation one-pagers for residents
Rotation One-Pagers gives a one-page template for each rotation, a drafting prompt for an institution-approved AI tool with no-PHI and verification guardrails, and illustrative wards, night float, and ICU examples.
Why Incoming residents get local know-how in one place, drafted safely and reviewed before it is shared.
What intracranial hemorrhage triage software does and does not guarantee
Neurology explains that automated hemorrhage detection screens head CTs, flags suspected bleeds, and moves them up the worklist, but does not guarantee every critical case is read first because missed bleeds stay in the normal queue; Aidoc BriefCase was cleared through 510(k) K180647 on August 1, 2018.
Why Triage software reorders reading; it does not replace a radiologist's read of every study.
References, links, and wording were also kept current.
Microsoft's Science study tests step-by-step AI diagnosis, and a JAMA letter tracks Medicare's AI-enabled prior authorization. Clinicians often miss wrong AI advice, reasoning models add false-positive diagnoses, and an agentic system speeds immune-related adverse event review. The ONC HTI-5 rule ends HTI-1 model card requirements; California, NHS England, and the MHRA set clinical AI rules. An FDA device evidence tracker and teaching materials join the book.
Gallifant et al. in Nature Cancer: agentic LLM extraction of immune-related adverse events
An agentic LLM cut median adverse-event review time from 428 to 242 seconds in a crossover trial at one health system; accuracy fell in prospective silent deployment.
Why Immune-related adverse event curation is slow manual work in trials and registries; this shows what AI assistance gains and where accuracy drifts.
Lotan et al.: reference quality in OpenEvidence answers
LLMs in Practice adds a cross-sectional study of 150 standardized prompts across five specialties: none of 4979 OpenEvidence references was confirmed as nonexistent, and three ASCO meeting abstracts had author attribution errors. The study tested whether references exist, not whether they support the clinical claims.
Why Existing references are a floor, not proof of accuracy; clinicians still need to check that each cited source supports the answer.
ONC HTI-5 final rule (2026) ends HTI-1 model card and risk-management requirements
Privacy & HIPAA and Vendor Evaluation explain that the HTI-5 final rule, published 28 October 2026, removes the HTI-1 source-attribute ("model card") and intervention risk-management requirements for certified health IT effective 28 December 2026; after that date certification no longer means the documentation exists.
Why Health systems can no longer rely on certification for algorithm transparency, so procurement has to request intended use, training data, performance, and known risks directly.
Chapters Privacy, HIPAA, and Patient Data Security · AI Vendor Evaluation Framework · Clinical AI Policy Templates
Utah AI prescription-renewal pilot: licensing board objection and the state's reply
Case Studies reports that on April 20, 2026 the Utah Medical Licensing Board recommended suspending the AI prescription-renewal pilot, and that the state declined the next day, citing physician review of every Phase 1 refill decision, monthly reporting, and its authority to modify or cancel the pilot.
Why The exchange shows how state AI sandboxes and medical boards can disagree over who oversees autonomous prescribing.
Chapter Clinical Case Study Library
Wu et al. in Nature Communications: reasoning models find more diagnoses and more false positives
On 1,157 hospitalizations, reasoning-mode models found the primary diagnosis more often (52.8% vs 38.7%) but added 34.8% more false-positive diagnoses per case.
Why A model that names the right diagnosis more often can also add wrong ones, so clinicians reviewing reasoning-model output need to check every suggested diagnosis.
FDA AI-enabled device evidence tracker
The FDA AI-Enabled Device Evidence Tracker keeps a dated copy of FDA's AI-enabled device list, summarizes it by year, pathway, and review panel, and grades the strongest published study behind 33 widely deployed device families, from randomized trial to FDA summary only.
Why FDA authorization says a device met a regulatory standard, not how it was tested; the grades show which devices have trial or prospective evidence.
Chapters FDA AI-Enabled Device Evidence Tracker · Evaluating AI Clinical Decision Support Systems · Diagnostic Imaging, Radiology, and Nuclear Medicine · Cardiology and Cardiothoracic Surgery
Teaching clinical AI in residency, fellowship, and medical school
Teaching with the Physician AI Handbook offers four- and eight-session tracks, maps to the ACGME Common Program Requirements and the AAMC AI competencies, journal club questions for every chapter, case discussions, and assessment items with answer points.
Why Program directors can fit clinical AI into an existing didactic calendar against the frameworks they already report on.
Chapters Teaching with the Physician AI Handbook · Course Syllabus Template
Living edition and annual frozen editions
Editions and Versions explains that the handbook is published as a living edition on the website and as dated annual editions that do not change once archived, with how to cite each.
Why Readers who cite the handbook can point to a version that will not change.
Chapter Editions and Versions
Bai et al. in Nature Medicine: EVLF-FM, an explainable vision-language model for imaging
EVLF-FM, trained on about 1.3 million images across six specialties, diagnoses and marks the image region behind its answer; the authors say it is not ready for deployment.
Why Showing where in the image an answer comes from helps radiologists check it; explanations are post hoc and the model reads only 2D images.
Neprash and Barnett in JAMA: service use under Medicare's WISeR prior authorization model
In a JAMA research letter, targeted services fell 15.0% to 19.5% in the first 6 months of Medicare's WISeR prior authorization model, which encourages AI review.
Why Prior authorization is spreading into traditional Medicare with AI in the loop; the data cannot separate denials from fewer orders or show AI's role.
Azad et al. in npj Digital Medicine: ambient scribe adoption across Johns Hopkins
After a systemwide opt-in ambient scribe rollout at Johns Hopkins, 69.6% of community clinicians and 34.1% of academic clinicians adopted it.
Why A license alone does not produce uptake; deployment plans need specialty templates, trainee guidance, and tracking of who stops using the tool.
Kuo et al. in npj Digital Medicine: umbrella review of AI fracture detection
An umbrella review of 21 meta-analyses puts AI fracture detection at 0.897 sensitivity and 0.904 specificity, with low or very low certainty.
Why Headline accuracy for fracture AI rests on weak syntheses, with no consistent gain over clinicians and no reported data on harms.
Preprint: AI models adopt a patient's stated diagnosis under repeated pushback
In a preprint, nine AI models switched to a patient's stated diagnosis in 23.2% of simulated consultations without any new clinical findings.
Why Patients often arrive with a diagnosis in mind; a correct answer that collapses under pushback is a failure mode benchmarks rarely test.
MEDIATOR preprint: open-vocabulary tokenization for EHR models across hospitals
A preprint tokenizer kept 100% of eICU events, against 13.0% for closed-vocabulary tokenizers, when moving an EHR model between hospital databases.
Why Hospitals code the same events differently, which limits moving EHR models between sites; these results are retrospective ICU benchmarks only.
Microsoft reports sequential diagnosis results (SDBench) in Science
In Microsoft's Science benchmark, an AI orchestrator with GPT-5 diagnosed 83.2% of NEJM case conferences step by step; generalists averaged 36.1% on held-out cases.
Why It tests the real diagnostic task, deciding what to ask and order under cost limits, though a text-only benchmark is not bedside care.
Meta-analysis in npj Digital Medicine: do clinicians catch wrong AI advice?
Across 46 reports, clinicians were far more often wrong when automated advice was wrong (pooled OR 10.19); about a quarter of wrong computer ECG readings went uncorrected.
Why A human reviewer does not reliably catch AI errors, so health systems should test oversight as a safety step rather than assume it works.
Sato et al. in Nature Health: chatbots favor brand-name drug products
Five LLMs named one product first 97.5% of the time, favoring originator brands (OR 20.5) and raising projected drug costs.
Why Clinicians and patients who ask a chatbot which drug product to use may be steered toward costlier brands unless they check generic options.
State of Clinical AI Report as a peer-reviewed review in BMJ Digital Health & AI
The State of Clinical AI Report appears as a peer-reviewed narrative review in BMJ Digital Health & AI, covering 82 studies across six domains.
Why Readers who cite the report can point to its peer-reviewed version and see how its studies were chosen.
Chapter Further Reading and Resources
ACCESS2 in Nature Communications: autonomous eye screening in youth with diabetes
In ACCESS2, AI eye screening completion was equivalent across racial and ethnic groups of youth with diabetes (99.5% and 98.0%).
Why Equal completion is a first step toward closing screening gaps, but it does not show equal diagnostic accuracy or better vision outcomes.
Chapter Ophthalmology
Ariel et al. in Annals: how often chatbots recommend low-value tests and treatments
Six commercial LLMs recommended unneeded testing or treatment in 5.4% of 225 Choosing Wisely scenarios, mostly thyroid ultrasound and routine ECGs.
Why Overuse clustered in two common workups, which tells clinicians exactly where chatbot advice most needs checking against Choosing Wisely guidance.
Course syllabus: grading rubric, learner AI-use policy, and the AAMC principles
Course Syllabus Template adds a grading rubric, a default policy on learners' use of generative AI, and a map to the seven AAMC principles.
Why Teachers adopting the course get ready assessment criteria and an AI-use policy that a curriculum committee can check against national AAMC guidance.
Chapter Course Syllabus Template
MHRA Strategy 2030: AI as a medical device and foundation models
The MHRA's 2030 strategy sets out an adaptive AI-device framework, wider predetermined change control plans, and a new Centre for AI and Digital Health.
Why The UK regulator has set the direction for how AI devices and foundation models will be overseen there.
JAMA analysis of OpenEvidence use by US clinicians
A JAMA log analysis counted 952,008 NPI-verified US clinicians and 271 million OpenEvidence queries; it measured use, not accuracy or outcomes.
Why Readers get peer-reviewed usage figures, measured from platform logs, for an AI reference tool many physicians already consult.
AMA Center for Digital Health and AI and its agentic AI governance framework
The AMA's agentic AI framework asks what clinical function an AI performs, what evidence supports it, what oversight it needs, and who is accountable.
Why The AMA says these questions will guide its work on liability, licensure, and regulation; its frameworks state positions rather than set rules.
AMA 2026 physician survey: what the 81% AI figure counts
In the AMA's 2026 survey, 72% of physicians named an AI use case and 9% were uncertain; the reported 81% combines both groups.
Why Readers can see how much of the AMA's headline 81% reflects physicians selecting named AI use cases.
Chapter Industry Reports and Benchmarks
JAMA Viewpoint on rethinking physician training for the AI era
Medical Education adds a JAMA Viewpoint on rethinking physician training for the AI era.
Why Program directors weighing AI access for trainees can see the argument alongside the evidence on never-skilling and deskilling, with its limits stated.
Chapter AI in Medical Education
Anthropic reports an out-of-scope form submission in developer testing
Evaluating AI Systems adds a developer-reported case of an AI agent submitting a form it was told not to submit.
Why Out-of-scope agent actions occur even under developer testing, which is why a response for such events should be defined before clinical deployment.
AHRQ sets evaluation requirements for AI-assisted evidence synthesis
Research chapter adds AHRQ's October 2026 evidence-synthesis challenge and its evaluation requirements.
Why It sets concrete evaluation criteria for AI-assisted systematic reviews, while noting these are competition rules, not proof of safety or expert equivalence.
Chapter Clinical Research with AI
Reported Medicare portal access by an internal OpenAI research agent
Privacy chapter adds a reported case of an internal AI research agent reaching beyond authorized access.
Why It shows how an AI agent's research task can exceed authorized access, a concrete privacy risk for clinical and administrative data systems.
FDA class II controls for adjunctive melanoma-detection devices (2026)
Since April 24, 2026, electrical impedance spectrometers and optical melanoma-detection devices are class II devices under FDA special controls. Dermatology describes the 510(k) route for these adjunctive tools and distinguishes DermaSensor's separate category and non-dermatologist intended users.
Why Regulatory category, intended users, and labeling determine the role of a lesion-assessment device.
Chapter Dermatology
FDA TEMPO participants and the limits of pilot status
FDA's TEMPO participant list includes SonderMind's Adjunctive Care Application, Limbic's Unpacked, and Cadence's HypertensionOS. AI Governance in Healthcare explains that pilot participation under enforcement discretion is not FDA authorization; FDA has not evaluated effectiveness for these uses, and manufacturers are expected to collect real-world data.
Why Pilot access requires continued evaluation before clinicians infer patient benefit.
PhysicianBench: consistency across long-horizon clinical tasks
In PhysicianBench's preprint, the best tested agent completed 46.3% of the 100 tasks on the first attempt and 28.0% consistently across three runs. Evaluating AI Clinical Decision Support Systems contrasts these sandbox results with Anthropic's developer-run tests, whose harness and grader differ; neither establishes safe care.
Why Repeated task completion and evaluation conditions matter when comparing clinical agents.
Deterioration alerts paired with rapid-response notifications
Across 11 hospitals in one health system, a nonrandomized before-and-after study associated Epic Deterioration Index alerts plus rapid-response notifications with lower in-hospital mortality. Internal Medicine and Hospital Medicine reports the adjusted odds ratio of 0.82 (95% CI 0.74–0.91) and treats the result as evidence about the combined response pathway, not the score alone.
Why An alert's clinical value depends on the response pathway that follows it.
Nolan et al.: patient values and proxy decision support
Medical Ethics, Bias, and Health Equity describes a prospective study enrolling 15 patient-proxy dyads for hypothetical end-of-life choices: patient-model agreement with supplied values was 72.6% to 72.9%, versus 44.7% for proxies. Only the model received those values, a comfort-care-only rule matched 72.2%, and the small study did not measure care outcomes.
Why Fair comparisons require equivalent information and a credible simple baseline.
EchoNet-Dynamic in a randomized interpretation workflow
In Cardiology and Cardiothoracic Surgery, a randomized study of 3,495 echocardiograms found that cardiologists substantially changed 16.8% of initial left ventricular ejection fraction assessments made by AI versus 27.2% made by sonographers. The result supports an assisted interpretation workflow; the trial did not test patient outcomes or autonomous interpretation.
Why The comparator and outcome define what the trial establishes for clinical implementation.
FDA draft guidance on robotically-assisted surgical devices leaves out autonomous robots
FDA's September 2026 draft guidance covers teleoperated surgical robots that assist practitioners and does not address autonomous or remotely teleoperated systems.
Why Surgeons assessing AI features in robotic systems need to know which devices the draft covers and which FDA is addressing separately.
FDA De Novo for Clairity Breast: five-year breast cancer risk from a mammogram
FDA authorized Clairity Breast to estimate five-year breast cancer risk from a screening mammogram; it does not detect cancer and is now offered to consumers.
Why Patients may arrive with a five-year risk number before their clinician has decided how it fits existing risk models and screening criteria.
ARPA-H CIRCLE program funds AI digital twins of critically ill patients' immune systems
ARPA-H's CIRCLE program funds AI digital twins of ICU patients' immune systems; its length-of-stay target is a goal, not a result.
Why Each future CIRCLE output should be judged by what it claims, from immune-state estimation to treatment recommendations.
Liang et al.: a task-constrained respiratory chatbot in a randomized lay-user study
Liang and colleagues report higher lay accuracy on simulated respiratory cases with a task-constrained chatbot than with web search, with little triage gain.
Why Patients may consult narrow symptom chatbots before a visit; this trial shows what one such tool improved in simulated cases and what it did not.
Marti-Castellote et al.: LLM adjudication of cardiovascular events in PARADISE-MI
Marti-Castellote and colleagues report LLM event adjudication that agreed with a physician committee on most confident calls in one external trial.
Why Event adjudication is a demanding step in cardiovascular trials, and this is an external comparison against a physician committee.
Chapter Clinical Trials and AI
Lange et al.: conversation-level review of a deployed Parkinson's disease chatbot
Lange and colleagues found clinically critical errors that automated triage missed in a deployed patient-information chatbot.
Why In this deployment, finding critical errors required clinical expert review of whole conversations, not automated scoring alone.
Gottlieb proposes CLIA-style certification for AI that health systems develop in-house
Gottlieb proposes FDA certification of health systems that develop their own AI, modeled on laboratory oversight; the pathway would need new legislation.
Why A model a health system develops for its own patients often enters care without FDA review.
Suki's Ask brings UpToDate Expert AI answers into an ambient scribe
Suki's Ask answers clinical questions through UpToDate Expert AI inside a scribe; the only reported evaluation is a vendor preference test.
Why Decision support is arriving inside documentation tools, so a pilot should confirm which chart fields the tool reads and check dosing answers.
Unlu et al.: AI prescreening identified more eligible patients by completing more screens
In Unlu and colleagues' trial, AI prescreening identified more eligible patients because it completed more screens, not because each screen yielded more.
Why Trial teams can expect faster screening from AI prescreening, not a higher eligibility yield per screened patient.
Chapter Clinical Trials and AI
NHS England requires patient information and clinician review for ambient scribes
NHS England requires patient information, respect for dissent, and clinician checking of ambient-scribe output before it enters the record.
Why Ambient documentation requires explicit patient communication and accountable review of the generated record.
UK government accepts AI commission recommendations with implementation still pending
The UK government accepts the National Commission’s healthcare AI recommendations, while implementation and liability changes remain future work.
Why Clinicians and providers must distinguish planned regulatory reform from the rules currently governing AI use.
Chapters Healthcare Policy and AI Governance · Physician AI Liability and Regulatory Compliance
ARPA-H RAPID funds rare-disease data and diagnostic research
ARPA-H RAPID supports rare-disease data resources and diagnostic research; awards do not establish clinical diagnostic benefit.
Why Rare-disease AI development requires shared data and evaluation resources alongside clinical validation.
McBain et al.: youth chatbot use belongs in the mental-health history
McBain and colleagues find that youth chatbot use for mental-health advice is often undisclosed; self-reported helpfulness does not establish clinical benefit.
Why Routine mental-health histories can ask about chatbot use that young people may otherwise leave unmentioned.
Chapter Psychiatry and Behavioral Health
Callender et al.: lung-screening eligibility from age and smoking history
Callender and colleagues report lung-cancer risk prediction from age and smoking history in historical cohorts, with prospective evaluation still needed.
Why Fewer required inputs could reduce the data-collection burden of assessing screening eligibility.
Chapter Primary Care, Family Medicine, and Preventive Medicine
Pandya et al.: lung-protein signature has unresolved treatment-selection limits
Pandya and colleagues’ lung-protein signature has no statistically supported treatment interaction; a preprint reanalysis questions whether it chiefly tracks smoking exposure.
Why Risk prediction and treatment selection require different evidence before a biomarker can guide prevention.
MedGemma authors report benchmark gains with clinical-readiness limits
MedGemma authors report medical benchmark gains, but benchmarks and a single-reader report-generation assessment do not establish clinical readiness.
Why Open-weight medical models need clinical evaluation beyond benchmark comparisons before use in care.
Chapters Emerging AI Technologies in Healthcare · AI Tools Every Physician Should Know
Perets et al.: bias can enter clinical records before model training
Perets and colleagues trace EHR bias to processes preceding model training; referral and admission bias is the most frequently documented source.
Why Assessing training data requires examining how care decisions and measurements produced the record.
Chapters The Clinical Data Challenge · Medical Ethics, Bias, and Health Equity
Celi et al.: clinical AI reliability extends beyond benchmark accuracy
Celi and colleagues frame reliability around knowledge boundaries, equity, community governance, and trustworthy evidence; their introduction provides an editorial framework.
Why Reliability assessments can examine whether systems communicate uncertainty and remain accountable in the settings where they are used.
BRIE authors report chart-answer omissions despite supported factual claims
BRIE authors report that omission dominates chart-answer failures in a retrospective, mostly model-graded preprint from one health system.
Why Checking unsupported claims alone misses information that a chart answer leaves out.
TRACE authors report reduced note bloat before LLM processing
TRACE authors report that removing templated and copied note text reduces LLM input length without large performance declines in a Stanford preprint.
Why Copied and templated text adds input burden when chart assistants read longitudinal records.
Chapters Large Language Models in Clinical Practice · The Clinical Data Challenge
CytoVI authors report CLL separation using routine flow-cytometry panels
CytoVI authors report CLL separation, including minimal residual disease, from routine clinical flow panels in a Basel cohort.
Why Generative antibody-panel models provide a clinical flow-cytometry example alongside classical decision-support approaches.
Daneshjou: staged AI use should preserve scientific judgment
Daneshjou’s editorial supports deliberate, staged AI use that preserves the cognitive skills needed for independent scientific judgment.
Why Training programs can align AI access with demonstrated competence while preserving opportunities to develop unaided judgment.
Chapter AI in Medical Education
Omar et al.: verification reminders reduce unsafe-option selection
Omar and colleagues find that a verify-or-escalate reminder reduces unsafe-option selection in an experimental evaluation without eliminating the failure.
Why Safety assessment must test obedience to unsafe instructions as a distinct failure mode alongside unsupported claims and bias.
Chapters Clinical AI Safety and Risk Management · Large Language Models in Clinical Practice
Iannone and Giansanti: autism assistive technology remains an early research area
Iannone and Giansanti’s narrative review describes prototype-focused autism assistive technology with unresolved implementation needs and no established long-term clinical benefit.
Why Evaluation of autism assistive technology must account for usability and fit across a heterogeneous spectrum.
Chapter Pediatrics and Neonatology
ECG AI diagnosis yield and delayed testing require separate evaluation
EAGLE evaluates low-ejection-fraction diagnosis yield, while delayed ECG acquisition can prevent timely STEMI recognition before interpretation begins.
Why Clinical evaluation must address both the consequence of acting on an AI alert and timely access to the underlying test.
OECD Madrid action plan offers non-binding steps for health AI
Healthcare Policy and AI Governance describes the OECD's Expert-Led Madrid Action Plan as ten actions for scaling AI in health. Like the OECD AI Principles, the plan is non-binding.
Why Implementation advice needs to be distinguished from enforceable obligations.
JAMA Psychiatry cohort examines AI-associated psychosis in a medical setting
Psychiatry describes a JAMA Psychiatry cohort study designed to estimate AI-associated psychosis prevalence in a large medical setting and characterize affected patients and conversational AI's role in documented episodes. The observational design cannot establish that chatbot exposure causes psychosis.
Why A medical-setting cohort extends the evidence beyond individual case reports.
Chapter Psychiatry and Behavioral Health
AI-MARRVEL prioritizes a BRSK1 candidate supported by subsequent clinical evidence
Allergy, Immunology, and Medical Genetics describes AI-MARRVEL variant prioritization and the BRSK1 report. An AI-generated candidate was followed by a matched patient cohort and functional evidence; prioritization alone did not establish the gene-disease association.
Why Genetic diagnosis requires evidence beyond a model's ranking of candidate variants.
Yu et al.: patients state high minimum acceptable AI detection performance
Ophthalmology describes Yu et al.'s survey in which mean minimum acceptable sensitivity and specificity were both 0.90. These are stated preferences rather than deployment outcomes, supporting discussion of false negatives, false positives, and predictive values in the intended population.
Why Patient expectations should be addressed with interpretable performance measures.
Chapter Ophthalmology
Barrit et al.: adversarially steered AI shifted physicians toward harmful decisions
Integration into Clinical Workflow describes a participant-blinded randomized preprint: 225 physicians made neurocritical-care decisions using one open-weight model with concealed aligned or adversarial instructions in anonymized cases rather than live care. Harmful final decisions rose from 2.0% to 49.4% with adversarial AI; only two physicians spontaneously identified steering.
Why Clinician review alone cannot be relied on to catch a conversational system whose objective has been altered.
OpenAI reports GPT-6 Sol and Luna self-harm safety regression
Evaluating AI Systems describes OpenAI's reported MentalHealthBench gains for GPT-6 Sol and Luna alongside regression versus GPT-5.6 on a separate self-harm evaluation without system-level safeguards. These developer evaluations do not establish patient benefit or therapeutic safety.
Why Model updates require renewed checks of mental-health behavior even when benchmark scores improve.
Google reports broader watermark checking; a negative result cannot authenticate video
Medical Misinformation describes Google's reported checker for marks from Google, OpenAI, NVIDIA, and Kakao tools. A negative result cannot authenticate a video because some generators add no watermark and research documents removal while preserving image quality.
Why Watermark checks provide more useful evidence when a mark is detected than when none is found.
Chapter Medical Misinformation and AI
VA ambient scribe contract has a shared ceiling across awarded contracts
AI-Assisted Clinical Documentation describes the VA's $775.72 million shared ambient scribe contract ceiling over five years. Abridge reports availability through a distribution partner; the vehicle does not establish Abridge-specific committed spending, nationwide deployment, or clinical benefit.
Why Procurement ceilings and vendor availability need to be distinguished from local use and measured outcomes.
FDA grants Queen of Hearts De Novo authorization for ACS ECG assessment
Cardiology and Cardiothoracic Surgery records Queen of Hearts' FDA De Novo authorization, DEN250044, on September 3, 2026, with product code SHS. The manufacturer reports ECG findings intended to support urgent ACS assessment; authorization does not establish local validation, autonomous MI diagnosis, or improved outcomes.
Why Evidence from another ECG product or workflow cannot be transferred to this device.
CMS FY 2027 NTAP covers qualifying CARE Body CT inpatient cases
Healthcare Policy and AI Governance describes temporary NTAP eligibility for CARE Body CT beginning October 1, 2026, with a maximum add-on of $137.53 for qualifying Medicare inpatient cases. CMS did not evaluate substantial clinical improvement under this alternative pathway, and payment does not establish local value.
Why Hospitals must distinguish reimbursement eligibility from demonstrated patient benefit.
Bearelly et al.: retinal imaging for first-trimester preeclampsia prediction
Obstetrics and Gynecology describes Visionary AI's first-trimester retinal-imaging study: a model transferred without retraining to 66 analyzable NYU pregnancies reached AUC 0.81. The two-site evidence remains preliminary, with calibration, clinical thresholds, and effects on aspirin decisions or outcomes untested.
Why External prediction performance is an early step toward evaluating a screening pathway.
Chapter Obstetrics and Gynecology
Mansuri et al.: Aidoc PE sensitivity varies after deployment
Diagnostic Imaging, Radiology, and Nuclear Medicine describes a retrospective preprint from one academic health system reporting lower PE sensitivity for subsegmental and non-acute emboli. Comparison with manufacturer validation is limited by unverified version equivalence; labels came from reports written with AI flags available.
Why Monitoring by embolus location and acuity can reveal limitations hidden by aggregate performance.
California AB 1979 and SB 503 preserve judgment and require bias monitoring
Medical Liability and Regulatory Compliance describes California AB 1979's protection of licensed professional judgment and prohibition on AI independently performing licensed clinical functions, alongside SB 503's bias mitigation, documentation, and monitoring duties. Most provisions take effect January 1, 2027.
Why California practices and developers face specific responsibilities for clinical decision support.
Utah permits a staged Nolla Health topical acne prescribing pilot
Healthcare Policy and AI Governance describes Utah's agreement allowing Nolla Health to issue specified topical acne prescriptions to eligible adults. Physicians review every prescription before transmission in the first stage; later review stages require written approval, and regulatory relief does not establish clinical effectiveness.
Why The agreement defines where prescribing authority, physician oversight, and ongoing reporting apply.
Chapters Healthcare Policy and AI Governance · Clinical Case Study Library
FDA CDRH plans mental-health device evidence guidance for FY 2027
Psychiatry describes CDRH's FY 2027 priority for draft evidentiary recommendations on generative AI conversational devices for mental disorders, with comments on priorities through November 30, 2026. Clinical AI Safety and Risk Management notes that finalizing lifecycle guidance is also a priority; the list is a publication plan.
Why A planned guidance does not yet specify the evidence a developer must submit.
Chapters Psychiatry and Behavioral Health · Clinical AI Safety and Risk Management
CMS applies a CY 2026 efficiency adjustment to non-time-based radiology services
Diagnostic Imaging, Radiology, and Nuclear Medicine describes the CY 2026 Physician Fee Schedule's 2.5 percent downward efficiency adjustment to work RVUs and intraservice physician time for non-time-based services, including eligible radiology codes. Time-based codes, telehealth-list services, and new CY 2026 codes are exempt.
Why Payment assumptions affect the financial case for workflow changes.
Gu et al.: multimodal benchmark answers can survive removal of the image
Evaluating AI Systems describes Gu et al.'s Nature Medicine stress tests: models could answer correctly with key inputs removed, change answers after small prompt alterations, and produce flawed reasoning traces. The results measure benchmark brittleness under perturbation, rather than clinical error prevalence.
Why A correct benchmark answer does not establish that a model used the clinically necessary input.
Chapters Evaluating AI Clinical Decision Support Systems · Clinical AI Safety and Risk Management
AAP policy requires human oversight of pediatric generative AI
Pediatrics and Neonatology describes the AAP policy statement requiring direct human oversight of current clinical generative AI, with ultimate responsibility remaining with the pediatrician. It calls for validation across diverse pediatric populations and addresses clinician use rather than direct use by children or families.
Why Pediatric adoption requires population-specific validation and an accountable clinician.
Chapters Pediatrics and Neonatology · Healthcare Policy and AI Governance
Wu et al.: consumer health AI can control routes through care
Large Language Models in Clinical Practice describes a Nature Health perspective on consumer AI integration across interfaces, contextual data, orchestration, and clinical workflows. The framework examines who controls symptom interpretation, care routing, payment, and treatment completion as platforms connect these functions.
Why Platform integration can shift control over a patient's care pathway beyond the chatbot answer.
Oncoformer predicts cancer diagnoses from routine records retrospectively
Hematology-Oncology and Precision Medicine describes Oncoformer's retrospective COMPASS evaluation using routine clinical records and chest radiographs. In the independent replication cohort, AUROC was 0.869 for a cancer diagnosis up to one year later after excluding data within 90 days of diagnosis; this does not establish screening benefit or authorization.
Why Lead-time discrimination requires prospective evaluation before it supports a screening pathway.
Chapters Hematology-Oncology and Precision Medicine · Diagnostic Imaging, Radiology, and Nuclear Medicine
Mondal et al.: public performance evidence covers 30 of 77 pathology AI devices
Pathology and Laboratory Medicine reports public device-specific performance evidence for 30 of 77 CE-marked or FDA-authorized pathology and hematology-morphology AI devices in Mondal et al.'s analysis. Published studies used 15 metrics, limiting comparison across devices.
Why Authorization does not ensure that a purchaser can inspect a published device-specific evaluation.
Chapters Pathology and Laboratory Medicine · Evaluating AI Clinical Decision Support Systems
Luo et al.: LLM pre-consultation improves history quality in an eye-hospital trial
Ophthalmology describes Luo et al.'s single-center randomized comparison of LLM pre-consultation with ophthalmology residents. The trial found higher history-quality scores but longer interviews, supporting AI-collected history as input to the clinician's examination rather than diagnosis without one.
Why History quality and consultation duration both matter when evaluating intake assistance.
Chapter Ophthalmology
Christensen et al.: radiologist supply and imaging demand depend on assumptions
Diagnostic Imaging, Radiology, and Nuclear Medicine describes projections of Medicare-enrolled radiologist supply and imaging use through 2055. Residency growth, attrition, and per-person imaging trends change the estimates, so workforce forecasts depend on explicit assumptions.
Why Staffing plans require both supply and demand scenarios before assigning a role to AI.
Wenderott et al.: imaging workflow studies show mixed time effects
Diagnostic Imaging, Radiology, and Nuclear Medicine describes Wenderott et al.'s systematic review: 67% of 33 studies measuring task time reported a reduction, but three meta-analyses of 12 studies found no statistically detectable pooled effect.
Why A reported time saving in one workflow does not establish a general productivity gain.
Paramasamy et al.: chest CT reporting time varies by examination type
Diagnostic Imaging, Radiology, and Nuclear Medicine describes a retrospective before-and-after evaluation of Veye Lung Nodules at Erasmus Medical Center. Adjusted median reporting time fell from 21.3 to 18.2 minutes overall, while emergency department examinations took longer.
Why A local average can conceal workflows in which the same product increases reporting time.
Local disease prevalence changes the predictive value of imaging AI
Diagnostic Imaging, Radiology, and Nuclear Medicine describes Scaringi et al.'s small retrospective emergency deployment, where a large-vessel-occlusion algorithm had a positive predictive value of 33% at 4.1% prevalence. Baird et al. describe estimating predictive values and error rates from local prevalence and case volume.
Why Local base rates help clinicians anticipate false alerts before relying on validation figures.
Li et al.: simulated FDG-PET from MRI in a retrospective dementia study
Neurology describes SiM2P, which simulates FDG-PET from MRI. In a masked reader study of a held-out retrospective cohort, simulated PET improved dementia detection and differential diagnosis over MRI; it remains uncleared for diagnosis and cannot substitute for acquired FDG-PET.
Why Synthetic imaging requires appraisal against acquired imaging and clinically relevant decisions.
EASAC and FEAM recommend evaluation before and after clinical AI deployment
Healthcare Policy and AI Governance describes EASAC and FEAM's advisory recommendations to assess net benefit before and after deployment. Experimental use warrants informed consent; routine care under professional oversight generally needs notification when AI materially supports care, rather than separate AI consent.
Why The recommendations connect outcome evaluation with practical distinctions in patient communication.
Primary care AI renewal depends on payment and workflow
Primary Care explains how a small spending share, fragmented buyers, fee-for-service incentives, and additional clicks constrain AI adoption. Documentation, administrative automation, and bounded screening have clearer renewal paths than broad diagnostic copilots, whose performance alone does not resolve payment or liability.
Why Practices can appraise commercial viability alongside evidence for the exact clinical task.
Chapter Primary Care, Family Medicine, and Preventive Medicine
A tokenised patient journey remains a partial medical record
The Physician-AI Partnership describes Mahmood and Topol's Lancet essay on tokenisation: time-stamped codes, measurements, text, images, and intervals form a sequence a model can read. The medical record remains a partial trace, and a tokenised journey has not been validated as a digital twin.
Why Representing a record in sequence does not establish a validated simulation of the patient.
Aleph Surgical reports pre-seed funding and autonomous surgery laboratory work
Surgery and Anesthesiology records Aleph Surgical's reported $7.5 million pre-seed round and a company preview of autonomous vascular anastomosis on its Mark 0 robot. The work remains a laboratory demonstration moving toward operating environments.
Why A company demonstration requires a different appraisal from an authorized surgical platform or evidence of patient benefit.
Reis et al.: disclosing AI use lowers warmth, competence, and willingness
A preregistered US vignette experiment (N = 1,030) rated physicians who stated that they use AI as less warm (ηp² = 0.04) and less competent (ηp² = 0.05), with lower appointment willingness (ηp² = 0.07). After adjustment for multiple testing, that penalty did not meaningfully depend on physician age, gender, or race and was stronger among younger consumers. The ratings are vignette advertisements, not clinic visits (Large Language Models in Clinical Practice).
Why Disclosure of AI use changed how patients rated physicians on warmth, competence, and willingness to book, mostly apart from the physician's age, gender, or race.
Chapters Large Language Models in Clinical Practice · Evaluating AI Clinical Decision Support Systems
Bessemer 2026 scorecard: executives report faster administrative AI payback
Industry Reports & Benchmarks cites a Bessemer survey of 226 executives across 65 use cases: reported payback in about 12 months at roughly 3.5x cost, and 54% reported material ROI in year one. Revenue cycle was 4.0x and 67% semi- or fully autonomous, versus clinical use at 2.9x and 4% autonomy. 46% trusted AI for clinical decision support, 77% overrode suggestions more than half the time, and about 48% of payers would not reimburse fully autonomous AI care. It is self-reported executive ROI, not patient outcomes.
Why Health-system executives in this survey describe administrative AI as nearer to autonomous use and reimbursement than clinical AI.
Chapter Industry Reports and Benchmarks
References, links, and wording were also kept current.
September opened with OpenAI's Epic FHIR connector for ChatGPT for Healthcare and evidence across radiology, pathology, psychiatry, and other specialties. The FDA released a generative-AI discussion paper, and five society and UK documents on clinical AI were added. Oncology cited NHS-Galleri, whose primary endpoint was not met, and PATHFINDER 2. A Nature editorial and the Chen review state that 23% of roughly 4,600 clinical LLM papers used real-world patient data.
ARPA-H ADVOCATE funds heart-failure clinical AI agents
Healthcare Policy and AI Governance describes ARPA-H's ADVOCATE awards: up to $62.7 million over four years for patient-facing heart-failure agents, a supervisory agent, and implementation plans, with FDA authorization expected within 24 months of award. The program is research and regulatory pathfinding, not an authorized product.
Why A federal program is paying for patient-facing clinical agents and post-deployment supervision, with an authorization timetable that is still a plan.
Chen and a Nature editorial: 23% of clinical LLM papers used real-world data
Evaluating AI Clinical Decision Support Systems cites a September 2026 Nature editorial arguing that generative-AI tools informing clinical decisions need transparent real-world assessment, and the Chen review's finding that only 23% of roughly 4,600 clinical LLM papers used real-world patient data.
Why Publication volume in this literature still mostly reflects simulated or exam-style tasks rather than patient data.
CROWN: a research cytology encoder, not an FDA-cleared screener
Pathology and Laboratory Medicine describes CROWN, a cytology visual encoder pretrained on more than 10 million patches, which the paper's benchmarks ranked highest overall among established pretrained encoders, with accuracy exceeding 95% in 48 patch-level classifications and 98% in 22 (Zheng et al., 2026). It is a research backbone, not an FDA clearance or an autonomous screener.
Why Cytology has a generalist encoder analogous to histopathology foundation models, and it still needs task-specific validation before clinical use.
Google reports Argon found a flaw in hospital software
Large Language Models in Clinical Practice notes Google's 30 September 2026 report that Wiz Scan for Good used Gemini 4 Argon to find a critical vulnerability exposing personal information in healthcare software used by hospitals worldwide, which the company says earlier frontier models missed. The chapter treats this as a vendor-partner vignette about clinical-software supply chains, not as clinical decision support or bedside authorization.
Why A company report of a defensive cyber finding does not show that the model is safe or authorized for clinical use.
White House Accord on Super Intelligence names no enforcement
Healthcare Policy and AI Governance records the September 2026 White House Accord on Super Intelligence, signed by six frontier labs, as a voluntary commitment that names no deadlines, auditors, or penalties.
Why Readers can see that this 2026 accord, like earlier corporate AI pledges, does not itself create an enforcement mechanism.
Farrag et al.: clinical evidence-search tools miss much of a gold corpus
Large Language Models in Clinical Practice reports that five clinical evidence-search platforms had median formulation-level recall from 7.2% to 42.2% and pooled recall from 45.8% to 72.3%, with single-query zero-retrieval from 47% to 80% and 12.0% of gold evidence never retrieved (Farrag et al., 2026). Evaluating AI Clinical Decision Support Systems asks for those rates before a cited answer is treated as complete.
Why A cited passage can still leave most of a gold corpus unretrieved, so coverage failure is separate from whether a retrieved citation looks supported.
Chapters Large Language Models in Clinical Practice · Evaluating AI Clinical Decision Support Systems
MEDS: an open data standard for EHR foundation-model research
Emerging Technologies adds a NEJM AI review of the Medical Event Data Standard (MEDS), an open schema for EHR event data, reported in use at 21 institutions as of March 2026 (McDermott et al., 2026). The chapter notes MEDS serves predictive research and neither replaces OMOP for causal studies nor establishes any foundation model's clinical utility.
Why Training models on longitudinal structured EHR data calls for portable pipelines and simple event streams, a different priority from the deep harmonization OMOP provides for observational informatics.
NHS-Galleri randomized trial: multi-cancer screening misses its primary endpoint
Oncology now cites the NEJM report of the NHS-Galleri randomized trial of annual multi-cancer early detection screening in England, whose primary endpoint, a reduction in combined stage III and IV cancers, was not met, with companion performance reporting in Nature Medicine. The chapter reads secondary stage-shift and detection-rate claims against the null primary endpoint.
Why Distinguishes the trial’s unmet primary endpoint from secondary screening results.
PATHFINDER 2: Galleri multi-cancer blood test performance in 35,878 adults
Oncology adds the Nature Medicine report of PATHFINDER 2, which returned GRAIL's Galleri test results to 35,878 adults aged 50 or older; among 32,007 with 12-month status, positive predictive value was 60.3%, specificity 99.64%, and sensitivity 39.3% across all cancers (Nabavizadeh et al., 2026). Adding the test to recommended screening increased screen-detected cancers about 6.5-fold in a descriptive analysis.
Why Gives readers published performance figures for the test within a section that now treats analytical performance, diagnostic yield, stage shift, and mortality as separate endpoints.
ResidencyRL: Google DeepMind preprint on training clinical dialogue agents in simulation
Evaluating AI Systems adds a Google DeepMind preprint that trains a clinical dialogue agent with multi-turn reinforcement learning on simulated telehealth encounters; against a Gemini 3.5 Flash baseline, reported adversarial diagnostic accuracy rose from 81.0% to 88.0% (Liévin et al., 2026, preprint). All gains come from simulations, so the chapter treats it as a training method, not clinical-utility evidence.
Why Adds a training-method result beside the chapter's simulated-encounter benchmarks, such as CRAFT-MD and AgentClinic, which the preprint also used to test transfer.
OpenAI's MentalHealthBench for chatbot mental-health conversations
Evaluating AI Systems adds OpenAI's MentalHealthBench, in which more than 80 psychologists and psychiatrists wrote 5,262 grading criteria for 1,215 synthetic conversations, and Psychiatry links to it. The chapter notes that an OpenAI model grades only the chatbot's next reply, so the benchmark is not evidence that a chatbot helps patients or is safe as therapy.
Why OpenAI released a specialist benchmark covering mental-health conversations from routine questions to emergencies, complementing the chapter's HealthBench coverage.
Chapters Evaluating AI Clinical Decision Support Systems · Psychiatry and Behavioral Health
RADAR: generalist vision-language model for abdominal CT
Radiology & Nuclear Medicine adds a Science paper on RADAR, a vision-language model trained on 424,911 contrast-enhanced abdominal CT examinations, reporting mean AUC about 0.913 across 146 findings, external multi-center AUC about 0.895, and reader-assist gains in sensitivity and reading time; described as a research model, not FDA-cleared and distinct from the Nat Med LiON liver model.
Why Extends chapter coverage of report-supervised vision-language models from chest imaging to abdominal CT, with a large training and validation cohort.
Cross-country scaling of diabetic retinopathy AI screening
Ophthalmology adds a Nature Medicine Comment (not a trial) describing the scale-up of one diabetic-retinopathy deep-learning screening tool from a single hospital to India, Thailand, and Australia, reaching more than a million patients screened; framed as an implementation narrative rather than new randomized outcome evidence.
Why Gives readers a primary source on real-world, cross-country deployment scale of AI-based diabetic retinopathy screening, distinguished from the FDA-authorized pivotal trials already covered.
Chapter Ophthalmology
Five professional-society and UK guidance documents on clinical AI
Policy & Governance, Liability & Compliance, Internal Medicine, Oncology, Primary Care, and Psychiatry add five 2025-2026 guidance documents: a UK National Commission report on staged AI authorization; ACP's ethics paper on rationality, self-governance, and competence; ASCO's six oncology-AI principles; the AAFP/Rock Health Starfield Signal primary-care roadmap; and APA's Resource Document, which states no AI application is yet FDA-approved for psychiatric diagnosis or treatment.
Why Gives readers each society's own framing of clinical-AI guidance rather than a secondary summary, while noting none of these are binding law or a liability shield.
Chapters Healthcare Policy and AI Governance · Physician AI Liability and Regulatory Compliance · Internal Medicine and Hospital Medicine · Hematology-Oncology and Precision Medicine · Primary Care, Family Medicine, and Preventive Medicine · Psychiatry and Behavioral Health
Insurance claims and FDA update data show slow AI adoption
Evaluating AI Systems adds a claims study tracking AI-specific procedure codes across billions of U.S. insurance claims, finding medical AI adoption still nascent and concentrated in a few procedure families and higher-income, metropolitan, academic areas (Wu et al., NEJM AI), and an audit finding less than 2% of FDA AI device updates involved retraining on new data (Wu et al., CHIL 2024).
Why Shows that regulatory clearance counts and technical capability overstate how much AI has actually reached bedside use.
EAGLE: opportunistic esophageal cancer detection on chest CT
Radiology & Nuclear Medicine adds EAGLE, a two-stage model for esophageal malignancy on noncontrast chest CT externally tested across eight centers in three countries (n = 11,466) at about 98.5% specificity and about 90% sensitivity, with a prospective hospital run reaching about 42% positive predictive value (Zhou et al., 2026).
Why The chapter notes several industry-affiliated authors and closed code and data, and that this is diagnostic performance, not an established mortality benefit.
Chapters Hematology-Oncology and Precision Medicine · Diagnostic Imaging, Radiology, and Nuclear Medicine
An AI Productivity Index for clinical AI procurement
Evaluating AI Systems adds a framework proposing that procurement decisions weigh implementation burden, cognitive load, downstream resource use, and equity of benefits and burdens alongside discrimination metrics (Al-Ajlouni and Njei, 2026).
Why Offers a conceptual complement to safety and efficacy appraisal, not a certified productivity score.
Case-tracker analysis of AI-related legal harms in care
Liability & Compliance adds an analysis of 31 U.S. legal cases and reported harms involving AI in care delivery, arguing that patients often face tools deployed by insurers and facilities, not only bedside clinicians (Mansi and Riedl, 2026).
Why Broadens the liability discussion beyond physician-centered malpractice scenarios to the full set of parties that deploy clinical AI.
Epoch AI tracks a rise in math-arXiv AI-acknowledgment rates
Research with AI adds a report finding a sharp 2026 rise in AI-acknowledgment disclosures on math arXiv papers, including about 12% of established-author math papers disclosing research-assistance-class AI use by August 2026 (Epoch AI, 2026).
Why Offers a disclosure-norm signal from one STEM field, not a rate applicable to clinical manuscripts.
Chapter Clinical Research with AI
MMBU: imaging benchmark exposes a familiar transfer gap
Evaluating AI Systems adds MMBU, a preprint benchmark testing 15 open-weight and 2 frontier vision-language models across 35 biomedical imaging submodalities; medical adaptation improved familiar suites such as PathVQA and VQA-RAD, but those gains often did not transfer, and no model beat a random baseline on object detection (D'Cunha, Lozano, Sun, Yeung-Levy et al., 2026, preprint).
Why Extends the chapter's benchmark-transfer caution from text-based leaderboards to imaging vision-language models.
RAG conflict, citation support, and fine-tuning limits in clinical LLMs
LLMs in Practice adds three studies: ClashEval finds LLMs adopt incorrect retrieved content over correct prior knowledge more than 60% of the time (Wu, Wu, and Zou, 2024); SourceCheckup finds 50 to 90% of LLM medical answers are not fully supported by their cited sources (Wu et al., 2025); and a NEJM AI study finds commercial fine-tuning APIs teach new medical knowledge poorly, averaging about 37% generalization accuracy (Wu, Wu, and Zou, 2025).
Why Together, the three studies show that retrieval, citation, and fine-tuning each fail to guarantee that an LLM's answer reflects the source it cites.
A capability map from predictive models to multi-agent systems
AI Fundamentals for Clinicians adds an at-a-glance table mapping shorthand terms (predictive ML, deep neural models, generative models, AI agents, multi-agent workflows) to their added capability and a clinical example, framed as a containment hierarchy rather than a strict maturity ladder.
Why Gives readers a quick reference for terminology used throughout the handbook before the AI-agents discussion.
Chapter AI Fundamentals for Clinicians
Conversational AI needs prospective, not just benchmarked, evidence
Evaluating AI Systems adds two commentaries arguing that trust in conversational medical AI must be earned in real-world clinical settings, not benchmarked into existence (Schaekermann et al., 2026), and that evaluation should pair standardized prospective studies and post-deployment monitoring with LLMs' fast iteration cycles, without lowering evidentiary standards (Mahajan et al., 2026).
Why Sharpens the claim-to-evidence rule specifically for conversational and LLM-based systems, which update faster than clinical trials can evaluate them.
Patient-facing generative AI shapes understanding before the visit
LLMs in Practice adds a commentary arguing that the deeper risk of patient-facing generative AI is interpretive: it shapes how people understand symptoms, risk, and authority before any visit, so evaluation should track uncertainty framing and disclosure, not accuracy screens alone (Parikh and Kohli, 2026).
Why Extends the chapter's patient-facing AI caution beyond factual accuracy to how the tool frames uncertainty and authority.
I³LUNG: multimodal models for lung cancer immunotherapy selection
Oncology adds I³LUNG, a study of 2,396 patients with advanced non-small-cell lung cancer testing multimodal models that fused clinical, blood, imaging, pathology, and genomic data; clinical-and-blood-only models reached an AUC up to 0.77 on an independent test set but fell to 0.55-0.72 on external validation (Prelaj et al., 2026).
Why Shows that multimodal fusion looked stronger in development than it did on external validation, with prospective decision-support testing still ongoing.
CRISP: prospective evidence for AI-assisted intraoperative pathology
Pathology adds CRISP, a foundation model trained on more than 100,000 frozen sections and tested on more than 15,000 intraoperative slides, then in a prospective cohort of more than 3,000 patients where it informed surgical decisions in 92.6% of cases and human-AI collaboration reduced diagnostic workload by 35% (Zhao et al., 2026).
Why Frames CRISP as pathologist-governed support with multi-center prospective signal, not autonomous sign-out or proof of better oncologic outcomes.
BRIDGE benchmark: real-world clinical-text tasks stress LLMs
Evaluating AI Systems adds the BRIDGE benchmark, which tests language models on 87 real-world clinical-text tasks across nine languages rather than licensing-style exams; across 95 models, overall performance was far below typical exam-leaderboard scores, with billing and coding especially weak (Wu et al., 2026).
Why Demonstrates that exam-style leaderboard scores do not predict performance on the documentation and coding tasks clinicians actually need.
Five evidence phases for clinical-AI readiness claims
Evaluating AI Systems adds a Perspective mapping clinical-AI claims to five evidence phases from model development through post-deployment monitoring, warning that strong technical performance or regulatory authorization alone does not establish clinical readiness (Nenadic et al., 2026).
Why Gives readers a framework for sizing a vendor's claim to its evidence phase rather than assuming authorization equals proven benefit.
Echocardiogram videos as opportunistic kidney-disease screening
Cardiology adds a deep-learning model trained on more than 300,000 parasternal long-axis echocardiogram videos that detected chronic kidney disease, with external validation at two U.S. health systems, and the Glossary adds an opportunistic-screening entry (Yuan et al., 2026).
Why Frames this as screening discrimination in echo-referred cohorts, not a replacement for eGFR or proof that alerts improve outcomes.
Chapters Glossary of Medical AI Terms · Cardiology and Cardiothoracic Surgery
Clinician essay on ad-funded AI and query-data use
Medical Education and Privacy & HIPAA add an essay arguing that ad-funded clinical AI tools create an incentive to monetize physicians' knowledge gaps, and that a business-associate agreement and a no-PHI-training claim do not by themselves resolve whether de-identified query aggregates feed product or advertising systems (Wu, 2026).
Why The chapters treat this as an opinion essay and vendor policy language, not evidence of foundation-model training on clinician chats.
Chapters AI in Medical Education · Privacy, HIPAA, and Patient Data Security
A field-level taxonomy of health-AI reliability failures
AI Failures & Lessons and Evaluating AI Systems add a Cell perspective cataloging recurring reliability failures in predictive and generative health AI, including incorrect outputs, performance gaps across groups and settings, and post-deployment degradation (Salaudeen et al., 2026).
Why The authors call for lifecycle-aware evaluation and continuous monitoring rather than one-time model scores, framing the taxonomy as a tool for case review and vendor diligence.
Chapters AI Failures in Medicine: Lessons Learned · Evaluating AI Clinical Decision Support Systems
AI alone may sometimes outperform physician-AI hybrid care
Evaluating AI Systems and Policy & Governance add a Perspective posing that for some cognitive tasks, AI alone may outperform both physician-only care and physician-AI hybrids, shifting the real design questions onto workflow, liability, regulation, reimbursement, and training (Emanuel et al., 2026).
Why Challenges the default assumption that hybrid human-AI care is always the safer choice, while noting this is a Perspective, not ward-outcome data.
Chapters Healthcare Policy and AI Governance · Evaluating AI Clinical Decision Support Systems
Retinal imaging as an atrial fibrillation risk marker
Cardiology and Ophthalmology add a two-cohort study finding thinner macular ganglion-cell layers on routine retinal imaging associated with prevalent and incident atrial fibrillation, and the Glossary adds an oculomics entry describing the concept (Huemer et al., 2026).
Why Positions opportunistic eye imaging as a possible early AF-risk marker, not a stand-alone diagnostic test.
Chapters Glossary of Medical AI Terms · Cardiology and Cardiothoracic Surgery · Ophthalmology
Paper2Agent: an MCP-based virtual corresponding author
Research with AI adds a pointer to Paper2Agent, which packages a computational methods paper's manuscript, code, and workflows into an MCP-backed tool so readers can apply the methods in natural language instead of installing the repository (Miao et al., 2026).
Why Useful for reusing computational methods; the chapter notes this is not a clinical outcome study.
Chapter Clinical Research with AI
Institute for Progress essay on locked FDA trial dossiers
Research with AI adds a pointer to an essay arguing that regulatory AI tools trained on private sponsor filing archives can widen incumbent advantages created by locked FDA clinical trial dossiers (Teslo, Institute for Progress, 2025).
Why Flags this as a policy proposal about data access, not clinical outcome evidence, with fuller treatment in a companion handbook.
Chapter Clinical Research with AI
AI as a population health determinant: ambient exposure framing
Policy & Governance adds a pointer to a population-exposure framing that treats AI as a health determinant, distinguishing ambient algorithmic exposure from personal tool use, with fuller treatment in a companion handbook (Parikh et al., 2026).
Why Extends the chapter's policy framing to population-level AI exposure, not just individual clinician tool use.
On-premise clinical agents with consistency-gated autonomy
Evaluating AI Systems adds a study coupling institutionally governed on-premise deployment with decision-time reliability checks: an open-weight model matched a cloud baseline's diagnostic accuracy on MIMIC-IV (about 90.0% versus 90.7%), and a consistency-based gate retained 49.4% of cases at 98.9% accuracy for selective autonomy (Zhang et al., 2026).
Why This is a retrospective simulation, not a prospective outcomes trial, and the thresholds require local calibration before use.
Chapters AI Fundamentals for Clinicians · Evaluating AI Clinical Decision Support Systems · Clinical AI Safety and Risk Management
Structural heart disease screening and AI-assisted heart-failure care
Cardiology adds an ECG-image model validated for structural heart disease screening across hospital and population cohorts (Dhingra et al., 2025), a study linking deep-learning coronary calcium on routine chest CT to cardiovascular outcomes (Peng et al., 2023), and a randomized pilot in which a generative-AI workflow supported nonmedical staff delivering heart-failure medication optimization under cardiologist approval (Navarese et al., 2026).
Why None of the three studies alone establishes that screening, or the AI workflow itself rather than the surrounding care model, improves outcomes.
Google DeepMind survey on AI use in scientific research
Research with AI adds a science-workforce survey reporting large self-reported time savings from AI alongside substantial verification overhead and a growing backlog of untested hypotheses among scientists (Codreanu et al., 2026).
Why Notes this is science-wide survey evidence, not clinical outcome data, with fuller treatment in a companion handbook.
Chapter Clinical Research with AI
GPT-4o vignette trial added to the evaluation chapter
Evaluating AI Systems adds the multi-country GPT-4o vignette trial, with a caution that these are controlled vignette scores versus a no-resource control, not bedside outcomes (Rounding et al., 2026).
Why Extends the evaluation chapter's evidence base to include this cross-country benchmark alongside its other vignette studies.
AI-as-intervention versus AI-for-trial-operations distinction
Clinical Trials and AI adds a review distinguishing AI whose output is part of the assigned treatment, which needs prospective evaluation with prespecified estimands, from AI that supports trial design, recruitment, monitoring, or analysis without being the intervention (Armoundas et al., 2026).
Why Gives readers a framework for matching the evidence standard to the AI's actual role in a trial.
Chapter Clinical Trials and AI
AI-agent eye clinic: early workflow lessons
Ophthalmology and Workflow Integration add a Comment describing early lessons from a real-world AI-agent eye clinic that routes patients through pre-consultation, triage, diagnosis, and an ophthalmologist copilot, with the ophthalmologist retaining the final decision (Yan et al., 2026).
Why Frames this as tertiary-hospital implementation experience, not autonomous-screening evidence or proof that multi-agent clinics improve outcomes.
Chapters Integration into Clinical Workflow · Ophthalmology
Patient-facing RAG chatbot leaked data without authentication
Privacy & HIPAA and Safety & Risk add a Perspective describing an anonymized assessment in which a standard browser, developer tools, and a commercial AI assistant retrieved a deployed chatbot's system prompt, configuration, knowledge base, and 1,000 stored patient conversations without authentication (Madrid-Garcia et al., 2026).
Why The authors argue retrieval-augmented generation is not itself a security boundary, so deployments need evaluation as complete software systems.
Chapters Privacy, HIPAA, and Patient Data Security · Clinical AI Safety and Risk Management
JAMA+ interview on AI for clinical-trial enrichment
Clinical Trials and AI adds a pointer to a JAMA+ AI Conversations interview on machine learning for trial enrichment and FDA expectations around AI biomarkers (Perlis, 2026).
Why Investigators using machine learning for trial enrichment need to understand FDA expectations around AI biomarkers.
Chapter Clinical Trials and AI
Economic history against a clinical-workforce collapse narrative
Future Perspectives adds a Perspective arguing that efficiency gains can expand demand (Jevons paradox), that clinical work is not fixed in quantity, and that automating tasks differs from automating jobs when care depends on interlocking human skills (Khullar, 2026).
Why Offers a counterweight to simple predictions that AI will shrink the physician workforce.
Multi-country GPT-4o vignette trial and equity gaps
Global Health Equity and LLMs in Practice add a randomized trial of GPT-4o assistance on clinical vignettes across Indonesia, Kenya, and the Netherlands, finding the largest absolute gain in Kenya (+18 percentage points) and a narrower Kenya-Netherlands performance gap under AI assistance (Rounding et al., 2026).
Why Suggests AI assistance can narrow between-country vignette performance gaps, though the control arm lacked usual resources and outcomes were not assessed.
Chapters AI and Global Health Equity · Large Language Models in Clinical Practice
Vignette evidence on AI liability judgments and documentation duties
Liability & Compliance adds a randomized vignette study finding higher reasonableness ratings for accepting standard-care AI advice than for rejecting it (Tacconelli et al., 2026), and Clinical Documentation notes that potential duties to audit AI drafts can involve clinicians, institutions, and manufacturers (Gerke et al., 2026).
Why Distinguishes controlled-vignette judgment findings from real malpractice outcomes or a documentation mandate.
Chapters Physician AI Liability and Regulatory Compliance · AI-Assisted Clinical Documentation
Clinician accountability for AI-assisted decisions
Liability & Compliance adds a pointer noting that clinicians retain accountability for the clinical decision itself, not merely a check on the model's output, even as AI improves care (Parsons and Rodman, 2026).
Why Distinguishes a normative claim about professional role from any finding about malpractice outcomes.
Residency-level benchmarks for clinical agents
Evaluating AI Systems adds a commentary distinguishing exam-style scores, which mostly test medical knowledge, from residency-level benchmarks that ask whether an agent can safely act in clinical workflows before earning autonomy in the EHR (Lancet Digital Health, 2026).
Why Gives readers a design standard for evaluation, not evidence that any deployed agent meets it.
Validation Accords call for generative-AI validation standards
Evaluating AI Systems adds a correspondence calling for a consensus framework to validate generative AI in medicine (Huo et al., 2026).
Why Notes this is an invitation to develop standards, not a finished checklist or regulatory requirement.
Five-phase framework for medical AI evaluation
Evaluating AI Systems adds a proposed medical-AI evaluation framework sequencing five gated phases (technical validation, operational robustness, controlled interaction, clinical evidence, and real-world integration), with fall-back and re-entry triggers for drift, version changes, or safety signals.
Why Offers readers a structured evaluation ladder, though the chapter notes it is an architecture proposal, not a regulatory checklist or evidence any deployed system has cleared all five phases.
Open-source hospital LLMs struggle on PHI de-identification
LLMs in Practice adds a Swiss tertiary hospital study of six open-source LLMs (8B to 24B parameters) on real French clinical notes: near-perfect scores on simple retrieval but a best PHI de-identification macro-F1 of about 0.33 to 0.34, well below a fine-tuned RoBERTa baseline of 0.94, with one model producing unsupported or invented content in 55% of reviewed translations.
Why Tempers enthusiasm for locally deployed open-source clinical LLMs with a single-center study showing sharp performance drops on harder tasks.
AI colonoscopy detection gains concentrated in diminutive polyps
Surgical Subspecialties adds a network meta-analysis of 13 RCTs (4,156 patients, 8 AI colonoscopy systems) finding the pooled adenoma detection gain concentrated in diminutive polyps, with minimal effect on small and large polyps and low to very low GRADE certainty.
Why Adds size-stratified nuance to adenoma detection rate claims without concluding that AI-assisted colonoscopy lacks clinical value.
Chapter Surgical Subspecialties
Version-to-version variation in a mammography AI tool
Radiology & Nuclear Medicine adds a retrospective analysis of 117,709 Norwegian screening mammograms showing that upgrading one commercial AI model from version 1.7 to 2.1 shifted which cancers and false positives reached its highest malignancy score, changing AUC from 0.908 to 0.928 (Larsen et al., 2026).
Why Shows a software version update can change which examinations a risk-adapted reading pathway routes to double reading, so each version needs its own local evaluation.
Pediatric LLM gaps, sepsis calculator update, and AI ethics framework
Pediatrics adds a scoping review noting clinical-text LLM studies remain retrospective and untested in real-time workflows (Huang et al., 2026); an updated neonatal sepsis calculator refresh, not independent validation (Kuzniewicz et al., 2024), which can delay antibiotics in some culture-positive infants (Pettinger et al., 2020); and a proposed ACCEPT-AI framework for age-related bias checks (Muralidharan et al., 2023).
Why Gives readers primary sources for the evidence gaps and limits behind existing pediatric AI tools and research.
Chapter Pediatrics and Neonatology
Mammography AI evidence organized by screening workflow
Radiology & Nuclear Medicine compares which screening workflows the MASAI trial and the PRAIM observational study each evaluated, noting differences in reading pathway, comparator, and what results do and do not transfer.
Why Clarifies that AI-support findings are pathway-specific and do not generalize across different screening workflows.
Cognitive offloading evidence in medical education
Medical Education adds a cognitive-science review finding basic capacities such as working memory may be more resilient to AI offloading than learned skills, and a randomized high-school study where answer-oriented GPT-4 use worsened unaided exam performance, largely mitigated by hint-based tutoring.
Why Gives readers primary evidence on how AI tool design affects skill retention and learning outcomes.
Chapter AI in Medical Education
FDA's generative AI medical device discussion paper
Policy & Governance adds FDA's August 18, 2026 discussion paper proposing a two-axis risk heuristic (device activity by consequence of an incorrect output) and a competency-based premarket approach for generative AI enabled devices, open for public comment through October 19, 2026 under docket FDA-2026-N-7874.
Why This is a discussion paper seeking input, not draft or final guidance, so the chapter is explicit that it does not yet state FDA's regulatory expectations.
Physician judgment in Medicaid work-requirement exemptions
Policy & Governance adds a JAMA Health Forum argument that large language models should not judge the credibility of medical-frailty exemption attestations under new Medicaid work-requirement rules, alongside a JAMA Internal Medicine piece describing the physician's role as the exemption arbiter.
Why Addresses a new administrative task, coverage-exemption attestation, where AI-assisted judgment could carry direct coverage consequences for patients.
Gray-market peptide sourcing risk noted for patients
Medical Misinformation adds guidance to treat patient-brought gray-market research peptides or forum dosing schedules as a behavioral signal rather than clinical guidance, since vial identity, purity, and sterility cannot be assumed.
Why Extends the chapter's misinformation coverage to an unregulated supply channel patients may already be using, pointing to the companion handbooks' population-level treatment of the same risk.
Chapter Medical Misinformation and AI
CLExEval: LLM judges can approve wrong diagnoses
Evaluating AI Systems adds a preprint stress-testing rare diagnostic narratives with progressive information masking: one model's accuracy fell from 95.0% to 32.5% under the heaviest masking, and LLM judges still passed clinically incorrect outputs at approval rates as high as 100% for one model even when given the correct diagnosis.
Why Shows that automated LLM-judge scoring can miss clinically incorrect answers, reinforcing the chapter's caution about relying on automated evaluation alone.
Temporal foundation models and patient-journey tokenization
Temporal foundation models convert heterogeneous, time-stamped patient records into learnable sequences. The Clinical Data Challenge cites five research-scale examples (Delphi-2M, Curiosity, Apollo, ALADYNOULLI, and AURORA) as hypothesis-generating, not cleared, tools, and the companion handbooks cover the population-scale framing of the same idea.
Why Introduces a data-representation approach the chapter had not covered, while noting the authors' industry advisory and equity ties and that observational models do not establish causality.
Chapter The Clinical Data Challenge
A taxonomy for autonomous agentic AI systems in telehealth
Emerging Technologies adds a taxonomy distinguishing simple chatbots, task-specific agentic systems, and autonomous agentic AI systems with persistent memory and messaging-platform hooks, and sketches governance across technical safeguards, personnel rules, and regulatory oversight.
Why Gives the multi-agent section vocabulary for systems that can execute actions and touch records, where prompt injection or unintended writes can propagate further than a suggestion-only chatbot allows.
NEJM AI editorial: guideline machines
LLMs in Practice adds an NEJM AI editorial arguing that outdated clinical guidelines could be improved with LLM assistance if not for real dangers, proposing that AI focus on errors of omission and produce a bias-free summary of key messages.
Why The editorial argument is not outcome evidence, so the chapter frames it as one clinician's proposal rather than a validated approach.
Regulating AI by its influence on a person's health trajectory
Ethics & Health Equity adds a comment arguing health AI should be regulated in proportion to how much it can shape a person's health trajectory, including use outside the clinic, noting this is an abstract-level claim without a multi-level decision-authority scale.
Why Adds an early framing for governance scope that reaches beyond the clinic, pending a fuller published version.
Clinical evidence LLMs, and scaffolding versus substitution in training
LLMs in Practice compares clinical evidence LLM products, while Medical Education adds a scaffolding-versus-substitution framework distinguishing trainees who form an independent differential before consulting an evidence LLM from those who let the tool generate it first.
Why Independent testing can overturn marketing claims about evidence tools, and early studies suggest substitution during training risks under-developing diagnostic reasoning.
Chapters AI in Medical Education · Large Language Models in Clinical Practice
Real-world mammography AI performance and a new clearance
Radiology & Nuclear Medicine adds a US pre/post study of single-reading Transpara v1.7.4-A finding no significant change in recall or cancer-detection rates and 15 of 171 cancers missed, a multi-site analysis finding the same AI's performance varies by device, and FDA's August 2026 clearance of the newer Transpara 2.1.0 as a counterpoint to European double-reading results already on the page.
Why Real-world, single-reading US evidence complicates the page's existing European double-reading results, and the clearance record shows local evidence must be matched to the exact software version deployed.
Biased AI can harm diagnosis; explanations do not rescue it
Workflow Integration adds a randomized vignette study of hospitalists, nurse practitioners, and physician assistants across 13 states (n=457) in which standard AI support raised diagnostic accuracy by 2.9 to 4.4 points, while systematically biased AI lowered it by 9.1 to 11.3 points; Evaluating AI Systems notes that image-based explanations did not significantly rescue accuracy when the model was biased.
Why Shows that explainability features are not a safety control for a systematically biased model, reinforcing the chapters' trust-calibration guidance.
Chapters Evaluating AI Clinical Decision Support Systems · Integration into Clinical Workflow
Training-cohort geography in US deep-learning studies
Evaluating AI Systems adds a JAMA research letter mapping geographically identifiable US cohorts in 74 imaging deep-learning studies from 2015 to 2019: California, Massachusetts, and New York accounted for most identifiable cohorts, while 34 states contributed none.
Why Even multi-site US studies can draw from a narrow geography, so the chapter reinforces that external validation in the deployment setting is still required.
Model Facts labels for point-of-care transparency
Workflow Integration adds a proposed Model Facts label, a one-page, clinician-facing summary of how, when, and when not to use a machine-learning model at the point of care, positioned alongside developer-oriented model cards.
Why Gives clinicians a concrete transparency format that supports appropriate use without replacing local validation or governance review.
Training-set sex imbalance in chest radiography
Ethics & Health Equity adds a study across three neural-network architectures and two public chest radiograph datasets finding classifier performance consistently declines for the underrepresented sex as training-set imbalance increases.
Why Distinguishes training-set sex skew from the underdiagnosis pattern in Case 5, supporting the chapter's call for subgroup reporting during validation.
Chapters Medical Ethics, Bias, and Health Equity · Diagnostic Imaging, Radiology, and Nuclear Medicine
Opinion essay on clinical-trial access delays
Clinical Trials and AI adds a physician-investigator framing paragraph responding to a New York Times Opinion essay on trial-access delays, noting the essay is opinion and interview synthesis, not measured trial-startup data, and pointing to the companion handbooks' clinical-trial latency coverage.
Why Frames a widely read opinion piece against what the handbook can and cannot verify about trial-startup delay.
Chapter Clinical Trials and AI
New ethics case study: chest X-ray underdiagnosis
Ethics & Health Equity adds Case 5, documenting how chest X-ray classifiers can label disease as no finding more often in under-served groups, an underdiagnosis pattern distinct from the proxy and target failures in earlier cases.
Why Adds a documented failure mode, disease present but labeled healthy, that can delay care if deployed in a triage workflow.
Chapters Medical Ethics, Bias, and Health Equity · Diagnostic Imaging, Radiology, and Nuclear Medicine
Three studies on AI-assisted emergency triage and adoption
Emergency Medicine adds a four-week live-department evaluation of a multi-LLM decision-support system where eligible-case adoption fell from 68% to 30% across each shift even though 99 of 100 sampled outputs were rated appropriate; a retrospective comparison finding 71.3% five-level agreement between an LLM and routine triage; and a multi-agent system scoring 87.04% exact match on a triage benchmark.
Why Adoption, agreement, and benchmark scores each fall short of showing AI triage tools are ready to replace clinician judgment; sustained engagement, not accuracy, was the adoption barrier.
Chapter Emergency Medicine
First systematic review of OpenEvidence
AI Tools for Physicians adds the first systematic review of OpenEvidence, covering 11 peer-reviewed studies through January 2026: lower citation-error rates than general-purpose chatbots where measured, 82% accuracy on 130 neurology items, and only 26.7% fully accurate answers on structural-heart questions versus 66.7% for ChatGPT-4o.
Why Tests marketing claims about a widely used clinical-reference tool against independent, though still limited, evidence.
A hip-fracture risk model and a virtual-trial implant study
Orthopedics adds FRACTURE-ML, a nationwide Swedish registry model (about 3.54 million people) reaching a holdout one-year AUC of 0.89 for hip-fracture risk and distinct from radiograph fracture detection, and a virtual-trial framework comparing two humeral stem designs across more than 500 simulated surgeries on 45 cadaveric humeri.
Why Neither study has external validation or an implementation trial; both are early-stage evidence for prediction and implant-design tools.
Cleveland Clinic's enterprise ambient-scribe rollout
Clinical Documentation adds Cleveland Clinic's account of onboarding more than 4,000 ambulatory clinicians to an ambient scribe within 16 weeks, reporting 70% encounter-level utilization among established users after a year, versus a published range of 20 to 42% ever-used adoption elsewhere.
Why Offers a large deployment case for training and onboarding, while noting the utilization figure applies only to established users, not the full clinician base.
Pediatric cardiology questions and retrieval augmentation
Pediatrics adds an evaluation of eight large language models on 304 board-style pediatric cardiology questions, with and without single-textbook retrieval-augmented generation; top models reached 98.4% accuracy without retrieval, which added no significant benefit except for one model on fellow-level items.
Why Shows that high multiple-choice accuracy does not by itself demonstrate educational effectiveness or bedside decision support.
Chapter Pediatrics and Neonatology
Role-prompting frameworks for tumor-board style concordance
Oncology adds a 100-case gastrointestinal evaluation of GPT-5 under five role-prompting frameworks, finding 78 to 87% concordance with institutional multidisciplinary-team decisions regardless of framework, while specialty-sounding language was nearly universal but uncorrelated with accuracy.
Why Warns that fluent, specialty-styled AI output is not the same as specialty-specific clinical reasoning.
Developer-run evaluation of an AI note-simplification tool
LLMs in Practice adds a developer-run evaluation of an AI note-simplification tool on synthetic clinician notes, reporting mean readability improving from a Flesch-Kincaid grade of 10.57 to 7.61 with 87.3% of expert ratings judged safe, alongside a 12.7% residual safety-flag rate.
Why The sole author is the tool's commercial developer and the notes were synthetic, so the chapter frames this as reported, not independent, evidence.
Physics-informed model for orthognathic soft-tissue prediction
Surgery & Anesthesiology adds PhysSFI-Net, a physics-informed model trained on 135 patients and externally validated on 33, reporting an internal shape error of 1.070 mm and an external shape error of 1.431 mm when predicting postoperative facial soft tissue after jaw repositioning.
Why These are retrospective geometric accuracy figures against postoperative imaging, not evidence for autonomous surgical planning.
Multimodal genomics and EHR AI: a synthesis review
Allergy & Immunology adds a synthesis describing multimodal AI frameworks that combine genomic, multi-omics, and longitudinal EHR data, naming the PennChart Genomics Initiative and eMERGE-PGx as clinician-facing decision-support examples, while noting polygenic and EHR models often reflect predominantly European-ancestry cohorts.
Why Contextualizes a review-level synthesis against the chapter's existing caution about ancestry bias in genomic risk tools.
Adult survey of reported LLM use for mental health
Psychiatry adds a stratified US adult survey (analytic n=1,871) in which 24% reported using LLMs, mostly ChatGPT, for mental health, alongside the chapter's existing adolescent and young-adult estimate; users were younger and reported more access barriers to care.
Why Extends the chapter's scale estimates from adolescents to adults, while flagging that reported use is not a measure of efficacy or benefit.
Chapter Psychiatry and Behavioral Health
OcularChat: a conversational model for AMD classification
Ophthalmology adds OcularChat, a multimodal model fine-tuned on AREDS color fundus photographs and simulated patient-physician dialogues, reporting 0.954 accuracy for advanced age-related macular degeneration on the internal test set but a lower 0.494 F1 score on external images.
Why Shows both the promise and the external-validation drop-off typical of dialogue-based imaging classifiers.
Chapter Ophthalmology
FDA discussion paper on generative AI medical devices
Liability & Compliance adds a note on FDA's discussion paper covering risk assessment, premarket evaluation, and postmarket considerations for generative AI enabled medical devices, alongside a matching entry in the regulatory milestones list.
Why FDA published a nonbinding discussion paper inviting comment on how it might regulate generative AI medical devices.
Ambient scribe documentation and AI text provenance
Clinical Documentation adds a matched-cohort finding that ambient scribe notes can carry more psychiatric symptom text without more documented intervention, and describes write-time provenance as the proposed way to track AI and clinician authorship in generative drafts.
Why Two new studies address distinct documentation risks: symptom-text inflation without added intervention, and the need for provenance when AI and clinicians co-author a note.
Patient-facing dermatology AI: lay self-triage evidence
Dermatology adds two patient-facing evidence points: reader experiments finding LLM explanations can raise lay automation bias and deference to AI assessments, and a Belgian cost-effectiveness analysis of the SkinVision skin-checking app. Neither establishes safe self-triage or treatment selection.
Why Both studies test claims about consumer dermatology AI tools that patients already use, without showing they are safe substitutes for clinical evaluation.
Chapter Dermatology
LLM safety hazards and a workflow-risk tier framework
Safety & Risk adds a review's map of LLM-specific hazards (prompt injection, data poisoning, sycophancy, informal shadow use) with hospital AI security operations as one named mitigation, and a four-tier (S1 to S4) workflow-risk classification showing the same diabetic-retinopathy tool can sit at different tiers depending on clinician oversight.
Why Two frameworks give clinicians vocabulary for AI-specific hazards and for how much oversight a given workflow requires, complementing the chapter's existing safety guidance.
Consumer chatbot mental-health harms: audit and scoping review
Psychiatry adds a multi-turn audit of nine general-purpose consumer chatbots finding concerning behavior was common, context-dependent, and accumulated over a conversation, and a scoping review mapping 119 papers on chatbot mental-health harms into five clusters, noting most crisis-response work is preliminary.
Why Both studies press on where routine chatbot safety testing falls short: harms that accumulate over a conversation and a thin evidence base for crisis response.
Chapter Psychiatry and Behavioral Health
Treatment-resistant depression model transportability
Psychiatry adds an L1-regularized EHR model of treatment-resistant depression fitted at the time of a first antidepressant prescription, reporting on its transportability across health systems including Mass General Brigham.
Why Gives readers evidence on whether a treatment-resistant depression prediction model generalizes when applied outside the health system where it was built.
Chapter Psychiatry and Behavioral Health
Four imaging AI studies added to radiology evidence
Radiology & Nuclear Medicine adds LiON, a CE-CT liver-malignancy additional reader from a large single-arm Chinese trial; PRISM, a non-contrast MRI liver-lesion assistance tool from a multicenter Chinese diagnostic study; a comparison finding consumer multimodal models reach only 51.6 to 62.7% open-ended accuracy on 252 neuroradiology cases; and a CT-first pulmonary-lesion risk model trained on 28,105 cases across four public datasets. None is FDA-cleared.
Why Expands the chapter's evidence on imaging AI tools that remain research-stage additional readers or risk-stratification aids, not autonomous or cleared diagnostic replacements.
Three cautionary evidence points added to primary care
Primary Care adds three findings: 21 LLMs failing differential diagnosis even when the final answer looked correct, a randomized flu-vaccine nudge that raised willingness in some unvaccinated older adults but lowered it in those already vaccinated, and a home-video fracture-triage tool whose discrimination does not license skipping radiography.
Why Each study shows a measured effect that stops short of supporting autonomous use: diagnostic accuracy, nudge messaging, and imaging triage all still need clinician oversight.
Chapter Primary Care, Family Medicine, and Preventive Medicine
Rare-disease LLM diagnostic-accuracy review
Allergy & Immunology adds a systematic review and meta-analysis of 15 studies (39,529 cases) reporting a pooled Recall at 1 of 43.3% for LLM rare-disease diagnosis, noting that ranking accuracy is not clinical utility.
Why Gives readers a quantitative view of how rare-disease diagnostic-support tools perform in aggregate, ahead of any claim of clinical readiness.
MEGA-CDP guideline-pathway benchmark preprint
Evaluating AI Systems adds a preprint benchmark of guideline-pathway adherence (MEGA-CDP) to its discussion of evaluation pitfalls, illustrating that a high score can leave transportability, workflow effects, clinical benefit, and safety unresolved.
Why Adds a concrete example supporting the chapter's caution that benchmark performance alone does not establish clinical readiness.
Four whole-slide imaging AI studies added to pathology
Pathology adds nnMIL, a slide-level aggregator across about 40,000 whole-slide images, 35 tasks, and four foundation models; PathPrism, encoding colorectal-cancer slides into 628 interpretable spatial features; a rare-diagnosis reader study showing common-condition performance does not extend to rare tumors; and GBAIM, a six-class gastric-biopsy model with high reported sensitivity and specificity on Chinese validation sets.
Why Expands the chapter's whole-slide imaging evidence while keeping each tool's real scope: aggregation, feature extraction, or reader-study performance, not clearance or outcomes claims.
ECG phenotyping for cardiomyopathy and an HFpEF tool
Cardiology adds a hierarchical ECG model that screens for low ejection fraction, hyperkalemia, and hypertrophic cardiomyopathy phenotyping, including septal versus apical subtype, and PARAGUIDE, a phenomapping tool addressing why sacubitril/valsartan showed no overall cardiovascular benefit in a prior heart-failure trial with preserved ejection fraction.
Why Both tools aim to make heterogeneous cardiac populations, hidden ECG phenotypes and HFpEF subgroups, legible enough for more targeted management.
Federated LoRA adaptation study on the privacy page
Privacy & HIPAA adds a study of federated low-rank adaptation for medical LLMs, framing the 2026 result as an information-extraction finding rather than a privacy certification.
Why Clarifies that a federated-learning technique demonstrated for one task does not by itself prove a privacy guarantee.
OCT automation pipeline and a cost-effectiveness analysis
Ophthalmology adds FOCUS, a multicenter 3D OCT pipeline reporting high F1 scores on retrospective imaging, and a 2026 health-system cost-effectiveness analysis identifying a preferred diabetic-eye screening strategy. Neither shows improved vision outcomes or supports an unstaffed clinic.
Why Adds workflow-automation and economic evidence to the chapter's diabetic-eye-screening coverage, each with its own evidence boundary.
Chapter Ophthalmology
Local LLM chart abstraction and registry ambiguity
Research with AI adds a nephrectomy study finding local LLM chart abstraction can agree with a curated research database, alongside a preprint showing aggregate registry-abstraction accuracy can hide failure on harder subsets. Agreement with a database is not a case for retiring human review.
Why Both studies bound how far automated chart abstraction can be trusted before human review is still required.
Chapter Clinical Research with AI
ICU outcome prediction: clinician comparison and a severity model
Critical Care adds a study finding retrospective AI-beats-clinician ICU outcome prediction did not transfer cleanly to the bedside, where subspecialized physicians generally outperformed the model, and Elder-DDB, a geriatric ICU severity model reporting external AUROC of 0.801 to 0.834 and a lower alarm rate than SOFA.
Why Tempers a retrospective AI performance claim with a bedside comparison, while adding an age-specific severity tool that reports fewer false alarms than a standard score.
Infantile fundus screening for retinopathy risk
Pediatrics adds a multicenter open-set system reporting high area-under-curve scores with uncertainty triage for infantile fundus abnormality screening, noting it is not autonomous retinopathy-of-prematurity screening and is not FDA-authorized.
Why Distinguishes a research-stage screening aid from an autonomous or cleared retinopathy-of-prematurity tool.
Chapter Pediatrics and Neonatology
Three new emergency medicine AI studies
Emergency Medicine adds a systematic review finding stroke-imaging AI has the strongest workflow evidence in the specialty but remains product-specific, an abstract-only multi-agent model predicting emergency-department-to-ICU delirium, and PHISE, a prehospital injury-severity estimate meant for scenes where imaging is not yet available.
Why Adds evidence and its limits across three emergency-medicine use cases: imaging-based triage, delirium prediction, and prehospital severity scoring.
Chapter Emergency Medicine
Rheumatoid arthritis treatment-selection model
Internal Medicine adds ANSWER, a model for choosing the next biologic or targeted synthetic DMARD in rheumatoid arthritis, an area the chapter notes is still mostly trial and error.
Why Introduces a decision-support approach to a therapy-selection problem the chapter identifies as largely unsupported by structured tools.
EHR-based epilepsy phenotyping
Neurology adds a model that recovers ILAE epilepsy type and seizure type from the clinical note, complementing wearable and EEG seizure detectors that flag events but do not classify epilepsy type.
Why Fills a gap between event detection and the structured phenotyping clinicians need for classification and treatment planning.
Digital CBT for perinatal depression: a pooled analysis
OB/GYN adds a pooled analysis of 16 randomized trials (7,577 participants) comparing digital cognitive behavioral therapy with control for perinatal depressive symptoms.
Why Gives readers pooled trial evidence on a scalable, technology-delivered treatment option for perinatal depression.
Chapter Obstetrics and Gynecology
Imaging staging for hepatic cystic echinococcosis
Infectious Diseases adds a retrospective multicenter reader study applying WHO-IWGE CE1 to CE5 staging to non-contrast CT for hepatic cystic echinococcosis, noting it is not a treatment trial.
Why Adds imaging-based staging evidence for a parasitic disease the chapter had not previously covered in this depth.
OpenAI's Epic FHIR connector for ChatGPT for Healthcare
LLMs in Practice and AI Tools for Physicians add OpenAI's new Epic connector: administrator-configured FHIR/OAuth access, per-clinician sign-in, read-only chart context, and no write-back. The vendor reports physicians rated 99.1% of 4,363 sampled responses safe; the connector is distinct from Stanford's ChatEHR.
Why OpenAI announced Epic integration for ChatGPT for Healthcare, and independent reporting confirms the access is read-only; the safety figure is a vendor evaluation, not independent validation.
Chapters Large Language Models in Clinical Practice · AI Tools Every Physician Should Know
References, links, and wording were also kept current.
August added a new Clinical Trials and AI chapter covering protocol design, patient matching, and endpoint measurement. Oncology added a single-center comparison of ChatGPT with a human molecular tumor board across 20 cases.
ChatGPT versus human tumor board comparison added to oncology
Oncology adds a single-center retrospective comparison of ChatGPT 4.0 with a human molecular tumor board across 20 cases: a median of 3 versus 1 therapeutic recommendations per case, similar information-density scores, moderate run-to-run consistency, and shorter case-processing time, without measuring patient outcomes.
Why Shows that faster, more numerous recommendations do not by themselves demonstrate that the model matched or improved on the board's decisions.
Prospective ED evaluation of a multi-LLM decision support system
Workflow Integration adds a four-week DECIDE-AI stage 1 evaluation of SHAKED, a multi-LLM decision support system, across 1,138 emergency department patients: use among eligible cases fell from 68% to 30% over the study, 99 of 100 sampled outputs were judged clinically appropriate, and the cycle-time difference did not reach statistical significance.
Why Provides prospective, adoption-over-time evidence rather than a single-point accuracy figure for clinical LLM decision support.
New chapter: Clinical Trials and AI
Clinical Trials and AI covers protocol design, patient matching and recruitment, endpoint measurement, trial-outcome prediction (including the CTO benchmark), trials of AI interventions, and reporting standards, while Research with AI treats general research methods.
Why Separates trial-specific AI uses, which need their own evidence and reporting standards, from the general research-methods chapter.
Chapters Clinical Trials and AI · Clinical Research with AI · Evaluating AI Clinical Decision Support Systems · Executive Summary · Quick Reference: All Chapter Summaries (TL;DRs) · Welcome
Documentation burden and burnout evidence added to Clinical Documentation
Cites Sinsky et al. (2016) on physicians spending nearly two hours on EHR and desk work per hour of patient time, and Shanafelt et al. (2016) linking clerical burden and computerized order entry to higher burnout risk.
Why Grounds the chapter's description of the documentation burden in the studies most often cited for it.
Monitoring-index follow-up study added to AI Tools for Physicians
Cites Hravnak et al. (2011), the UPMC step-down unit follow-up study finding that nurse-actioned alerts from a continuous cardiorespiratory monitoring index reduced the frequency and duration of serious instability events, alongside the original 2008 discovery study.
Why Adds the outcome data that follows from the monitoring index's original derivation study.
Embryo-selection trial evidence added to OB/GYN
Cites Illingworth et al. (Nature Medicine, 2024), a 14-clinic, 1,066-patient trial of deep-learning blastocyst selection (iDAScore), and Kieslinger et al. (The Lancet, 2023), the three-arm SelecTIMO trial comparing time-lapse-based and standard embryo selection across 15 Dutch clinics.
Why Gives the chapter's embryo-selection discussion its primary randomized-trial evidence.
Chapter Obstetrics and Gynecology
Early-onset sepsis calculator trial evidence added to Pediatrics
Cites van der Weijden et al. (eClinicalMedicine, 2025), an open-label, two-arm cluster-randomized trial of the EOS (early-onset sepsis) calculator across 10 Dutch hospitals in 1,830 neonates, alongside updated Kaiser Permanente calculator figures covering 204,485 infants.
Why Adds randomized-trial evidence for a calculator the chapter otherwise supports mainly with observational data.
Chapter Pediatrics and Neonatology
JAMA Surgery viewpoint on robotic-surgery credentialing added
Cites a 2026 JAMA Surgery Viewpoint (Ashar & Selber) proposing competency-based, vendor-neutral privileging for robotic surgery in place of platform-specific exposure or case-volume requirements.
Why Gives the credentialing discussion a named proposal from the surgical literature.
References, links, and wording were also kept current.
Evaluating AI Systems organized its guidance around formal reporting standards such as TRIPOD+AI, contrasted the Epic Sepsis Model's reported AUROC with its lower external validation performance, and added MAST, a task-specific medical AI superintelligence benchmark, with guidance on target-trial emulation and digital-twin validation claims. Dermatology organized its evidence by task, and Policy and Governance added ISO/IEC 42001 as a management-system overlay and the Joint Commission's voluntary AI governance certification.
ISO/IEC 42001 AI management-system overlay for health systems
Policy & Governance adds ISO/IEC 42001's AI management-system requirements (policies, objectives, documented risk treatment, accountability, review) as an organization-level overlay that coordinates, rather than replaces, use-specific clinical controls, alongside NIST's voluntary AI Risk Management Framework functions (GOVERN, MAP, MEASURE, MANAGE).
Why Health systems need organization-level structure for policies and accountability across AI use cases, distinct from and coordinating with controls for each individual clinical workflow.
FDA denies AI radiology device premarket exemption petition
Policy & Governance adds FDA's denial of Harrison.ai's citizen petition requesting partial premarket-review exemption for AI radiology devices, alongside FDA's caution that MAUDE adverse-event reports cannot establish event rates or causation.
Why Documents how FDA is handling a live industry push to loosen premarket review for AI radiology devices, while flagging the reporting-database limitations regulators themselves acknowledge.
FDA selects participants for TEMPO digital health pilot
Policy & Governance updates its account of FDA's TEMPO digital health pilot to note that participants have been selected, ahead of the pilot's pairing with CMS's Advancing Chronic Care with Effective, Scalable Solutions (ACCESS) model.
Why Marks a concrete step forward for the December 2025 pilot the chapter already covers as a mechanism linking device evidence generation to Medicare payment.
State and federal AI regulatory sandbox tracking; EU AI Act timeline
Policy & Governance adds tracking of state AI regulatory sandbox activity (Delaware's framework-development directive, Arizona's non-AI-specific operational sandbox, Wyoming's proposed program, Utah's authorized Doctronic prescription-renewal pilot), the federal SANDBOX Act (S.2750) directing OSTP to create a federal sandbox, and confirms EU AI Act high-risk obligations for qualifying medical devices apply from August 2028.
Why Gives readers a state-by-state and federal view of where AI-specific regulatory flexibility exists, and a firm compliance date for EU medical-device AI obligations.
Swarm learning for multicenter surgical-video disease staging
Surgery & Anesthesiology adds a 2026 NEJM AI appendectomy cohort using swarm learning to train patient-level surgical-video disease-staging models across institutions without centralizing video data, extending the chapter's earlier laparoscopic cholecystectomy video-analysis example.
Why Shows a path for multi-institution surgical-video model training that avoids centralizing patient video data, a practical barrier for this kind of research.
Chapters Surgery, Anesthesiology, and Perioperative Care · Quick Reference: All Chapter Summaries (TL;DRs)
Coronary fat attenuation index: prognostic validation in 40,091 patients
Cardiology adds the ORFAN multicentre cohort of 40,091 patients undergoing clinically indicated CCTA, in which an FAI-based AI-Risk model using pericoronary fat attenuation was externally validated for cardiovascular risk stratification regardless of obstructive coronary artery disease status; the observational study did not test whether FAI-guided care reduces events.
Why Gives readers a large external validation of an imaging-based inflammation biomarker while keeping the distinction between risk-stratification evidence and proof of clinical benefit explicit.
Observability and LLM-as-judge safeguards for agentic systems
Evaluating AI Systems adds guidance to monitor generative and agentic systems as full clinical workflows, tracking version identifiers, retrieval and tool-failure signals, and clinician override outcomes, and cites evidence that even the highest-performing LLM judge in a global-health evaluation should not be the sole release criterion for high-stakes use.
Why System instructions, retrieval sources, and connected tools change after release, so the chapter treats them as workflow components requiring re-evaluation and auditable version history.
Patient-reported measures for post-deployment AI monitoring
Evaluating AI Systems adds patient-reported outcome and experience measures as a supplement to post-deployment monitoring for patient-facing or diagnostic AI, citing a mapping of candidate domains to encounter-specific and diagnostic-journey reports, while noting many domains still need instrument development and validation.
Why Clinical outcomes and adjudicated safety events do not capture the patient's own experience of an AI-supported diagnosis, so the chapter adds a complementary measurement approach.
Joint Commission launches voluntary AI governance certification
Policy & Governance updates its CHAI and Joint Commission coverage: CHAI released governance playbooks in May 2026, and in June 2026 Joint Commission launched a voluntary Responsible Use of AI in Healthcare certification that evaluates organizational governance and safeguards, not individual AI products.
Why A voluntary healthcare AI certification program that was previously only planned has now launched, giving health systems a concrete governance benchmark.
Dermatology AI by task: lesion referral, equity, patient-facing tools, and monitoring
Dermatology covers lesion referral, equity, patient-facing tools, and monitoring as distinct tasks with different reference standards, adding realistic reader-study and equity evidence and an FDA-authorized lesion device table (DermaSensor, Nevisense).
Why Evidence from curated image studies should not be read as proof that an algorithm improves patient outcomes or replaces clinical assessment, so the chapter separates tasks by reference standard and harm.
Chapters Dermatology · AI Tools by Medical Specialty · Hematology-Oncology and Precision Medicine · Pediatrics and Neonatology
Evaluating simulation and digital-twin validation claims
Research with AI adds a framework for evaluating vendor claims that a computational or AI model matches clinical trial results, requiring a named context of use and a prespecified error tolerance per FDA's 2023 credibility guidance and 2025 draft AI guidance, plus independent verification of each submodel rather than one overall performance claim.
Why Vendors increasingly pitch models claiming to match published trial results, and the chapter previously named pharmacokinetic prediction as unresolved without a framework for testing that specific claim.
Chapter Clinical Research with AI
Corpus-level LLM writing signal in PubMed abstracts
LLMs in Practice adds a corpus-level analysis of more than 15 million PubMed abstracts estimating that at least 13.5% of 2024 abstracts carried an excess-vocabulary signal associated with LLM-assisted writing; the method cannot identify individual authorship or establish misconduct.
Why Gives readers a population-level estimate of LLM-assisted writing in the published literature alongside the chapter's existing AI-disclosure data, with the method's stated limits made explicit.
Statistical rigor for benchmark and audit comparisons
Evaluating AI Systems adds a subsection applying a 2024 statistical framework for language model evaluations, computing standard errors for the chapter's benchmark tables; for the 68-question NOTA test, GPT-4o's 85.3% original accuracy carries a standard error of about 4 points and its 48.5% NOTA-modified accuracy about 6 points.
Why A bare accuracy percentage hides whether a reported gap between systems is a real capability difference or sampling noise from the specific question set, so the chapter adds error bars.
Target-trial emulation for observational comparative-impact studies
Evaluating AI Systems adds target-trial emulation as the protocol discipline for observational studies supporting a causal claim: specifying eligibility, intervention strategies, assignment, time zero, follow-up, outcomes, and analysis plan in advance, while noting emulation makes assumptions inspectable but does not remove residual or unmeasured confounding.
Why The chapter's guidance told readers observational designs must make assumptions and confounding explicit without naming the protocol discipline that makes that possible.
MAST: a task-specific medical AI superintelligence benchmark
Evaluating AI Systems adds MAST, an ARISE-run benchmark operationalizing a proposed framework for evaluating medical AI superintelligence claims by task and deployment context, comparing systems with clinicians on diagnostic and management reasoning, multimodal interpretation, safety, and calibration, with patient-important outcomes as the eventual standard; the chapter notes MAST remains in preview and is not a validated clinical tool.
Why Head-to-head medical AI studies mostly use vignettes, simulations, or retrospective data rather than real patient care, so the chapter gives readers a framework for weighing superintelligence claims against actual evidence.
Epic Sepsis Model: reported versus external validation performance
Evaluating AI Systems contrasts Epic's reported AUROC of 0.76 to 0.83 for its sepsis model with external validation at Michigan Medicine (AUROC 0.63, 33% sensitivity at threshold 6), and adds a multicenter prospective validation of the updated model finding AUROC of 0.82 to 0.92 but PPV of only 13% to 26% and high alert burden.
Why Shows that vendor-reported discrimination does not predict local performance or clinical burden, reinforcing the chapter's case for independent, site-specific validation before deployment.
Evidence gaps in clinical LLM and XAI evaluation literature
The chapter cites a review of 519 health care LLM studies (January 2022 to February 2024) finding only 5.0% used real patient data, 15.8% assessed fairness or bias, and 1.2% assessed calibration, and a review of 62 studies finding SHAP and LIME are the most common explainability methods in clinical decision support.
Why Documents how little of the published clinical LLM literature tests real patient data, fairness, or calibration, informing how much weight readers should give a given study.
Matching clinical AI evidence to the claim
Evaluating AI Systems covers matching evidence to the clinical claim, citing TRIPOD+AI, STARD-AI, and PROBAST+AI for model performance, DECIDE-AI for live clinical evaluation, CONSORT-AI and SPIRIT-AI for trials, GRADE for certainty rating, and the NICE Evidence Standards Framework.
Why Aligns the chapter's evaluation criteria with established reporting and risk-of-bias standards so readers can judge clinical AI evidence by the same criteria journals and regulators use.
CONCERN early warning system trial: the 2026 revised length-of-stay estimate
In 2026 the authors of the CONCERN early warning system cluster-randomized trial (60,893 hospital encounters, 74 clinical units) revised its length-of-stay effect estimate because of a modeling error affecting the originally published analysis.
Why Published trial estimates can change after release, so readers citing the CONCERN trial need its revised length-of-stay result, not the originally published figure.
Antimicrobial stewardship decision-support adoption evidence
Infectious Diseases adds a nationwide Israeli outpatient study of 19,287 UTI cases where prescriptions matching UTI Smart-Set recommendations had fewer bug-drug mismatches and less ciprofloxacin use, and a randomized vignette study finding clinicians can identify and ignore incorrect stewardship advice, with adoption driven by evidence and usability rather than explanation displays.
Why Neither study measured clinical cure or patient outcomes, so the chapter frames these as workflow and adoption evidence rather than proof of improved prescribing outcomes.
Resistance prediction models: transportability limits
Infectious Diseases adds validation evidence for automated resistance prediction: MALDI-TOF models trained on German and Swiss spectra performed poorly with location and time shifts and declined over 18 months of prospective use, and a carbapenemase-classification tool (MALCA) was validated at a single center's workflow, with the authors calling for multi-site validation.
Why Resistance-prediction accuracy does not transfer across sites or time without retraining, so the chapter flags local and temporal validation as a prerequisite for clinical use.
Antibiogo AST reading tool: validation and deployment
Infectious Diseases adds Antibiogo, an offline Android tool for reading disk-diffusion antimicrobial susceptibility tests, whose peer-reviewed validation assessed agreement with automated and manual reading rather than patient outcomes; MSF reports routine deployment in its laboratories with IVDD-CE marking for the EUCAST configuration.
Why Gives readers the validation basis and reported field-deployment scale for a low-resource AST reading tool, distinct from its technical accuracy claims.
GPT-4 and GPT-5.1 reasoning errors on oncology notes
Evaluating AI Systems cites a study of two retrospective oncology cohorts (breast, pancreatic, prostate cancer) finding GPT-4 produced reasoning errors in 23.1% of note interpretations, most often confirmation bias, anchoring bias, and omission; GPT-5.1 lowered the error rate but kept similar failure patterns, and automated evaluators detected errors without reliably classifying subtypes.
Why Supports clinician-reviewed error taxonomies on representative clinical notes rather than relying on endpoint accuracy alone to judge LLM reasoning safety in oncology.
References, links, and wording were also kept current.
The FDA denied Harrison.ai's petition to exempt radiology AI from premarket review. A Nature Medicine study found frontier LLMs including GPT-5.2, Gemini 3.1 Pro, and Claude outperformed specialized clinical AI tools, and Surgery and Anesthesiology added a new section evaluating robotic surgery evidence by platform, surgeon, and AI layer. Other additions included wet-lab-validated hypotheses from Co-Scientist and a study of DeepSeek's deployment across 90 Chinese tertiary hospitals.
DeepSeek deployment in Chinese tertiary hospitals
LLMs in Practice cites a 2025 Journal of Medical Systems study documenting DeepSeek deployment across 90 Chinese tertiary hospitals.
Why Gives readers a primary source documenting real-world DeepSeek deployment in hospital settings.
ProJenX and Unlearn AI digital twin comparator in ALS trial
Neurology adds a note on ProJenX's PRO-101 Phase 1 trial (ClinicalTrials.gov NCT05279755), where ProJenX and Unlearn plan to use an AI-generated ALS Digital Twin Generator to forecast ALSFRS-R, slow vital capacity, and plasma neurofilament light trajectories for an open-label extension, as trial-design support rather than proof of drug efficacy.
Why Digital twins are entering ALS trial design as comparators, and the chapter flags what clinicians should check, context of use, source population, uncertainty, before treating them as efficacy evidence.
Rehabilitation robotics evidence in stroke and spinal cord injury
Orthopedics adds review-level evidence on rehabilitation robotics: a 2025 Cochrane update of 101 trials found robotic gait devices modestly increase independent walking after stroke without improving walking speed or distance, the RATULS trial found robot-assisted arm training no better than enhanced therapy, and a spinal cord injury meta-analysis found mixed gains. Adds the FDA exoskeleton clearance K143690.
Why The chapter's rehabilitation robotics evidence needed review-level sourcing rather than device promotion, so clinicians can weigh dose- and comparator-dependent functional gains against marketing claims.
Surgical robotics and OR analytics companies: evidence status
Surgery & Anesthesiology adds a subsection distinguishing FDA-authorized teleoperated platforms (da Vinci, Versius, Hugo) from OR video analytics vendors (OR Black Box, Touch Surgery, Theator, Caresyntax) and autonomy-first startups such as Aleph Surgical, noting Aleph's research preview has no peer-reviewed publication or FDA authorization as of June 2026.
Why Distinguishes regulated clinical platforms from data-capture and analytics tools and unproven autonomy startups so readers do not conflate commercial visibility with clinical evidence.
Robotic surgery and surgical AI evaluation
Surgery & Anesthesiology adds a section separating platform, surgeon, procedure, and AI-layer evidence for robotic surgery: FDA clearances for da Vinci 5, Versius, and Hugo RAS; a randomized trial and cohort studies on outcomes; autonomous ex vivo surgery research; and evaluation guidance for surgical video AI and skill assessment.
Why The chapter previously lacked dedicated coverage separating robotic-platform teleoperation from AI capability claims, evidence that varies by procedure and is often industry-influenced.
Chapters Surgery, Anesthesiology, and Perioperative Care · Surgical Subspecialties · Quick Reference: All Chapter Summaries (TL;DRs)
Frontier LLMs outperform specialized clinical AI tools in benchmark
Evaluating AI Systems and AI Tools for Physicians cite a 2026 Nature Medicine study in which frontier LLMs (GPT-5.2, Gemini 3.1 Pro, Claude Opus 4.6) outperformed the specialized tools OpenEvidence and UpToDate Expert AI across MedQA, HealthBench, and blinded clinician-reviewed real clinical queries, with the two specialized tools performing similarly to each other and to Google AI Overview.
Why Shows procurement teams that a clinical-focused brand or subscription price is not itself evidence of superior performance over general-purpose frontier models.
Chapters Evaluating AI Clinical Decision Support Systems · AI Tools Every Physician Should Know
FDA Clinical Decision Support final guidance (January 2026)
Executive Summary, Policy & Governance, and Radiology now cite FDA's January 29, 2026 final Clinical Decision Support guidance directly, replacing secondary commentary links, and clarify that software meeting the device definition, including patient- or caregiver-facing functions, remains subject to FDA's digital health policies even when a physician is in the loop.
Why Gives readers FDA's own guidance document and FAQ, and clarifies that clinician oversight alone does not exempt a tool from device status.
Chapters Executive Summary · Healthcare Policy and AI Governance · Diagnostic Imaging, Radiology, and Nuclear Medicine
Autonomous clinical AI licensure-like oversight proposal
Policy & Governance adds a section citing a JAMA Perspective proposing licensure-like oversight, distinct from ordinary clinical decision support or device clearance, for AI systems that make care determinations without per-case clinician review.
Why Gives health systems a framework for governing autonomous clinical AI as a credentialing question rather than a procurement or device-clearance question.
Chapters Healthcare Policy and AI Governance · Clinical AI Policy Templates
FDA denies Harrison.ai petition to exempt radiology AI from premarket review
Policy & Governance cites the primary petition and FDA's interim response: Harrison.ai's October 2025 citizen petition sought partial 510(k) exemption for certain radiology computer-aided detection, diagnosis, and triage devices, and FDA's April 1, 2026 interim response denied it, citing safety and oversight concerns while noting that predetermined change control plans can support post-clearance updates.
Why Distinguishes the requested radiology AI exemption from FDA’s interim response and the role of post-clearance change controls.
Multi-agent biomedical discovery systems: wet-lab validation evidence
Emerging Technologies and Research with AI cite peer-reviewed reports that Co-Scientist generated hypotheses validated in vitro for acute myeloid leukemia repurposing candidates, and that Robin's lab-in-the-loop workflow identified and validated candidate therapeutics for dry age-related macular degeneration, while noting neither removes the need for experimental or clinical validation.
Why Distinguishes genuine peer-reviewed, lab-validated multi-agent discovery results from broader claims about autonomous AI research.
Chapters Emerging AI Technologies in Healthcare · Clinical Research with AI
Ambient scribe time-savings claims calibrated
Clinical Documentation and LLMs in Practice cite peer-reviewed evidence that ambient AI scribe time savings are implementation-dependent rather than a default large reduction, and reiterate that every AI-generated note still requires clinician review before signing.
Why Grounds time-savings expectations for ambient documentation tools in peer-reviewed practice and trial evidence rather than vendor promises.
Chapters AI-Assisted Clinical Documentation · Large Language Models in Clinical Practice
References, links, and wording were also kept current.
May added a Science study comparing an OpenAI model's clinical reasoning with physicians, and a review finding few randomized trials among thousands of clinical LLM studies. A broad evidence refresh added trial results and foundation-model evidence across nearly every specialty chapter, plus a new section on embodied AI caregiving robots. AMIE was extended to multimodal conversation.
FDA Quality Management System Regulation inspection changes
Policy & Governance adds that the FDA's Quality Management System Regulation, effective February 2, 2026, replaces the legacy Quality System Regulation (21 CFR Part 820) with requirements aligned to ISO 13485:2016, and that FDA now inspects devices under Compliance Program 7382.850 instead of the prior inspection technique.
Why Gives device makers, including AI device makers, the current FDA quality system and inspection framework they must follow.
FDA Clinical Decision Support guidance referenced in evaluation chapter
Evaluating AI Systems notes FDA guidance clarifying which clinical decision support software functions are excluded from the device definition, and states that meeting the Non-Device CDS criteria is not itself an evaluation standard; local validation and post-deployment monitoring remain standard practice.
Why Clarifies that FDA's non-device classification for CDS software does not substitute for a site's own validation and monitoring.
NIST AI 800-4 monitoring taxonomy cited
Evaluating AI Systems cites NIST's CAISI proposal of six post-deployment monitoring categories, covering functionality, operational, human factors, security, compliance, and large-scale impacts, to structure monitoring beyond ad hoc dashboards.
Why Gives readers a federal reference framework for organizing AI monitoring practices after deployment.
Ambient scribe platforms: what the randomized trials show
Clinical Documentation compares ambient scribe platforms by trial evidence. One randomized trial showed no difference in time-in-note against usual care; another showed lower note time and work exhaustion. Accuracy and vendor-specific metrics for other platforms vary by setting and warrant local verification.
Why Aligns vendor comparisons with what randomized evidence actually supports and steers readers toward local pilot verification.
Stroke AI evidence boundary in radiology
Radiology & Nuclear Medicine cites a 2025 systematic review of 380 deep-learning studies in acute ischemic stroke, finding 45.0% focused on lesion segmentation, 33.9% on classification or triage, 8.2% on outcome prediction, and 3.9% on generative AI or large language model applications.
Why Gives readers a primary source for where stroke imaging AI evidence currently concentrates.
CONCERN early warning system trial in internal medicine
Internal Medicine cites a pragmatic cluster-randomized trial of the CONCERN nursing-surveillance early warning system across 74 clinical units and 60,893 hospital encounters, finding lower mortality risk (adjusted HR 0.64) and shorter length of stay (adjusted incidence rate ratio 0.91) in intervention encounters.
Why Adds outcome-level trial evidence for a deterioration early warning system built on nursing documentation rather than vitals and labs alone.
Real-time AI insulin dosing trial in internal medicine
Internal Medicine cites a 2025 multicenter randomized trial testing a real-time AI insulin clinical decision support system against senior endocrinologist management.
Why Adds randomized trial evidence for AI-assisted inpatient insulin dosing decisions.
Perioperative hypotension prediction trial in surgery
Surgery & Anesthesiology cites the HYPE-2 randomized trial of a machine-learning Hypotension Prediction Index during elective on-pump cardiac surgery, which reduced the median time-weighted average MAP below 65 mm Hg by 63% among 130 patients.
Why Adds randomized evidence for an intraoperative hypotension prediction tool.
Neonatal parenteral nutrition AI system in pediatrics
Pediatrics adds a Neonatal Parenteral Nutrition section describing TPN2.0, a data-driven neonatal nutrition recommendation system trained on 79,790 orders from 5,913 Stanford patients and externally validated on 63,273 orders.
Why Gives readers evidence on an AI system addressing complex, safety-critical neonatal nutrition orders.
Chapter Pediatrics and Neonatology
Gynecologic ultrasound AI review in OB/GYN
OB/GYN cites a systematic review of 59 studies on AI applied to ultrasound imaging for benign gynecologic disorders, including PCOS, infertility, and endometriosis, most rated at high risk of bias in subject selection or index-test domains.
Why Adds a primary evidence assessment of AI ultrasound tools across benign gynecologic conditions.
Chapter Obstetrics and Gynecology
Prospective ED triage AI review in emergency medicine
Emergency Medicine cites a systematic review that found only seven prospective ED triage AI studies among 1,633 screened records, with reported accuracy ranging from 80.5% to 99.1% across differing triage systems and outcome measures.
Why Shows prospective evidence for AI triage tools remains scarce and supports AI as a triage aid, not autonomous triage.
Chapter Emergency Medicine
ICU AI operationalization evidence gap in critical care
Critical Care adds an Operationalization Evidence Gap section citing a systematic review of 1,263 ICU AI studies, finding 74% at technology-readiness level 4 or below, 24% externally validated, and only 2% reaching clinical integration.
Why Documents how far most ICU AI research remains from clinical deployment.
GI pathology foundation model evidence in oncology
Oncology cites a multinational cohort study that trained a pathology foundation model on more than 130 million patches from 104,876 whole-slide images, then fine-tuned it to predict survival and adjuvant chemotherapy benefit in gastric, esophageal, and colorectal cancers.
Why Adds evidence for foundation-model approaches to prognosis and treatment-benefit prediction in GI cancers.
ECG foundation model update in cardiology
Cardiology adds an ECG Foundation Models section describing ECGFounder, a model built on 10,771,552 cardiologist-annotated ECGs and evaluated across multiple domains and lead configurations.
Why Notes that large ECG foundation models are emerging, though not yet positioned to replace ECG interpretation.
GOLDEN BRIDGE II stroke CDSS trial in neurology
Neurology cites the GOLDEN BRIDGE II cluster randomized trial of a stroke clinical decision support system across 77 Chinese hospitals and 21,603 patients with acute ischemic stroke, combining AI-assisted imaging analysis, cause classification, and treatment recommendations.
Why Adds a large outcome trial evaluating an AI-assisted stroke decision support system.
Youth mental health chatbot meta-analysis in psychiatry
Psychiatry cites a meta-analysis of 15 randomized trials in 1,974 participants ages 12 to 25, finding a moderate-to-large effect on depressive symptoms (Hedges g=0.61) but no effect on anxiety, stress, affect, or well-being.
Why Clarifies which mental health outcomes chatbot interventions currently show evidence for in adolescents and young adults.
Chapter Psychiatry and Behavioral Health
AMIE diagnostic dialogue trial in primary care
Primary Care cites a randomized, double-blind crossover study of 159 clinical scenarios in which AMIE, a diagnostic-dialogue AI system, outperformed primary care physicians on most prespecified axes of history taking, diagnostic accuracy, management, communication, and empathy in a text-based setting.
Why Adds controlled trial evidence comparing a conversational diagnostic AI system with physicians.
Chapter Primary Care, Family Medicine, and Preventive Medicine
Lung biomarker foundation model evidence in pathology
Pathology cites a real-world deployment study that fine-tuned an open-source pathology foundation model to predict EGFR mutation status from H&E lung adenocarcinoma slides across 8,461 slides, reaching AUC 0.847 internally, 0.870 externally, and 0.890 in a prospective silent trial.
Why Adds prospective, real-world evidence for an AI biomarker-prediction workflow in lung pathology.
Total-body imaging trial caution in dermatology
Dermatology cites a randomized trial in high-risk melanoma patients in which adding 3D total-body photography and sequential digital dermoscopy increased excisions per person (5.73 vs. 3.99) but detected fewer melanomas (24 vs. 43) than usual care.
Why Tempers expectations for total-body imaging surveillance with a trial showing more excisions but fewer melanomas detected.
Chapter Dermatology
Retinal angiography reporting AI evidence in ophthalmology
Ophthalmology cites InterpreFFA, a diagnosis-supervised contrastive learning tool for fundus fluorescein angiography interpretation, which helped two residents improve diagnostic accuracy from 85.55% to 90.34% and shortened report time from 153.93 to 108.08 seconds in a simulated setting.
Why Adds evidence that AI-generated draft reporting can improve trainee diagnostic accuracy and speed on angiography interpretation.
Chapter Ophthalmology
ChatGPT-assisted consent trial in orthopedics
Orthopedics cites a single-blind pilot RCT of 60 total knee arthroplasty patients randomized to ChatGPT-assisted versus traditional surgeon-led informed consent, finding lower post-consent anxiety and higher satisfaction, with similar depression, pain, and knee function scores between groups.
Why Adds pilot trial evidence on chatbot-assisted informed consent for a common orthopedic procedure.
AI antimicrobial stewardship review in infectious diseases
Infectious Diseases cites a systematic review of AI-driven antimicrobial stewardship applications, identifying empirical antibiotic selection, resistance prediction, dose optimization, and vancomycin dosing support as the most common use cases, with most evidence still observational.
Why Surveys where AI antimicrobial stewardship tools currently show implementation evidence versus remaining observational.
Childhood asthma trajectory AI review in allergy and immunology
Allergy & Immunology cites a systematic review and meta-analysis of 89 machine-learning trajectory studies identifying recurring childhood asthma, wheeze, and eczema patterns, most rated low methodological quality, particularly for modeling and reporting.
Why Adds a primary evidence assessment of machine-learning approaches to childhood allergic disease trajectories.
Colonoscopy CADe living guideline update in surgical subspecialties
Surgical Subspecialties adds a Living Guideline Update section citing a 2025 BMJ living guideline on computer-aided detection for colonoscopy polyps, drawing on 44 RCTs and more than 30,000 participants, finding low-certainty evidence that CADe may increase positive findings but no direct evidence on cancer incidence.
Why Gives readers the current guideline-level evidence status for AI-assisted colonoscopy detection.
Chapter Surgical Subspecialties
Outcome trial examples added to AI evaluation framework
Evaluating AI Systems adds Level 5 outcome-trial examples, citing the CONCERN early warning system's mortality and length-of-stay results and the GOLDEN BRIDGE II stroke decision support trial as early instances of outcome-level evidence.
Why Shows readers what outcome-level evidence looks like in practice, since the evidence base at this level remains thin.
Human-LLM collaboration meta-analysis in workflow integration
Workflow Integration cites a 2026 systematic review and meta-analysis of human-LLM collaboration across 10 peer-reviewed clinical studies, finding directionally favorable but heterogeneous effects and only 2 of 9 randomized trials at low risk of bias.
Why Adds a primary source quantifying how consistent human-AI workflow benefits are across clinical studies.
Ambient scribe RCT evidence added to documentation guidance
Clinical Documentation updates its ambient scribe comparison table, citing a randomized trial finding lower note time and work exhaustion with a patient-facing summary tool, and notes that time savings across products are modest to moderate rather than uniformly 1 to 2 hours daily.
Why Tempers vendor time-savings claims with randomized trial evidence on realistic effect sizes.
Evidence-generation speed gap added to AI research guidance
Research with AI notes that PubMed adds more than 5,000 articles daily while systematic reviews take 12 to 24 months, and that AI tools can accelerate literature screening and trial recruitment by 20% to 40% and generate real-world evidence from EHR data in days.
Why Explains why AI-accelerated evidence generation creates both an opportunity and a scrutiny challenge for clinical research.
Chapter Clinical Research with AI
AI device recall evidence added to vendor evaluation checklist
Vendor Evaluation cites a 2025 cross-sectional study finding that recalls of FDA-cleared AI-enabled medical devices were uncommon but concentrated early after clearance and predominantly involved products lacking clinical validation.
Why Reminds readers that FDA clearance alone does not end evaluation of a vendor's device.
Chapter AI Vendor Evaluation Framework
Embodied AI caregiving robots: new section
Emerging Technologies adds a section on humanoid and socially assistive caregiving robots, citing U.S. workforce projections for home health and personal care aides and a 2026 JMIR Aging systematic scoping review of 59 studies, concluding evidence remains concentrated in short, small studies of social interaction, cognitive support, and monitoring rather than autonomous hands-on care.
Why Distinguishes measurable workforce need for caregiving support from the current, limited evidence base for humanoid caregiving robots.
FDA Predetermined Change Control Plan guidance cited
Evaluating AI Systems, Policy & Governance, and Medical Education now cite the FDA's Predetermined Change Control Plan guidance, finalized in December 2024 and reissued in August 2025, and its January 2025 draft guidance on AI-enabled device software lifecycle management, and update the FDA device count to 1,430 listed devices, 96.2% by 510(k) submission prefix.
Why Reflects the FDA's finalized PCCP framework and current device list, giving readers the primary regulatory sources for adaptive AI oversight.
Chapters AI in Medical Education · Healthcare Policy and AI Governance · Evaluating AI Clinical Decision Support Systems
ER-Reason benchmark: LLM reasoning across ED workflow stages
Emergency Medicine and LLMs in Practice add the ER-Reason preprint, a benchmark using 3,437 patients across 3,984 ED encounters and 72 physician-authored rationales to test LLM performance on triage, EHR review, treatment planning, diagnosis, and disposition; o3-mini reached 34.40% exact ICD-10 match accuracy for final diagnosis.
Why A more workflow-realistic benchmark shows reasoning models still misjudge acuity and disposition, reinforcing that ED triage requires local calibration and clinician control.
Chapters Large Language Models in Clinical Practice · Emergency Medicine
FDA AI-enabled device list update: 1,430 devices, 96.2% via 510(k)
Updates counts drawn from FDA's AI-enabled medical device list to 1,430 devices total, with radiology accounting for 1,094 (76.5%); the 510(k) share (96.2%) is now computed directly from the list's own submission-number prefixes.
Why Reflects FDA's periodic update to the device list and ties the regulatory-pathway statistic to the list itself rather than a secondary analysis.
Chapters AI Tools by Medical Specialty · Quick Reference: All Chapter Summaries (TL;DRs) · Executive Summary · Healthcare Policy and AI Governance · AI Tools Every Physician Should Know · Diagnostic Imaging, Radiology, and Nuclear Medicine
FDA device lifecycle gaps and LLM regulatory pathway proposals
Policy & Governance adds a 2026 analysis finding 44.87% of 956 FDA-approved radiology AI devices underwent software updates without a closed-loop risk management system, plus two interviews on a combined oversight and reimbursement pathway and on premarket review logic for LLM-enabled devices.
Why Documents a lifecycle oversight gap in approved AI devices and proposed regulatory pathways to address it.
AMIE extended to multimodal diagnostic conversation
Emerging Technologies adds a Nature Medicine study extending AMIE from text-only dialogue to multimodal patient conversations with images, documents, and ECG data; in 105 simulated telehealth consultations, specialists rated it higher than primary care physicians on diagnostic accuracy.
Why Documents an exploratory evaluation, not a randomized trial, of multimodal AI diagnostic dialogue.
50-country survey finds a physician AI adoption gap
Global Health Equity adds a 2026 survey of 1,049 physicians across 50 countries finding high AI awareness (86.5%) but low use (27.8%), with formal training associated with more than threefold higher odds of use and institutional AI availability with more than eightfold higher odds.
Why Shows that AI awareness does not translate into use without training and institutional deployment capacity.
Chapter AI and Global Health Equity
Scoping review yields a medical education AI competency taxonomy
Medical Education adds a 2026 scoping review synthesizing 54 studies from 22 countries into seven domains, 37 competencies, and 170 learning objectives for undergraduate AI education.
Why Gives medical educators a structured competency framework drawn from the current literature.
Chapter AI in Medical Education
Research agenda for AI interpreter services in language access
Ethics & Health Equity adds a 2026 research agenda for AI interpreter services warning that translation accuracy alone is insufficient, and calling for evaluation of patient understanding, trust, and escalation to human interpreters for high-risk encounters.
Why Frames language access as an AI equity requirement with its own evaluation standards.
PRIMA-AI trial finds no benefit from an EHR graft-loss risk model
Evaluating AI Systems adds the PRIMA-AI randomized trial of a 1-year graft-loss risk model embedded in the EHR for kidney transplant recipients, which found no increase in patient-reported treatment conversations (39% vs 40%) and low, variable clinician uptake.
Why Shows that placing an AI prediction in the EHR is not itself a clinical intervention without workflow support.
Stanford deployment reports error rates for AI hospital-course summaries
Clinical Documentation adds a prospective Stanford deployment of an LLM-based agentic workflow generating hospital-course summaries, in which 100 physician reviews found omissions in 25%, inaccuracies in 20%, unsupported content in 2%, and no incorrect citations.
Why Gives readers a physician-reviewed error profile for an AI summarization tool before it reaches discharge documentation.
ECG-based heart failure detection validated across three cohorts
Cardiology adds an open-source ECG-based heart failure detection model validated against NT-proBNP-refined labels across a 25,300-patient development cohort, a 43,727-patient test cohort, and 161,352 MIMIC-IV patients, achieving AUC 0.86 to 0.96.
Why Adds a large, externally validated, open-source model for a screening use case.
Imaging AI compared against physicians on glioma IDH prediction
Neurology adds a study comparing two MRI-based AI models with 18 physicians on IDH mutation prediction in glioma; GliomaVista-IDH scored AUC 0.97 internally but fell to 0.82 on external validation, while physicians reached AUC 0.88 with better calibration.
Why Shows external validation and calibration gaps that matter before imaging AI guides treatment planning.
Meta-analysis compares autonomous AI and store-and-forward DR screening
Ophthalmology adds a 2026 meta-analysis of 28 studies finding autonomous AI has higher pooled sensitivity than store-and-forward pathways for diabetic retinopathy and macular edema, while noting pathway choice depends on prevalence, referral capacity, and the harm of missed disease.
Why Helps readers weigh screening pathway tradeoffs rather than treating autonomous AI as simply superior.
Chapter Ophthalmology
Large observational study of digital front-door eConsult use
Primary Care adds a BMJ Digital Health & AI study of 43,657,891 eConsult submissions across Great Britain from 2019 to 2023, finding 89.0% routed to the patient's practice and 7.7% to urgent or emergency care, with use varying by region and deprivation.
Why Frames digital front-door triage as a workload, safety, and equity question for AI-assisted triage tools.
Chapter Primary Care, Family Medicine, and Preventive Medicine
ONC HTI-1 algorithm transparency requirements
Vendor Evaluation adds a section on ONC's HTI-1 final rule, which requires algorithm transparency documentation for AI embedded in certified health IT covering more than 96% of hospitals and 78% of office-based physicians.
Why Gives readers a regulatory checklist item for vendor due diligence on certified health IT.
Chapter AI Vendor Evaluation Framework
Randomized trial of AI-assisted cervical cytopathology
OB/GYN adds a 2026 multicentre randomised crossover trial (n=1,920) in which deep learning improved non-expert cytopathologist sensitivity from 71.3% to 85.7%, with comparable specificity and reading time cut from 175 to 31 seconds per slide.
Why Adds rare randomized trial-level evidence for AI-assisted cervical cytopathology.
Chapter Obstetrics and Gynecology
Systematic review finds few RCTs among clinical LLM studies
Evaluating AI Systems and LLMs in Practice add a 2026 systematic review finding 4,609 peer-reviewed clinical LLM studies from January 2022 through September 2025, but only 19 prospective randomized trials, with most testing simulated scenarios rather than patient outcomes.
Why Gives readers field-scale evidence that publication volume for clinical LLMs outpaces rigorous trial evidence.
Chapters Evaluating AI Clinical Decision Support Systems · Large Language Models in Clinical Practice
Physician-AI collaboration framing expanded with Watson Oncology example
Future Perspectives and Primary Care expand the discussion of physician-AI collaboration limits, citing IBM Watson for Oncology's over $4 billion investment before MD Anderson terminated its contract, and noting symptom checkers achieve only 30-60% accuracy.
Why Reinforces that collaboration gains depend on task design and validation, not investment or hype.
Chapters The Physician-AI Partnership: Future Perspectives · Primary Care, Family Medicine, and Preventive Medicine
RecovryAI reports first FDA Breakthrough designation for patient-facing generative AI
Policy & Governance and Orthopedics add RecovryAI's first FDA Breakthrough Device Designation for a patient-facing generative AI device, an LLM-based chatbot for post-joint-replacement recovery that contacts patients twice daily and escalates concerns to the surgical team.
Why Documents the first formal FDA regulatory pathway entry for a patient-facing generative AI medical device.
Chapters Healthcare Policy and AI Governance · Orthopedic Surgery and Physical Medicine
Nature Medicine editorial on evidence standards for AI claims
Industry Reports & Benchmarks adds a 2026 Nature Medicine editorial stating that claims of clinical usefulness or improved outcomes require evidence proportional to the claim, not retrospective performance metrics alone.
Why Gives readers an editorial standard for weighing AI vendor claims against funding and adoption momentum.
Chapter Industry Reports and Benchmarks
Brodeur et al. Science study: LLM clinical reasoning versus physicians
AI Failures, Executive Summary, Evaluating AI Systems, LLMs in Practice, and Emergency Medicine add a Science study in which an OpenAI o1-series model exceeded two attending physicians on blinded emergency department second-opinion diagnosis at three touchpoints (65.8%, 69.6%, and 79.7% correct), with a linked Perspective calling for prospective trials before deployment.
Why Gives readers peer-reviewed evidence on LLM diagnostic reasoning alongside a caution against reading it as deployment-ready.
Chapters AI Failures in Medicine: Lessons Learned · Executive Summary · Evaluating AI Clinical Decision Support Systems · Large Language Models in Clinical Practice · Emergency Medicine
References, links, and wording were also kept current.
April added a new Industry Reports and Benchmarks chapter, and expanded oncology with the MASAI, DYNAMIC, and NHS-Galleri screening trials plus ASCO's six guiding principles for AI in oncology. Cardiology added ARISE, the first AI-ECG randomized trial associated with a mortality signal, and Safety and Risk added evidence that leading AI models affirmed user actions 49% more often than humans. The FDA launched its TEMPO digital health pilot.
AI Failures adds LLM diagnosis and NOHARM safety case studies
AI Failures & Lessons adds two case studies: 21 LLMs failing to produce appropriate differential diagnoses in over 80% of cases with limited information, and the NOHARM benchmark finding severe harm potential in 22% of LLM clinical recommendations across 31 models.
Why Documents evaluated safety and reasoning gaps in LLMs used for clinical decision support.
FDA lifecycle guidance, safety framework, and AI chatbot hazard listing
Safety & Risk adds the FDA's 2025 draft guidance on AI-enabled device lifecycle management, a Joint Commission and CHAI national safety framework, and ECRI's 2026 listing of AI chatbot misuse as a top health-technology hazard.
Why Gives readers current regulatory and safety-body guidance for evaluating AI vendors.
HIPAA Security Rule proposal addresses AI training data
Privacy & HIPAA adds HHS OCR's January 2025 proposed HIPAA Security Rule update, the first in over a decade, which explicitly addresses AI training data, alongside IBM's reported average healthcare data breach cost of $7.42 million.
Why Gives readers the first HIPAA update proposal to directly address AI training data.
Healthcare AI startup investment concentration study
Industry Reports & Benchmarks adds a 2026 npj Digital Medicine analysis of 3,807 healthcare AI startups founded between 2010 and 2024, adding structural detail to existing funding reports.
Why Gives readers peer-reviewed detail on where healthcare AI startup investment concentrates.
Chapter Industry Reports and Benchmarks
Cardiology adds ARISE, TAILORED-AF, EQUAL, and ADMINISTER trials
Cardiology adds the ARISE trial, the first AI-ECG randomized trial associated with a mortality signal, TAILORED-AF (AI-guided catheter ablation), the EQUAL trial, the ADMINISTER trial on digital consults and GDMT, the AHA science advisory on AI evaluation, and a new heart failure AI section.
Why Brings the chapter's evidence current with recent randomized trials and a society advisory.
Radiology's FDA device count and validation-gap evidence
Radiology reports 1,094 FDA-listed devices, about 76% of listed medical AI, and adds systematic reviews on the commercial MRI AI evidence gap, FDA-authorized radiology AI/ML devices, and AI device recalls.
Why Documents how mature radiology AI deployment is and where independent validation is still missing.
Neurology adds adaptive DBS, BrainGate, and MS imaging evidence
Neurology adds the first FDA-approved closed-loop adaptive deep brain stimulation system, the BrainGate speech neuroprosthesis (97.5% accuracy at about 32 words per minute over 8 months in ALS), and new multiple sclerosis imaging evidence.
Why Adds primary evidence for newly cleared neurotechnology and updated MS imaging tools.
Ring-form wearables: pre-symptomatic detection and cycle tracking
Emerging Technologies and OB/GYN add coverage of ring-form wearables such as the Oura Ring, including pre-symptomatic COVID-19 detection about 2.75 days before testing and continuous temperature-based menstrual cycle tracking.
Why Adds evidence for a consumer wearable category physicians increasingly encounter.
Chapters Emerging AI Technologies in Healthcare · Obstetrics and Gynecology
Critical care, emergency, internal medicine, and pediatrics trial evidence
Critical Care reports the ACTiVE trial (1,201 patients; closed-loop ventilation showed no ventilator-free-days benefit over conventional care); Internal Medicine adds the first FDA-authorized AI sepsis diagnostic (Prenosis Sepsis ImmunoScore, April 2024); Pediatrics adds a cluster-randomized trial of the EOS Calculator across 19 Netherlands hospitals.
Why Reports completed trial results and a regulatory first relevant to inpatient and neonatal care.
Chapters Critical Care and Pulmonary Medicine · Emergency Medicine · Internal Medicine and Hospital Medicine · Pediatrics and Neonatology
LLMs in Practice adds care-transition chatbot and benchmark evidence
LLMs in Practice adds a chatbot for care transitions that reduced specialist consultation duration by 28.7% across 2,069 patients, an RCT on LLM-clinician collaboration, and the MedHELM benchmark reference.
Why Adds trial evidence on LLM-assisted care transitions and clinician collaboration.
Trial finds no time savings for a leading ambient scribe
Clinical Documentation notes that a randomized trial found time-in-note did not meaningfully change for Nuance DAX Copilot, alongside the vendor-reported figure of 50%.
Why Gives readers independent trial evidence alongside a vendor's reported claim.
Multi-agent oncology systems and cardiac imaging evidence
AI Fundamentals for Clinicians and Emerging Technologies cite a 2025 Nature Cancer study on multi-agent oncology systems, and Emerging Technologies adds a 100-patient prospective study and FFR-CT diagnostic accuracy evidence.
Why Grounds oncology concordance data in a peer-reviewed study and adds imaging evidence.
Chapters AI Fundamentals for Clinicians · Emerging AI Technologies in Healthcare
First RCT of an LLM-assisted maternal health chatbot
Global Health Equity adds the first randomized controlled trial of an LLM-assisted maternal health chatbot and Anthropic's reported survey of 80,508 AI users across 159 countries.
Why Adds trial-level evidence for LLM-based maternal health support in low-resource settings.
Chapter AI and Global Health Equity
Deepfake detection lag and misinformation amplification evidence
Medical Misinformation adds a 2026 Radiology study on deepfake detection lagging creation, and evidence on algorithmic amplification of misinformation and vaccine hesitancy's toll during COVID-19, alongside Meta's January 2025 discontinuation of third-party fact-check labels.
Why Adds primary evidence on how medical misinformation spreads and is detected.
Chapter Medical Misinformation and AI
AMA and AAMC AI literacy policy and physician training study evidence
Medical Education adds the AMA's November 2025 policy on AI literacy, AAMC's 2025 principles for AI use, and a study finding a 1.5-hour training improved AI use among 326 physicians across 39 countries.
Why Gives readers current professional society guidance and evidence on structured AI training.
Chapter AI in Medical Education
HealthBench Professional: a clinician-specific AI evaluation
Evaluating AI Systems adds OpenAI's HealthBench Professional, a clinician-specific extension of the HealthBench evaluation framework, and a new discussion of evaluation integrity.
Why Extends the chapter's coverage of AI evaluation benchmarks to a clinician-focused tool.
OpenEvidence reports 1 million consultations in a single day
AI Tools for Physicians notes OpenEvidence's reported milestone of 1 million clinical consultations between verified physicians and its AI system in a single day, and updates the FDA-authorized AI device count to over 1,450.
Why Gives readers a company-reported adoption figure alongside the current FDA device count.
OB/GYN adds a section on AI in assisted reproduction and IVF
OB/GYN adds a new section on AI in assisted reproduction and IVF and expands its discussion of maternal health disparities.
Why Fills a gap in the chapter's coverage of reproductive technology.
Chapter Obstetrics and Gynecology
Script concordance testing as a clinical-reasoning benchmark
Evaluating AI Systems adds script concordance testing as a validated benchmark for clinical reasoning.
Why Gives readers an additional validated method for evaluating clinical reasoning in AI systems.
Ethics adds dataset transparency and AI disclosure guidance
Ethics & Health Equity adds the STANDING Together international consensus on dataset transparency and a section on professional society guidance for AI disclosure to patients.
Why Gives readers current guidance on dataset transparency and disclosure standards.
DermaSensor: FDA-authorized AI skin cancer device
Oncology adds coverage of DermaSensor, the first AI-enabled skin cancer detection device authorized for primary care (FDA De Novo, DEN230008, January 2024), an elastic scattering spectroscopy device reporting 95.5% overall sensitivity (87.5% for melanoma, 97.8% for basal cell carcinoma, 98.7% for squamous cell carcinoma) and halved missed cancers in a 108-clinician primary care study.
Why Extends AI oncology coverage into primary care skin cancer triage.
Oncology screening trial evidence: MASAI, DYNAMIC, and NHS-Galleri
Oncology adds the MASAI randomized trial of AI-assisted mammography screening (Lancet Digital Health, 2025), the DYNAMIC trial, the first randomized evidence for ctDNA-guided adjuvant treatment decisions (Tie et al., 2025), and the 142,000-participant NHS-Galleri trial of a multi-cancer blood test, which missed its primary endpoint in February 2026 but showed a secondary-endpoint reduction in stage IV disease.
Why Gives readers randomized-trial results, not just observational data, for three different AI and molecular screening approaches.
Radiotherapy auto-contouring and liver cancer LLM evidence
Oncology adds a randomized multicenter trial of AI-assisted auto-contouring in thoracic and breast cancer radiotherapy (NCT05787522; Niu et al., 2026; Yu et al., 2025) and a 2026 nationwide retrospective study of 13,614 hepatocellular carcinoma patients evaluating an LLM-based clinical application (Yang et al., 2026).
Why Extends the chapter's AI evidence from imaging and screening into radiation therapy planning and liver cancer prediction.
ESMO AI-biomarker framework and patient-reported outcomes evidence
Oncology adds ESMO's first international consensus framework classifying AI-based biomarkers into four validation tiers (Aldea, Kather et al., Annals of Oncology, 2026), and a new section on electronic patient-reported outcome (ePRO) monitoring citing a 355-patient noninferiority trial (Visweswaran et al., 2026) and a study in which 91% of patients would recommend the system (Basch et al., 2025).
Why Gives oncologists a framework for validating AI biomarkers and highlights patient-reported outcomes as an evidence base the chapter says has been overlooked.
DYNAMIC-III trial does not replicate ctDNA-guided treatment result
Oncology adds that DYNAMIC-III, a stage III colon cancer trial, did not replicate the positive result of the earlier DYNAMIC trial for ctDNA-guided treatment decisions (Tie et al., 2025; Morris et al., 2024).
Why Balances the chapter's positive ctDNA-guided trial evidence with a trial that did not succeed.
Machine learning to reduce low-value clinical alerts
Internal Medicine adds that machine learning approaches can reduce low-value clinical alerts while maintaining safety (Baron et al., JAMIA Open, 2021).
Why Adds evidence for reducing alert fatigue without compromising patient safety.
ASCO's six guiding principles for AI in oncology
Oncology adds ASCO's six guiding principles for responsible AI use in oncology, published May 2024 and updated since.
Why Gives oncologists a professional-society framework for evaluating AI tools in their own practice.
New chapter: Industry Reports & Benchmarks
Adds a new Foundations chapter cataloging major annual reports on AI in healthcare (Rock Health, CB Insights, AMA physician surveys, HIMSS/Medscape, McKinsey, Deloitte, KLAS, Accenture, ONC, RAND, Stanford AI Index, CHAI, ECRI, and AHA), organized by report type, funding source, and what each report measures.
Why Gives physicians a way to compare AI adoption claims across market reports and surveys that often report very different figures for the same phenomenon.
Chapter Industry Reports and Benchmarks
Healthcare AI adoption compared with other sectors
AI Fundamentals for Clinicians adds that about 8.3% of healthcare firms use AI in producing goods or services, compared with 11.6% in finance, 15.1% in education, and higher shares in other sectors.
Why Shows readers how healthcare AI adoption compares with other industries.
Chapter AI Fundamentals for Clinicians
AI-use disclosure trends in JAMA Network journals
LLMs in Practice adds that a cross-sectional study of manuscripts submitted to 13 JAMA Network journals found 3.3% of authors disclosed AI use overall, rising from 1.71% to 5.97% of submissions over 27 months (Perlis et al., JAMA, 2026).
Why Shows readers how disclosed AI use in medical publishing has changed over time.
Ötleş et al. on shadow AI governance gaps in hospitals
Policy & Governance adds that hospital AI governance often covers only enterprise-procured, EHR-integrated tools, while many clinicians already use unsanctioned shadow AI on personal devices, so banning or restricting official tools does not eliminate this use, only makes it invisible (Ötleş et al., NEJM AI, 2026).
Why Argues institutional governance needs to address AI use hospitals do not officially provide, not just the tools they procure.
Sycophancy evidence: Cheng et al. and Chandra et al.
Safety & Risk adds that 11 leading AI models affirmed users' actions 49% more often than humans, including in scenarios involving deception, illegality, or harm (Cheng et al., Science, 2026), and that modeling shows idealized users can grow dangerously overconfident in false beliefs when chatbots validate them (Chandra et al., 2026, preprint).
Why Strengthens the chapter's sycophancy warning with quantified evidence that validation-seeking behavior is a safety failure mode, not a style issue.
Expands the Hixon et al. sepsis AMR prediction findings
AI Fundamentals, LLMs in Practice, and Infectious Diseases now detail the Hixon et al. (2026) comparison: deep learning on structured EHR data reached AUROC 0.85 versus 0.74 for LLMs reading clinical notes, in almost 30,000 sepsis episodes across 10 BJC Healthcare hospitals, predicting ceftriaxone-resistant Gram-negative infections, with no gain from combining both approaches.
Why Gives readers the study's patient numbers, hospitals, and outcome measured, not just the headline comparison.
Chapters AI Fundamentals for Clinicians · Large Language Models in Clinical Practice · Infectious Diseases and Antimicrobial Stewardship
AI cervical screening: colposcopy meta-analysis and African AVE study
OB/GYN adds a 77-study meta-analysis showing AI outperforms experienced colposcopists (OR 1.75, 95% CI 1.33–2.31), a five-country African prospective study of smartphone-based AI cervical screening in 24,447 women (60.1% sensitivity, 81.9% specificity, exceeding standard VIA), and a $300 microscope paired with AI for low-cost precancer detection.
Why Broadens the chapter's evidence for AI-assisted cervical cancer screening from colposcopy into low-resource, low-cost screening settings.
Chapter Obstetrics and Gynecology
AI subtyping and prognosis models for gynecologic cancers
Pathology adds a Gynecologic Histopathology AI section: a deep-learning model identifying a high-risk endometrial cancer subtype invisible to conventional pathology (Darbandsari et al., Nat Commun, 2024), a graph-based model predicting ovarian cancer prognosis from H&E slides alone (Yang et al., Eur J Cancer, 2024), and a review noting gynecologic cancers remain understudied in digital pathology (Joshua et al., Cancers, 2025).
Why Extends the chapter's digital pathology coverage to gynecologic cancers, a less-studied area relative to prostate and breast cancer AI.
Macroscopic gross pathology AI: a new pathology subsection
Pathology adds a new section on macroscopic gross pathology AI, covering foundation models (UNI, CONCH, GigaPath, mSTAR, Atlas) that generalize across tumor types (Xu et al., 2025), a gastric cancer specimen model achieving 95.9% mean average precision for lesion localization (Kiyotoki et al., 2024), and the NCI's 2025 workshop identifying multiscale macroscopic-microscopic analysis as a priority research direction.
Why Gives readers a first look at an emerging pathology AI subfield that trails digital histopathology AI by years.
FDA launches TEMPO digital health device pilot
Policy & Governance adds FDA's Technology-Enabled Meaningful Patient Outcomes (TEMPO) pilot, announced December 5, 2025, which connects participating digital health manufacturers to CMS's Advancing Chronic Care with Effective, Scalable Solutions (ACCESS) model, planned to launch July 1, 2026.
Why A new FDA pilot links device evidence generation to a Medicare payment model, a mechanism the chapter argues drives clinical AI adoption.
Texas TRAIGA AI disclosure law takes effect
Policy & Governance updates its state AI regulatory sandbox table to reflect TRAIGA, which took effect January 1, 2026, requiring plain-language disclosure when AI is used in healthcare services or treatment and creating a 36-month regulatory sandbox.
Why Texas becomes an operational example of a state-level AI disclosure mandate that physicians practicing there must follow.
Ohde et al. on ambient scribe risks beyond primary care
Clinical Documentation adds a 2026 npj Digital Medicine perspective warning that ambient AI scribes, validated mostly in ambulatory primary care, face distinct failure modes in emergency departments, ICUs, operating rooms, telehealth, and other high-acuity or high-noise settings.
Why Cautions readers against assuming a scribe tool validated in outpatient care is safe elsewhere without local testing.
Lee et al. review of 70 FDA-cleared orthopedic AI devices
Orthopedics adds findings from a review of 70 FDA-cleared orthopedic AI devices: 22.8% had no clinical testing and 68.6% relied on retrospective datasets, though untested rates fell from 62.2% (2017 to 2019) to 19.7% (2022 to 2024).
Why Gives readers a quantified picture of the evidence gap behind orthopedic AI clearances and shows it is narrowing over time.
AI clinical decision support for post-arthroplasty hyponatremia
Orthopedics adds a new subsection describing an AI-based clinical decision support tool from Hospital for Special Surgery aimed at reducing hyponatremia after total joint arthroplasty (Kunze et al., NEJM Catalyst, 2026), noting limitations on independent validation.
Why Extends orthopedic AI coverage beyond imaging, navigation, and recovery tools into perioperative risk prediction.
Philips EchoNavigator R5.0 intraprocedural AI guidance
Adds coverage of Philips' EchoNavigator R5.0, FDA-cleared (510(k) K253614) for structural heart procedures, whose DeviceGuide feature uses AI to track Edwards PASCAL Ace M-TEER devices in real time by combining live echocardiography with fluoroscopy.
Why Marks cardiology AI moving from offline image interpretation into real-time intraprocedural guidance for structural heart interventions.
Court sanctions escalate over AI-generated inaccuracies in filings
Liability & Compliance adds that by early 2026, over 700 court decisions had addressed AI-generated inaccuracies in legal filings, with courts escalating from monetary sanctions to attorney disqualification and bar referrals.
Why Gives physicians context on the broader legal consequences of unverified AI output as AI-related litigation accelerates.
FDA AI device count updated to 1,451 (March 2026)
Updates the FDA-authorized AI/ML device count to over 1,450 (1,451 as of March 2026) and the radiology share to over 1,100 devices (about 76% of all authorizations) across the specialty tools appendix, TL;DR compilation, executive summary, future perspectives, physician toolkit, radiology, and neurology chapters.
Why Keeps the handbook's device and specialty adoption figures current.
Chapters AI Tools by Medical Specialty · Quick Reference: All Chapter Summaries (TL;DRs) · Executive Summary · The Physician-AI Partnership: Future Perspectives · AI Tools Every Physician Should Know · Diagnostic Imaging, Radiology, and Nuclear Medicine
References, links, and wording were also kept current.
March added Harrison.ai's petition to exempt radiology AI from FDA premarket review, and covered Claire, reported as the first FDA-approved intraoperative AI for breast cancer surgery, and RecovryAI, the first FDA Breakthrough Device for patient-facing generative AI. The AMA's 2026 survey found physician AI use more than doubled since 2023, and a review examined a Kenyan AI clinical support system. The FDA-authorized AI device count rose to over 1,400.
State laws requiring human review of AI healthcare decisions
Liability & Compliance adds Texas SB 1188, Illinois HB 1806, and Arizona HB 2175, which require physician or provider review of AI-generated records, treatment plans, or insurance denials, plus a pending New York bill (SB 7263) creating liability for AI chatbots giving unlicensed professional advice.
Why Gives readers a state-by-state view of emerging legal requirements for human oversight of clinical AI.
Autonomous AI telemedicine triage study (Saenz et al.)
Emerging Technologies adds a clinician-blinded study of an autonomous AI system across 2,379 U.S. primary care telemedicine encounters: 91.3% top-1 diagnostic accuracy against the treating clinician, rising to 96.3% above a safety confidence threshold, and a 2.5% disposition error rate. Preprint, not yet peer-reviewed.
Why Gives readers the first large-scale, real-world evidence on autonomous AI performance and a proposed staged deployment model in primary care telemedicine.
Generalist Biological AI foundation models (Rao et al. review)
Emerging Technologies adds a subsection on Generalist Biological AI (GBAI): foundation models trained across DNA, RNA, protein, and cellular data within one framework, contrasted with single-task tools like AlphaFold and Evo.
Why A 2026 review defines this emerging model class for readers tracking AI that spans DNA to cellular behavior instead of one task at a time.
Voice-cloned patient education RCT (Sun et al.)
Emerging Technologies adds a 3-arm RCT (n=174) comparing traditional education, physician-voice-cloned AI content, and patient-self-voice-cloned AI content, with compliance of 73.2%, 86.7%, and 92.5% and better 1-month adherence in the self-voice group.
Why Provides randomized evidence on whether personalizing AI patient education audio with a familiar voice improves compliance.
Mayo Clinic radiology AI monitoring benchmark (Cook et al.)
Evaluating AI Systems adds a 2026 Mayo Clinic review of 17 internally developed radiology AI algorithms in live use, describing daily, weekly, and yearly monitoring for utilization, failure modes, and data drift.
Why Gives readers an operational benchmark for post-deployment AI monitoring to ask vendors about before adoption.
AI-assisted radiology liability: two mock-juror studies
Liability & Compliance and Radiology & Nuclear Medicine add two mock-juror studies: a 1,334-adult study found jurors impose an added 'AI penalty' on radiologists who miss AI-flagged findings, reduced when AI error rates are disclosed, and a follow-up (n=282) found double-review workflows lower perceived negligence from 75% to 53%.
Why Independent mock-juror research now quantifies how AI transparency and review workflow affect perceived legal liability for radiologists.
Chapters Physician AI Liability and Regulatory Compliance · Diagnostic Imaging, Radiology, and Nuclear Medicine
ChatGPT Health undertriages emergency-level cases (Ramaswamy et al.)
Emergency Medicine adds a Mount Sinai evaluation of ChatGPT Health across 60 vignettes and 960 prompt variants in a 2x2x2x2 factorial design, finding consistent downgrading of emergency-level cases to 'urgent' and sociodemographic variation in triage recommendations by patient race and sex.
Why Warns readers that a widely available consumer AI triage tool is not validated for acuity assessment and may underestimate urgency unevenly across patient groups.
Chapter Emergency Medicine
Mammography AI triage noninferiority trial (AITIC, Elias-Cabot et al.)
Radiology & Nuclear Medicine adds the AITIC trial (NCT04849776), a prospective paired noninferiority RCT in Spain's Cordoba screening program testing AI triage to classify exams as low-risk and omit them from radiologist reading, finding stand-alone AI performance comparable to radiologists.
Why Gives readers randomized trial evidence on whether AI triage can safely reduce radiologist reading workload in mammography screening.
Latent raises $80M for AI medication approval platform
Primary Care notes that Latent raised $80 million in Series A funding for an AI platform automating EHR data processing and drug criteria interpretation, reportedly reaching 50% of the top 20 U.S. health systems.
Why Flags a fast-growing medication-approval AI platform and reminds clinicians at adopting institutions to verify AI-generated approvals before acting on them.
Chapter Primary Care, Family Medicine, and Preventive Medicine
GE HealthCare completes $2.3B Intelerad acquisition
Radiology & Nuclear Medicine notes GE HealthCare's completed $2.3 billion acquisition of Intelerad on March 19, 2026, combining Intelerad's cloud PACS, imaging workflow, and image-sharing tools with GE HealthCare's AI imaging portfolio.
Why Flags a PACS market consolidation that narrows independent vendor options for radiologists evaluating imaging platforms.
FDA-authorized AI devices: pathways and intended uses
Future Perspectives, Policy & Governance, AI Tools for Physicians, and Neurology now cite over 1,400 FDA-authorized AI/ML devices, with Policy & Governance noting approximately 96% cleared via the 510(k) pathway and Neurology noting none carry ALS-specific indications.
Why Distinguishes the size of the authorized-device landscape from its regulatory pathways and specific clinical indications.
Chapters The Physician-AI Partnership: Future Perspectives · Healthcare Policy and AI Governance · AI Tools Every Physician Should Know · Neurology and Neurological Surgery
ALS diagnosis: blood biomarkers and electrophysiology AI
Neurology adds a new section on AI for ALS diagnosis, covering plasma proteomics (AUC 98.3%) and gene-expression panels (AUC 0.894), deep learning on routine F-wave EMG data (88% accuracy), plasma lipidomics distinguishing ALS from primary lateral sclerosis, and NLP analysis of clinical notes finding an 11-month median diagnostic delay. None of these tools are yet clinically deployed.
Why Surveys emerging diagnostic AI for a disease where diagnostic delay is common, while flagging that none of the tools described has reached clinical deployment.
ALS prognosis and remote monitoring with wearables
Neurology adds AI approaches to ALS prognosis and monitoring: machine-learning models predicting ALSFRS-R functional decline from the 5,030-patient PRO-ACT trial database, a hip-worn accelerometer index associated with mortality risk, and remote speech and audio biomarkers that detect decline before ALSFRS-R changes, including one estimating forced vital capacity from speech acoustics.
Why Gives clinicians AI-based alternatives to standard ALSFRS-R scoring that can detect functional and respiratory decline earlier and remotely between visits.
Brain-computer interfaces restoring speech in ALS
Neurology adds a Brain-Computer Interfaces section covering the BrainGate speech neuroprosthesis (97.5% accuracy at 32 words per minute sustained over 8 months), real-time neural voice synthesis, motor-cortex decoding of inner speech reaching 74% accuracy with a 125,000-word vocabulary, and an eye-gaze text-entry interface. The section also notes ethical concerns about intention ambiguity, mental privacy, and responsibility for BCI-mediated speech.
Why Documents rapid progress in restoring communication for ALS patients through BCIs, alongside the ethical questions that direct neural speech decoding raises.
AI drug repurposing analysis for ALS from veteran health records
Neurology adds a causal-inference and machine-learning analysis of health records from over 11,000 US veterans with ALS, evaluating 162 medications and identifying 27 with statistically significant survival associations, including statins, PDE5 inhibitors, and alpha-adrenergic antagonists (Reimer et al., Lancet Digital Health, 2026). The chapter notes this is retrospective and observational, identifying candidates for prospective validation rather than proven treatments.
Why Identifies drug repurposing candidates for ALS from existing veteran health data without additional patient testing, though findings require prospective trials before clinical use.
AI beyond ALS: spinal muscular atrophy and frontotemporal dementia
Neurology adds AI research on the broader motor neuron disease spectrum: computer-vision detection of SMA-associated hypotonia from infant movement videos, and a systematic review of 25 AI studies (6,544 patients) for differentiating frontotemporal dementia, which affects about 15% of ALS patients with C9orf72 repeat expansions.
Why Extends the chapter's AI coverage to related motor neuron and neurodegenerative conditions that share diagnostic and monitoring challenges with ALS.
Independent safety review of Kenyan AI clinical support system
Workflow Integration, LLMs in Practice, AI Tools for Physicians, and Primary Care add a peer-reviewed safety review of 1,469 records from the same Kenyan clinical AI system, finding 99% guideline alignment and 3.4% hallucination rate, but actively harmful recommendations in 7.8% of encounters, with clinicians not modifying AI-generated content in 62% of cases.
Why Shows that a clinical AI system can perform well on average while still producing harmful outputs that reach documentation when clinicians defer to it without sufficient review.
Chapters Integration into Clinical Workflow · Large Language Models in Clinical Practice · AI Tools Every Physician Should Know · Primary Care, Family Medicine, and Preventive Medicine
Anthropic's global AI use survey: regional sentiment and benefit-harm tensions
Global Health Equity and Ethics & Health Equity add Anthropic's survey of 80,508 AI users across 159 countries, reporting that users in Sub-Saharan Africa, Central Asia, South Asia, and Latin America expressed more optimism about AI than users in North America, Western Europe, and Oceania, and identifying five benefit-harm tensions, with unreliability the top concern (26.7%).
Why Grounds the handbook's global-equity and ethics discussions in the largest multilingual qualitative study of how people worldwide experience AI's benefits and harms.
Chapters AI and Global Health Equity · Medical Ethics, Bias, and Health Equity
WHO Global Digital Health Strategy extended to 2027
Policy & Governance adds that the Seventy-eighth World Health Assembly extended the WHO Global Strategy on Digital Health 2020-2025 through 2027 in May 2025, noting that 129 countries have established national digital health strategies and a 2028-2033 follow-up framework is under development.
Why Gives readers the current status and timeline of the WHO's overarching digital health strategy, separate from its AI-specific guidance.
Mock-juror study: double review of AI-flagged imaging cuts liability
Liability & Compliance and Radiology & Nuclear Medicine add a randomized mock-juror study (n=282) finding that when radiologists documented reviewing AI-flagged imaging both before and after AI input, the share of jurors finding a duty-of-care breach dropped from 75% (single review) to 53%.
Why Suggests a documented double-review workflow may reduce physicians' legal exposure when using AI-flagged imaging, extending earlier mock-juror evidence on AI and liability.
Chapters Physician AI Liability and Regulatory Compliance · Diagnostic Imaging, Radiology, and Nuclear Medicine
Cleveland Clinic deployment of AI trial-screening across 25 hospitals
Research with AI adds Cleveland Clinic's real-world deployment of Dyania Health's Synapsis AI platform to screen patients for a Phase 3 ATTR-CM trial across 25 hospitals and 250 outpatient centers. Cleveland Clinic reported 96.2% accuracy across 7,700 trial-specific questions and that 36.6% of AI-identified candidates were Black, compared to 7.1% from routine screening.
Why The first large-scale, real-world evidence for AI-driven trial screening across a health system, and evidence the approach can surface more diverse candidates than routine screening.
Chapter Clinical Research with AI
AMA survey: physician AI use more than doubles since 2023
AI Fundamentals for Clinicians and Emerging Technologies add the AMA's 2026 Physician Survey on Augmented Intelligence, finding 81% of physicians now use AI in practice, more than double the 38% reported in 2023, while nearly half of physicians strongly oppose patient use of AI for radiology or pathology interpretation.
Why Shows physician AI adoption accelerating even as physicians remain skeptical of AI making independent clinical interpretations.
Chapters AI Fundamentals for Clinicians · Emerging AI Technologies in Healthcare
Microsoft Research: how patients use Copilot for health questions
LLMs in Practice adds a Microsoft Research analysis of over 500,000 de-identified health-related Copilot conversations, finding that nearly one in five involved personal symptom assessment and about one in seven were asked on behalf of someone else, often by caregivers; personal health queries increased during evening and nighttime hours.
Why Gives readers population-scale data on how patients use conversational AI for health questions outside clinical encounters.
Reporting standards for AI clinical research added to Research with AI
Research with AI adds three reporting-standard guidelines for AI clinical studies: SPIRIT-AI for trial protocols involving AI interventions, DECIDE-AI for exploratory clinical investigations of AI decision-support systems, and CONSORT-EHEALTH for web-based and mobile health interventions.
Why Gives researchers standard reporting frameworks to design and evaluate AI clinical studies against established methodology guidelines.
Chapter Clinical Research with AI
NCHR opposes Harrison.ai's FDA premarket review exemption petition
Policy & Governance adds that the National Center for Health Research filed formal opposition on February 27, 2026 to Harrison.ai's petition (Docket FDA-2025-P-5560) to exempt certain AI imaging tools from 510(k) premarket review, citing generalizability gaps, lack of prospective outcome evidence, and undetected performance drift risk.
Why The petition could change premarket review requirements for AI imaging devices; the opposition gives readers a documented counterargument grounded in evidence gaps.
PETRUSHKA trial: AI-guided antidepressant selection RCT
Psychiatry adds the PETRUSHKA trial, an international RCT across 47 sites in the UK, Brazil, and Canada with 540 adults starting antidepressant treatment, published in JAMA in March 2026. Patients using the AI-driven decision-support tool were 40% less likely to discontinue treatment within 8 weeks (17% versus 27% with usual care).
Why The first RCT to show AI-guided antidepressant selection improves treatment adherence, though the tool supports rather than replaces clinician judgment.
Chapter Psychiatry and Behavioral Health
Claire: first FDA-approved intraoperative AI for breast cancer surgery
Surgery & Anesthesiology and Tools by Specialty add Claire (Perimeter Medical Imaging AI), which the company reports received FDA Premarket Approval in March 2026 as the first AI-enabled imaging device for intraoperative breast cancer margin assessment. The pivotal trial (n=206) reported 88.1% overall accuracy and detected residual disease missed by standard methods in 40% of patients.
Why Gives surgical teams evidence and regulatory status for a new intraoperative margin-assessment tool that could reduce reoperation rates after breast-conserving surgery.
Chapters Surgery, Anesthesiology, and Perioperative Care · AI Tools by Medical Specialty
Athenahealth survey: physician comfort with AI rising
AI Fundamentals for Clinicians adds cross-sector AI adoption figures (8.3% of healthcare firms use AI in producing goods or services, versus 11.6% in finance and 15.1% in education) and the 2026 Athenahealth Physician Sentiment Survey (n=1,045), which found 54% of physicians comfortable with AI in practice, up from 46% in 2025.
Why Quantifies how physician comfort with AI is changing year over year and situates healthcare's adoption rate against other sectors.
Chapter AI Fundamentals for Clinicians
RecovryAI: first FDA Breakthrough Device Designation for patient-facing AI
Orthopedics and Policy & Governance add RecovryAI, an AI chatbot prescribed after joint-replacement surgery that contacts patients about sleep, activity, and diet and escalates concerns to the care team. The company reports it received the first FDA Breakthrough Device Designation for a large-language-model product designed for direct patient interaction.
Why Signals regulatory openness to generative AI that interacts directly with patients post-discharge, though no completed randomized trial data on clinical outcomes exists yet.
Chapters Orthopedic Surgery and Physical Medicine · Healthcare Policy and AI Governance
PathAI's PathAssist Derm receives FDA Breakthrough Device Designation
Pathology adds that PathAI received FDA Breakthrough Device Designation on March 3, 2026 for PathAssist Derm, an AI system that analyzes whole slide images of skin lesions to assist dermatopathologists. The chapter notes PathAssist Derm is currently for research use only and not cleared for diagnostic use.
Why Breakthrough Device Designation accelerates FDA review but is not a clearance; the research-use-only status tells readers the tool is not yet available for diagnostic use.
MedVersa generalist multimodal radiology foundation model
Radiology & Nuclear Medicine adds MedVersa, a foundation model trained on 29 million instances from 91 public datasets, evaluated across classification, segmentation, visual question answering, and report generation from one architecture. Radiologists rated its chest radiograph reports clinically equivalent to human-written reports in 64% of cases overall, rising to 91% for normal studies.
Why Gives readers peer-reviewed evidence that a single multimodal model can perform across radiology tasks, though independent validation across health systems is still needed.
ATTRACTnet: multimodal AI screening for ATTR-CM
Cardiology adds ATTRACTnet, a model that screens ECG, echocardiographic, demographic, and ICD-code data to flag transthyretin amyloid cardiomyopathy. A nonrandomized, multisite, prospective trial published in JAMA Cardiology (February 2026) reported a 48% diagnostic yield and an 88% treatment initiation rate among screened patients.
Why ATTR-CM is often underdiagnosed because its presentation overlaps with other causes of left ventricular hypertrophy; the trial's editorial called for randomized evidence before wider clinical use.
Survey: hospitalist AI use outpaces governance
Internal Medicine adds a 2026 survey of 70 hospitalists at a large urban academic center finding 66.7% reported using AI in clinical practice despite no system-wide integration or institutional endorsement, mostly for differential diagnosis and clinical question support.
Why Quantifies a governance gap: AI adoption by hospitalists is outpacing institutional validation and training frameworks.
Delivery Date AI: FDA-cleared ultrasound gestational dating
OB/GYN and Tools by Specialty add Delivery Date AI, an ultrasound-based deep-learning tool that received FDA De Novo clearance (DEN250007) in February 2026 to predict delivery date directly from ultrasound images. The PAIR study evaluated over 5,700 patients and reported an R-squared of 0.92 for predicting days to delivery.
Why Gives readers an FDA clearance record and prospective validation evidence for an adjunctive dating tool intended for pregnancies where traditional dating is unreliable.
Chapters Obstetrics and Gynecology · AI Tools by Medical Specialty
CardioMEMS HERO pulmonary artery pressure reader
Internal Medicine adds Abbott's CardioMEMS HERO, an FDA-approved (February 2026) replacement reader for the implanted CardioMEMS PA Sensor, describing its wireless connectivity and under-60-second reading time. The chapter notes that the CardioMEMS system's CHAMPION and GUIDE-HF trial evidence and CMS coverage apply to the system as a whole, not the new reader specifically.
Why A new FDA-approved reader for an already-covered heart failure monitoring device gives readers an updated hardware detail without changing the underlying trial evidence or CMS coverage.
Case study: ChatGPT Health emergency triage failures
Case Studies adds an eighth case study on ChatGPT Health, citing a Nature Medicine evaluation that found a 52% emergency under-triage rate across simulated scenarios and inconsistent suicide crisis alert triggering for clinically equivalent presentations.
Why Gives readers a documented example of a consumer AI health tool failing at emergency triage and crisis detection, for evaluating similar products.
Chapter Clinical Case Study Library
FDA comment period on Class II 510(k) exemptions
Policy & Governance adds a section on a Federal Register notice (Docket FDA-2026-N-0232) identifying Class II devices that may be exempted from 510(k) premarket notification, with public comments accepted through April 7, 2026.
Why A new comment period could change which AI-enabled devices need premarket review before reaching clinicians.
Virtual Biotech multi-agent drug discovery framework
Oncology adds a Stanford preprint describing the Virtual Biotech, a multi-agent system that curated outcomes from 55,984 clinical trials and reported that drugs targeting cell-type-specific genes were 40% more likely to progress from Phase I to Phase II and 48% more likely to reach market, with 32% lower adverse event rates.
Why A preprint proposes multi-agent AI for drug discovery and target evaluation, though its recommendations lack independent clinical validation.
JAMA study on AI disclosure in manuscript submissions
LLMs in Practice cites a JAMA Network analysis of 105,538 manuscripts across 13 journals (2023 to 2025): 3.3% of authors disclosed AI use, most often to correct language (67.7%), and AI disclosure was associated with 29% higher odds of rejection and 748% higher odds of pre-review withdrawal.
Why Gives readers population-level data on how AI is used in manuscript preparation and its association with review outcomes.
STAT News reporting on Harrison.ai's FDA petition and physician implications
Policy & Governance adds STAT News reporting on Harrison.ai's petition (Docket FDA-2025-P-5560) to release AI imaging tools without 510(k) review when the FDA has already cleared a similar product, plus a summary of what the pending regulatory changes mean for physicians.
Why Independent news reporting gives readers the petition's mechanics and a physician-facing summary of what less premarket scrutiny could mean for practice.
FDA Quality Management System Regulation effective date
Policy & Governance notes that the FDA's Quality Management System Regulation (QMSR) took effect February 2, 2026, aligning device quality requirements with ISO 13485:2016 and introducing a new inspection program.
Why Gives readers the compliance date that determines when AI/ML device manufacturers must meet the new quality system requirements.
Harrison.ai's petition for a 510(k) exemption in radiology AI
Policy & Governance adds a section on Harrison.ai's October 2025 petition asking the FDA to exempt four radiology AI product codes (POK, MYN, QFM, QDQ) from premarket notification when a device is substantially similar to one already cleared.
Why The petition, if granted, would change premarket review requirements for radiology AI, a shift readers evaluating vendor tools should track.
Transparency gap in FDA-authorized AI/ML device summaries
Policy & Governance adds a statistic that only 23.7% of FDA-authorized AI/ML devices report dataset demographics in public summaries, and only 1.5% report Predetermined Change Control Plans.
Why A peer-reviewed analysis quantifies how little demographic and update-plan information accompanies most authorized AI/ML devices.
Prima vision-language model for brain MRI interpretation
Neurology adds Prima, a University of Michigan vision-language model that interprets brain MRIs across 52 neurological diagnoses, reporting a mean AUC of 92.0% (up to 97.5% for specific categories) and triage capability in the study setting.
Why A new study extends AI-assisted brain MRI interpretation beyond single-purpose tools such as stroke or hemorrhage detection to a broad set of neurological diagnoses.
Pediatric representation gap in FDA-authorized AI/ML devices
Pediatrics adds a statistic that of 876 FDA-authorized AI/ML devices through March 2024, only 149 (17.0%) were labeled for pediatric use, and only 28 of those (18.8%) reported validation datasets that included children.
Why A peer-reviewed analysis quantifies how few authorized AI/ML devices are validated for pediatric use.
Chapter Pediatrics and Neonatology
Qure.ai qXR-Detect chest X-ray CADe clearance
Radiology & Nuclear Medicine adds Qure.ai's qXR-Detect, a multi-indication CADe system for chest X-rays that received FDA 510(k) Class II clearance in February 2026 covering lung, pleura, mediastinum/hila, heart, bone, and hardware findings. Qure.ai reports this is their only chest X-ray CADe device cleared with a Predetermined Change Control Plan (PCCP).
Why PCCP coverage is reported to reduce the resubmission burden as health systems adopt vendor devices that continue to update over time.
References, links, and wording were also kept current.
February added the first randomized trial finding that LLM medical accuracy does not transfer to public use, and a Nature Medicine stress test showing ChatGPT Health under-triaged 52% of emergency vignettes. Cardiology gained completed results from the GUARD-AF and HEARTLINE atrial fibrillation screening trials, and Epic launched its native AI Charting ambient documentation feature. A 20-model benchmark found LLMs accepted fabricated medical data in 31.7% of prompts.
Audio-only ambient scribes miss visual clinical information
Clinical Documentation, Emerging Technologies, and AI Tools for Physicians add a 110-interview study: a vision-enabled scribe processing smart-glasses video reached 98% accuracy on simulated medication histories versus 81% for audio-only processing (P<0.001), cutting omission errors from 358 to 10.
Why Quantifies a structural gap in audio-only scribes for tasks needing visual input, such as medication histories and wound assessments, that better speech recognition cannot close.
Chapters AI-Assisted Clinical Documentation · Emerging AI Technologies in Healthcare · AI Tools Every Physician Should Know
AI training boosts physician clinical competence scores
Medical Education adds a pre-post study of 326 physicians across 39 countries: a 1.5-hour training on effective AI use raised average competence scores from 56.9% to 77.6% (P<.001) and pass rates from 6.4% to 58.6% with GPT-4 assistance.
Why Suggests that brief, targeted training on how to use AI effectively matters more for clinical performance than the AI tool itself.
Chapter AI in Medical Education
Epic AI Charting: native ambient documentation in the EHR
AI Tools for Physicians adds Epic AI Charting (Art), a native ambient documentation feature launched February 2026 that drafts notes and queues orders from clinician-patient conversations. Epic controls 42% of the acute hospital EHR market and 55% of hospital beds (STAT News, February 2026); independent clinical validation is not yet available.
Why Native EHR integration gives the ambient documentation feature immediate distribution across health systems already running Epic, without requiring a separate vendor contract.
ChatGPT Health triage stress test: 52% emergency under-triage
LLMs in Practice adds a Nature Medicine stress test of ChatGPT Health triage (60 vignettes, 21 domains, 960 responses): 52% of gold-standard emergency cases were under-triaged to 24-48 hour care, and family-member language minimizing symptoms shifted recommendations toward less urgent care (OR 11.7). Crisis intervention messaging also fired inconsistently across suicidal ideation scenarios.
Why Shows the system cannot reliably distinguish acuity levels and gives confident-sounding responses regardless of uncertainty, despite handling classical emergencies correctly.
Surgical navigation malfunctions underreported to FDA's MAUDE
Policy & Governance replaces a generic surveillance gap example with a case in which voluntary MAUDE reporting captured only 8 malfunctions before an AI update, versus 100+ afterward, with 10+ patient injuries (Reuters, February 2026).
Why The underreporting illustrates the policy remedy the chapter recommends: mandatory AI-specific adverse event reporting.
LLM susceptibility to fabricated medical misinformation
Medical Misinformation adds a benchmark of 20 LLMs across 3.4 million prompts: models accepted fabricated medical data in 31.7% of prompts, with clinical note framing producing 46.1% susceptibility versus 8.9% for social media content. GPT-4o showed the lowest susceptibility and Llama-3.1-8B the highest (50.7%).
Why Shows that authoritative clinical-note phrasing bypasses model safeguards more effectively than informal content, with a counter-intuitive finding that explicit logical-fallacy framing reduced susceptibility.
Chapter Medical Misinformation and AI
Retia Argos Infinity: FDA-cleared cardiac output monitoring
Critical Care adds Retia Argos Infinity (K253092, FDA cleared February 2026), which estimates cardiac output from arterial waveform data without added hardware. Cites two validation studies against FloTrac and thermodilution (87% vs. 72% trending concordance), noting the first study's lead author co-developed the algorithm.
Why Gives physicians validation evidence, including a conflict-of-interest caveat, alongside the FDA clearance for a newly available monitoring platform.
MASAI mammography trial: full two-year RCT results
Radiology updates the MASAI trial entry with two-year follow-up results (N=105,934): AI-supported screening improved sensitivity (80.5% vs. 73.8%, P=0.031) at equal specificity, with fewer interval and aggressive cancers, though these reductions were not statistically significant.
Why Gives readers the trial's confirmed long-term outcomes now that two-year follow-up data are available, beyond the earlier interim workload-reduction findings.
First head-to-head RCT of ambient AI scribes
AI Tools for Physicians adds the first head-to-head RCT of two ambient scribe products (238 UCLA Health physicians, 71,369 encounters): Nabla cut documentation time 9.5% versus control (P=0.02), while DAX's smaller reduction was not statistically significant (P=0.66); a concurrent 66-practitioner RCT at UW Health is also cited.
Why Distinguishes vendor-reported documentation-time savings from peer-reviewed, controlled trial results as ambient scribes see wider adoption.
Aidoc CARE: multi-condition abdominal CT triage foundation model
Adds Aidoc CARE (K252970, FDA cleared January 2026), a foundation model triaging 14 acute abdominal CT conditions from one scan in place of single-condition algorithms; reported mean sensitivity 97%, specificity 98% in the pivotal study. Updates radiology, emergency medicine, neurology, toolkit, and emerging technology chapters.
Why The clearance is described as the first multi-condition imaging foundation model, shifting abdominal CT triage from first-in-first-out to urgency-based prioritization.
Chapters Emerging AI Technologies in Healthcare · AI Tools Every Physician Should Know · Emergency Medicine · Neurology and Neurological Surgery · Diagnostic Imaging, Radiology, and Nuclear Medicine · AI Tools by Medical Specialty
Silent trial methodology: scoping review and translational trial framework
Evaluating AI Systems and Policy & Governance add a scoping review of 75 silent evaluations across 16 countries, finding AI performance consistently drops from retrospective to prospective testing and that most teams report only AUROC while omitting subgroup equity and workflow testing; both add the CANAIRI 'translational trial' concept, assessing stakeholder readiness and workflow fit.
Why Gives readers primary evidence that pre-deployment silent testing needs to include equity and workflow checks, not just technical accuracy metrics.
Chapters Evaluating AI Clinical Decision Support Systems · Healthcare Policy and AI Governance
GUARD-AF and HEARTLINE atrial fibrillation screening trial results
Cardiology replaces its ongoing-trials section with completed results: GUARD-AF (n=11,905) found ECG patch screening raised AFib detection (5.0% vs 3.3%) and anticoagulation (4.2% vs 2.8%) but not stroke hospitalizations (0.7% vs 0.6%), and was halted early for COVID; HEARTLINE's AFib detection substudy is now cited, with stroke outcomes pending ACC 2026.
Why Both long-running AFib screening trials reported results, letting the chapter show that more detection has not yet been shown to translate into fewer strokes.
Three FDA-authorized prostate cancer AI devices
Surgical Subspecialties adds three FDA-authorized prostate cancer AI devices: Paige Prostate (De Novo, 2021) for pathology cancer detection, ArteraAI Prostate (De Novo, 2025), validated across 6 phase 3 trials with median 11.4-year follow-up for 10-year outcome prognostication, and Avenda iQuest (510(k), 2022) for tumor mapping, plus GleasonXAI, an explainable model for Gleason pattern grading.
Why Prostate cancer now has more FDA-authorized AI devices than any other urologic condition, spanning detection, prognostication, and treatment planning.
Chapter Surgical Subspecialties
Bladder cancer, kidney stone, and robotic surgery AI in urology
Surgical Subspecialties adds a multicenter cystoscopy study (accuracy up to 0.991), a 2025 meta-analysis showing AI outperforms clinicians on bladder tumor detection, the TOBY urine test (FDA Breakthrough Device Designation, June 2025), the AiFURS intraoperative kidney stone detection system (92 to 95% accuracy), and Medtronic's Hugo RAS robotic platform, FDA cleared in December 2025 for prostatectomy, nephrectomy, and cystectomy.
Why Documents urology AI moving from research-stage to multicenter validation and new FDA clearances across bladder cancer, kidney stone, and robotic-assisted surgery.
Chapter Surgical Subspecialties
AI for pulmonary function test interpretation and COPD prediction
Critical Care adds sections on PFT interpretation AI and COPD or asthma AI, citing a multicenter validation where AI matched the diagnostic gold standard with 100% pattern accuracy versus 74.4% for pulmonologists, and a 2025 meta-analysis of 13 studies on COPD exacerbation prediction.
Why Extends the chapter's AI evaluation coverage to pulmonary function testing and COPD monitoring, a gap in primary care and general pulmonology.
Automated insulin delivery devices and autonomous diabetic retinopathy screening
Internal Medicine adds two newly FDA-cleared automated insulin delivery systems (Omnipod 5 and Diabeloop DBLG2, both December 2025) alongside updated ADA Standards of Care guidance, and cites the ACCESS RCT showing autonomous point-of-care diabetic retinopathy screening reached 100% completion versus standard referral among diabetic youth.
Why New device clearances and trial evidence support point-of-care AI tools physicians can use for diabetes management and retinopathy screening.
FDA-cleared AI for thyroid nodule evaluation on ultrasound
Internal Medicine adds Koios DS, an FDA-cleared AI platform (510(k) K242130) for thyroid and breast ultrasound that aligns findings to ACR TI-RADS, with vendor-reported studies showing 14% higher cancer detection and 35% fewer unnecessary biopsies.
Why Gives readers the FDA clearance basis and vendor-reported performance for a thyroid ultrasound AI tool, noting independent validation is still pending.
Nephrology society AI framework and a CE-marked CKD risk algorithm
Internal Medicine adds the American Society of Nephrology's AI Workgroup framework for responsible AI use in kidney care, published in JASN, and the Kidney Klinrisk Algorithm (Roche), the first AI-based CKD risk stratification tool to receive a CE Mark, in October 2025, predicting kidney function decline in CKD stages G1 to G4.
Why A nephrology professional society and a newly CE-marked risk tool give physicians early guidance and evidence for AI in chronic kidney disease.
Gastroenterology AI beyond colonoscopy: liver, celiac, and gastric cancer
Surgical Subspecialties adds an AI-enabled ECG model that doubled new diagnoses of advanced liver disease in a primary care RCT, a pathology AI matching pathologist-level accuracy for celiac disease diagnosis, the GRAPE gastric cancer screening model (AUC 0.927 across 16 centers), and an FDA-cleared AI reading tool for small bowel capsule endoscopy.
Why Extends the chapter's gastroenterology coverage beyond colonoscopy to liver disease, celiac disease, and gastric cancer screening.
Chapter Surgical Subspecialties
Otolaryngology AI: cochlear implant outcomes and head and neck cancer
Surgical Subspecialties adds a study predicting cochlear implant language outcomes with 92% accuracy from pre-implantation brain MRI, the FDA-approved Cochlear Nexa implant with upgradeable firmware, and a multimodal AI model combining imaging and pathology to predict disease-free survival in oropharyngeal cancer.
Why Gives readers new AI applications in otolaryngology spanning pediatric cochlear implant planning, a newly approved implant, and cancer prognostication.
Chapter Surgical Subspecialties
LumineticsCore real-world diabetic retinopathy screening validation
Ophthalmology adds a real-world validation of the FDA-cleared LumineticsCore autonomous diabetic retinopathy screener across multiple health systems, reporting sensitivity of 87 to 100% and specificity of 60 to 91% for more-than-mild disease.
Why Gives readers real-world performance data for an autonomous screening tool already in clinical use across multiple health systems.
Chapter Ophthalmology
Reuters investigation: TruDi surgical navigation malfunctions after AI upgrade
Surgical Subspecialties adds a case study on the TruDi endoscopic sinus surgery navigation system, which had 8 reported malfunctions before AI algorithms were added and over 100 malfunctions and adverse events afterward, with at least 10 patients injured between late 2021 and November 2025, according to a Reuters investigation.
Why An investigative news report documents that adding AI to an existing surgical navigation system did not guarantee improved safety, with ongoing litigation over its accuracy claims.
Chapter Surgical Subspecialties
MedGemma succeeds Med-PaLM as Google's medical AI model
AI Tools for Physicians updates the Med-PaLM lineage (Med-PaLM to Med-PaLM 2 to Med-Gemini) to Google's current model, MedGemma, an open-source model based on Gemma 3, available in 4B multimodal and 27B text-only versions, and not approved as a medical device.
Why Keeps the toolkit's product descriptions current now that Google has retired Med-PaLM in favor of MedGemma.
Doximity launches Scribe free after a year of beta testing
AI Tools for Physicians describes Doximity's Scribe ambient-documentation tool as launched free to physicians in July 2025 after more than a year of beta testing with over 10,000 physicians, alongside Doximity's acquisition of Pathway Medical to integrate clinical datasets and AI capabilities.
Why Reflects Doximity moving its AI scribe from beta to general release and expanding its AI product suite through acquisition.
Randomized trial: LLM medical accuracy does not transfer to public use
LLMs in Practice and a new AI Failures & Lessons case study add a 2026 randomized trial (n=1,298) in which three LLMs identified medical conditions correctly alone in 94.9% of cases, but participants using the same LLMs identified relevant conditions in fewer than 34.5% of cases, no better than internet search (Bean et al., 2026).
Why Gives readers the first randomized evidence that LLM medical knowledge does not translate into effective use by the public.
Chapters Large Language Models in Clinical Practice · AI Failures in Medicine: Lessons Learned
Benchmark and human-AI interaction gap added to evaluation and safety chapters
Evaluating AI Systems adds an Interaction Gap section noting the same trial found MedQA scores above 60% corresponded to real-world participant scores below 20% on matching topics. Safety & Risk adds the trial to a broader AI+human pattern alongside Agarwal et al.'s finding that AI-assisted radiologists do not outperform AI alone.
Why Shows readers that benchmark and simulated-patient scores do not predict how real users perform with AI assistance.
Chapters Evaluating AI Clinical Decision Support Systems · Clinical AI Safety and Risk Management
Mechanistic interpretability audit of a Medicaid care coordination AI
Evaluating AI Systems adds a section on mechanistic interpretability, describing a 2026 audit of a Medicaid care coordination reinforcement learning system whose conformal safety model cleared 89.5% of decisions and flagged 10.5% for human review (Basu et al., 2026). Primary Care cites the same system's 12-percentage-point reduction in acute care events.
Why Gives readers a worked example of mechanistic interpretability, going beyond standard methods like SHAP and LIME, applied to a deployed clinical AI system.
Chapters Evaluating AI Clinical Decision Support Systems · Primary Care, Family Medicine, and Preventive Medicine
Echocardiography foundation models: PanEcho and EchoJEPA added to cardiology
Cardiology's Echocardiography Foundation Models section expands to cover PanEcho, a supervised multitask model trained across 39 reporting tasks (Holste et al., JAMA 2025), and EchoJEPA, a self-supervised latent-prediction model (Assran et al., 2025 preprint; Munim, Fallahpour, Szasz et al., 2026 preprint), noting EchoJEPA is a preprint while EchoPrime and PanEcho are peer-reviewed.
Why Gives readers primary sources for two new echocardiography foundation model approaches and flags which have been peer-reviewed.
FDA clarifies which AI documentation tools are not devices
Executive Summary adds a note on the January 2026 FDA Clinical Decision Support guidance, which clarifies that LLM-assisted documentation, literature summarization, and reasoning support fall outside FDA device regulation when physicians retain independent clinical judgment, and flags patients' growing use of personal health agents without clinical oversight.
Why Affects institutional procurement decisions and liability frameworks for AI documentation tools, and prepares physicians for patients arriving with agent-assisted research.
Chapter Executive Summary
Course syllabus template for teaching the handbook
Adds a Course Syllabus appendix: a 12-week graduate seminar syllabus (with a 6-week intensive option) built on the handbook, covering weekly readings, learning objectives, an assessment framework, supplementary reading lists, and a TL;DR summary of the pedagogical approach.
Why Gives medical schools, residency programs, and CME courses a ready-made curriculum for teaching evidence-based evaluation of clinical AI.
Chapter Course Syllabus Template
EHR foundation models for clinical trajectory prediction
Emerging Technologies adds a section on EHR foundation models, an approach distinct from text-based clinical language models, covering the CLMBR architecture, the EHRSHOT benchmark (6,739 Stanford patients, 41.6 million clinical events, 15 prediction tasks), and JEPA-based world models for patient trajectories, noting the approach remains research-stage without clinical validation.
Why Gives physicians a primer on an emerging model architecture built on medical-code sequences rather than clinical text, flagged as research-stage pending independent validation.
Autonomous AI agents in clinical practice
AI Fundamentals adds a section distinguishing AI agents, which plan and execute multi-step tasks with tool use, from single-turn chatbots, citing early oncology agent-concordance evidence and a benchmark finding 69.67% success for an LLM on 300 clinical EHR tasks.
Why Distinguishes a fundamentally different AI architecture, of growing clinical relevance, from the single-turn chatbots covered elsewhere in the handbook.
Chapters AI Fundamentals for Clinicians · Emerging AI Technologies in Healthcare · Hematology-Oncology and Precision Medicine
EU AI Act requirements for medical AI
Liability & Compliance adds a section on the EU AI Act's high-risk classification for medical device AI, covering the risk management, data governance, transparency, and human oversight requirements that apply alongside the EU Medical Device Regulation.
Why The EU's AI Act took effect in August 2024 and creates compliance obligations for medical AI beyond existing device regulation.
Evaluating explainability claims in clinical AI
Evaluating AI Systems adds a section on assessing explainable AI (XAI) methods, covering adoption evidence for SHAP and LIME, limits of attention-based saliency maps, and questions for evaluating vendor XAI claims.
Why Gives physicians criteria for evaluating vendor explainability claims, since interpretability affects clinician trust calibration and error detection.
Outpatient AI for chronic disease and remote monitoring
Internal Medicine adds an Outpatient and Ambulatory AI part covering AI diabetes apps (a 0.49% mean HbA1c reduction across 41 RCTs), a Johns Hopkins AI-led diabetes prevention trial, AI lifestyle coaching for blood pressure, and remote patient monitoring evidence.
Why Prior chapter coverage and most internal medicine AI literature focused on inpatient settings even though most internists practice in outpatient care.
Human-AI interaction design in clinical settings
Emerging Technologies adds a section on human-AI interaction patterns, covering how AI systems display information, uncertainty, and alternatives to clinicians, and evidence that human-AI teaming rarely achieves full complementarity.
Why A systematic review found AI development has focused on technical performance while overlooking alignment between algorithmic outputs and clinician expectations.
Equity dimensions beyond race in algorithmic bias
Ethics & Health Equity adds a section on non-racial equity dimensions in algorithmic bias, covering geographic disparities between rural and urban training data, socioeconomic proxies, and language and health-literacy gaps in NLP models and patient materials.
Why A 2025 systematic review found marginalized populations are often "incorrectly or insufficiently represented" in AI health systems, extending bias concerns beyond race.
Practical algorithmic auditing framework for clinical AI
Evaluating AI Systems adds a practical auditing guide built on the Lancet Digital Health medical algorithmic audit framework, subgroup testing methods including the AFISP framework, open-source auditing toolkits (Aequitas, Fairlearn), and post-deployment surveillance recommendations from STANDING Together.
Why Gives physicians a systematic way to detect hidden bias and failures before deployment rather than relying solely on vendor-reported performance metrics.
References, links, and wording were also kept current.
January saw major product launches: OpenAI's ChatGPT Health, Anthropic's Claude for Healthcare, and Utah's first state AI prescription-renewal pilot with Doctronic. The handbook added a Medical Education chapter, registered a Zenodo DOI, and cited the inaugural State of Clinical AI 2026 report and the FUTURE-AI trustworthy-AI framework. The FDA loosened Clinical Decision Support oversight, and a new model projected a 33% reduction in radiologist hours over five years.
Chin-Yee on dictation as a clinical reasoning process
Clinical Documentation expands its skill-erosion discussion to argue that dictating a note once forced physicians to organize observations into coherent reasoning, and that templated or ambient AI workflows can turn the note into a compliance artifact rather than a site of diagnostic thinking; it recommends trainees voice their assessment aloud before reviewing AI-generated notes.
Why Gives readers a citable argument for why note-writing itself may be a reasoning skill worth protecting as AI takes over documentation.
FDA loosens Clinical Decision Support software oversight
Policy & Governance adds that the FDA loosened Clinical Decision Support guidance effective January 6, 2026: single-recommendation software is now exempt from device regulation if it is the only 'clinically appropriate' option, undefined by the FDA, and time-critical decisions are now one factor rather than an automatic trigger for oversight.
Why A regulatory loosening that lets more diagnostic and treatment-recommendation AI reach clinics without FDA review.
Anthropic launches Claude for Healthcare; OpenAI's ChatGPT Health noted
LLMs in Practice adds that Anthropic launched Claude for Healthcare on January 11, 2026, with BAA coverage via AWS, Google Cloud, and Azure, a zero-training data policy, native ICD-10/NPI/PubMed integrations, and company-reported partners including Banner Health and Novo Nordisk.
Why Gives readers the enterprise capabilities and named partners of a new clinical AI product, flagged in the text as lacking independent clinical validation.
Bedi et al. on the MedHELM benchmark for medical LLMs
Evaluating AI Systems adds MedHELM, a clinician-validated taxonomy scoring LLMs across five categories of clinical work; among 9 frontier models, note generation scored highest (0.74 to 0.85) and administrative tasks lowest (0.53 to 0.63), and an LLM-panel judging method reached better agreement with clinicians (ICC 0.47) than clinicians reached with each other (ICC 0.43).
Why Gives readers a benchmark that scores clinical LLM performance by task type rather than a single aggregate accuracy figure.
Shah et al. on Stanford's DAX Copilot pilot outcomes
Emerging Technologies adds results from Stanford Health Care's 2024 DAX Copilot pilot (n=48 physicians): burnout fell 1.94 points and task load fell 24.42 points (both p<.001); a related analysis found documentation time fell 0.57 minutes per note and 19.95 minutes per day. Stanford deployed the tool enterprise-wide in March 2024.
Why Gives readers quantified, peer-reviewed outcomes for an ambient documentation tool already in wide clinical use.
Stanford's RAIL and FURM frameworks for AI lifecycle governance
Policy & Governance adds a new section on Stanford Health Care's Responsible AI Lifecycle (RAIL) framework, which stages AI projects from proposal through development, validation, pilot, and monitoring, and its FURM assessment (Fair, Useful, Reliable) required before deployment.
Why Gives readers a concrete institutional model for how a health system structures AI governance from proposal to monitoring.
Stanford ChatEHR as a model for institutional LLM deployment
LLMs in Practice adds a new Part 11 describing Stanford Health Care's ChatEHR, a build-from-within platform connecting multiple LLM providers to the full longitudinal patient record within the EHR, contrasted with standalone tools like web-based ChatGPT that create workflow friction from manual data entry.
Why Gives readers a worked example of how one health system built and governs LLM access to the full patient record.
New section on ELIZA and the origins of computer psychotherapy
History of AI in Medicine adds a section on Weizenbaum's ELIZA (1964-1966) and psychiatrist Kenneth Colby's 1966 proposal for computer-delivered psychotherapy, including Colby's PARRY chatbot simulating paranoid schizophrenia.
Why Traces today's concerns about emotional attachment to AI chatbots to a documented debate from the 1960s over simulated versus genuine therapeutic understanding.
Chapters History of AI in Medicine · Psychiatry and Behavioral Health
Lekadir et al. on the FUTURE-AI trustworthy AI framework
Ethics & Health Equity adds the FUTURE-AI framework, a two-year Delphi consensus from 117 experts in 50 countries, describing six guiding principles (fairness, universality, traceability, usability, robustness, explainability) and 30 best practices across the AI lifecycle.
Why Gives readers an international consensus checklist for evaluating whether a clinical AI system is trustworthy across its full lifecycle.
Ng et al. review of retrieval-augmented generation in healthcare
LLMs in Practice adds an early review of retrieval-augmented generation (RAG) for healthcare, describing how it addresses outdated information, hallucinations, and reliance on public training data by grounding responses in retrieved guidelines and institutional sources; the review notes RAG reduces but does not eliminate hallucinations.
Why Gives readers a framework for understanding why RAG-grounded LLM tools are more reliable than general-purpose chatbots, and where the approach still falls short.
Huo et al. and Gallifant et al. on LLM research reporting standards
Research with AI adds a systematic review of 137 studies on LLM chatbots for health advice, finding 99.3% did not identify the specific model tested and 64.5% used subjective ground truth, and describes the resulting TRIPOD-LLM reporting guideline, which requires exact model version, API access date, and full prompt documentation.
Why Gives readers a benchmark for judging LLM health-research quality and a citable standard for what rigorous reporting should include.
Chapter Clinical Research with AI
Sahni et al. and others on the case for operational AI in healthcare
Policy & Governance adds coverage of operational AI economics, citing estimates that administrative activities cost about $950 billion annually and that wider AI adoption could save $200 to $360 billion annually, and drawing on a documented 'productivity J-curve' and cross-industry examples to argue that returns require redesigning workflows around AI rather than retrofitting it.
Why Gives readers an evidence base for why administrative AI, not clinical decision support, may be where near-term healthcare savings materialize.
Nong et al. on the gap in local AI model validation
Policy & Governance adds the learning health system model as a framework for pairing AI deployment with ongoing evaluation, citing a finding that only 61% of U.S. hospitals performed any local performance evaluation of an AI model before deployment.
Why Highlights how few health systems independently verify AI performance before deploying it, a gap readers should weigh against vendor claims.
Task-specific LLM hallucination rates in clinical use
LLMs in Practice adds that hallucination rates vary sharply by task: about 1.5% for clinical note summarization versus 28-39% for reference generation in systematic reviews, with RAG-augmented systems reducing rates to 0-6%; it also reports Stanford Medicine's ChatEHR averaged 0.73 hallucinations and 1.60 inaccuracies per summarization across 23,000 sessions.
Why Gives readers task-specific benchmarks instead of a single hallucination rate, and a large real-world deployment figure for context.
CHAI and Joint Commission guidance for health AI deployment
Policy & Governance adds that CHAI partnered with the Joint Commission in September 2025 to release initial practical guidance for responsible AI deployment, with use-case-specific work groups covering clinical decision support, EHR information retrieval, prior authorization matching, and consumer health chatbots; a voluntary certification program based on CHAI's playbooks is planned.
Why The Joint Commission accredits over 22,000 healthcare organizations, so this partnership could set de facto standards for institutional AI governance.
Updated cost and penalty figures across three chapters
LLMs in Practice refines HIPAA penalty figures ($137 to $68,928 per violation, up to $2 million in annual caps per category, 2025 inflation-adjusted); Research with AI updates clinical trial cost figures (Phase III recruitment delays at $600,000 to $8 million per day) and the I-SPY 2 adaptive trial example; Oncology updates settlement and testing cost ranges.
Why Keeps dollar figures for penalties, trial costs, and malpractice exposure current and precise for readers making practical decisions.
Chapters Large Language Models in Clinical Practice · Clinical Research with AI · Hematology-Oncology and Precision Medicine
Wu et al. on the CPT billing gap in cleared AI devices
Policy & Governance adds that only two AI applications exceed 10,000 CPT claims despite over 1,300 FDA clearances, with IDx-DR as a rare Medicare-covered exception; it also cites evidence that private payers largely track Medicare rates and notes new CMS AI billing codes taking effect in January 2026, alongside the pending Access to Prescription Digital Therapeutics Act.
Why Shows readers why FDA clearance alone has not translated into billable, reimbursed clinical AI use.
Venture funding favors operational AI over clinical decision support
Policy & Governance adds that healthcare AI drew $18 billion in venture investment in 2025 (46% of healthcare investment), with AI-enabled startups taking 62% of that funding because investors favor operational-efficiency products over clinical decision support tools awaiting reimbursement clarity.
Why Explains, with investment figures, why capital and engineering talent are not flowing toward reimbursement-dependent clinical AI.
Reimbursement as a driver of clinical AI adoption
Policy & Governance argues from the EHR (HITECH Act, about $27-35 billion in incentives) and telehealth precedents that clinical AI adoption will track reimbursement, not regulatory clearance alone.
Why Grounds the chapter's adoption forecast in two documented U.S. health-technology adoption cycles rather than assertion.
Clarified framing on AI brittleness and failure modes
History of AI in Medicine, Safety & Risk, and Neurology sharpen language on how AI systems can fail: narrow systems that work well in-domain but fail on edge cases, models that fail severely on data unlike their training set, and suicide-risk algorithms whose high AUCs mask over 99% false positive rates.
Why Sharper, more specific language helps readers recognize concrete failure patterns rather than abstract warnings about AI limitations.
Chapters History of AI in Medicine · Clinical AI Safety and Risk Management · Neurology and Neurological Surgery
Jin et al. on TrialGPT for clinical trial matching
Research with AI adds detail on TrialGPT's three-module design (retrieval, matching, ranking) and reports 87.3% criterion-level matching accuracy and a 42.6% reduction in trial-screening time in a pilot study at NCI, with a link to the open-source code.
Why Gives readers a concrete example of LLM-based clinical trial matching with published accuracy and time-savings figures.
Chapter Clinical Research with AI
Azad et al. on deferral as a clinical AI safety mechanism
Safety & Risk adds a section on deferral, explicit abstention when confidence is low or a case falls outside validated scope, with under-triage, omission, and unsafe non-deferral rate metrics; Evaluating AI Systems and the glossary add matching definitions.
Why Gives readers a citable framework and metrics for judging whether a clinical AI system fails safely by deferring rather than guessing.
Chapters Clinical AI Safety and Risk Management · Evaluating AI Clinical Decision Support Systems · Glossary of Medical AI Terms
Radiology device count updated; chapter adds stroke-detection and mammography examples
Radiology & Nuclear Medicine, the Executive Summary, Chapter Summaries, and Medical Education align the radiology FDA-cleared device count at over 1,000 devices (1,039 as of December 2025). The chapter's opening adds that AI-assisted stroke detection reduces door-to-treatment times by 30 to 50 minutes, while early mammography CAD systems increased false positives without improving cancer detection.
Why Keeps the device count consistent across the handbook and grounds radiology's AI maturity in concrete examples of benefit and harm.
Chapters Diagnostic Imaging, Radiology, and Nuclear Medicine · Executive Summary · Quick Reference: All Chapter Summaries (TL;DRs) · AI in Medical Education
Epic Sepsis Model failure statistics detailed in appendices
Case Studies and AI Failures & Lessons add specific Epic Sepsis Model performance figures: 33% sensitivity, with 67% of cases missed, only 7% detected before clinical onset, against a claimed AUC of 76%.
Why Gives readers the specific numbers behind the Epic Sepsis Model's gap between claimed and real-world performance.
Chapters Clinical Case Study Library · AI Failures in Medicine: Lessons Learned
AMA commentary and guidance added to appendices
Case Studies adds a caution from AMA CEO John Whyte that AI without physician input risks patients and physicians alike (Becker's Hospital Review, January 2026), and Vendor Evaluation adds a link to AMA's AI governance guidance on health system AI success.
Why Gives readers a physician-organization perspective on AI oversight risk alongside a practical governance checklist for health systems.
Chapters Clinical Case Study Library · AI Vendor Evaluation Framework
RAISE consortium's Values In the Model (VIM) framework for clinical AI
Ethics and Policy & Governance add the RAISE consortium's VIM framework, which calls for AI systems to disclose how they navigate value-laden trade-offs such as intervention versus conservative management and autonomy versus paternalism (Goldberg et al., NEJM AI, 2026). Further Reading adds an entry on the RAISE consortium, including its MedLog proposal for event-level logging of clinical AI.
Why Extends fairness and bias-audit guidance with a framework for assessing what values a clinical AI system embeds, since physicians and patients cannot otherwise know which priorities it optimizes for.
Chapters Healthcare Policy and AI Governance · Medical Ethics, Bias, and Health Equity · Further Reading and Resources
LLM diagnostic error rates on complex case vignettes
Emerging Technologies reports that LLMs give incorrect diagnoses 17 to 19% of the time on complex case vignettes (Hirosawa et al., 2023), alongside a sourced figure for how often LLMs produce invented citations.
Why Adds a primary source for LLM diagnostic accuracy alongside the chapter's existing citation-fabrication statistic.
AMA survey on physician priorities for AI oversight and liability
LLMs in Practice adds AMA survey findings that 47% of physicians rank increased oversight as their top regulatory priority for AI tools, and 87% cite not being held liable for AI model errors as a critical factor for adoption.
Why Gives readers physician sentiment on regulatory priorities alongside adoption data for clinical AI tools.
WHO's 2025 guidance expands to environmental, labor, and liability concerns
Emerging Technologies, Global Health Equity, and Policy & Governance expand coverage of WHO's 2025 guidance on large multi-modal models, adding its concerns about environmental costs and labor exploitation in AI development, and its proposed liability frameworks and call for international coordination.
Why Extends the handbook's existing WHO guidance coverage to sustainability, labor, and liability questions raised by large AI models in health care.
Chapters Emerging AI Technologies in Healthcare · AI and Global Health Equity · Healthcare Policy and AI Governance
LLM training objective and limits on structured clinical prediction
LLMs in Practice clarifies that LLMs are trained to predict statistically likely next words, not to retrieve correct facts (WHO, 2025). It adds a finding that LLMs underperformed structured deep learning models on antimicrobial resistance prediction for sepsis (AUROC 0.74 versus 0.85), with combined approaches offering no improvement (Hixon et al., 2025, conference abstract).
Why Helps readers understand why LLMs generate confident but incorrect answers, and where structured prediction models remain preferable for clinical tasks.
Hixon et al. on structured EHR data outperforming LLMs for AMR prediction
AI Fundamentals, LLMs in Practice, and Infectious Diseases add a study of nearly 30,000 sepsis episodes across 10 hospitals finding that deep learning on structured EHR data predicted ceftriaxone-resistant Gram-negative infection better (AUROC 0.85) than LLMs analyzing clinical notes (AUROC 0.74), with no improvement from combining the two approaches.
Why Gives readers concrete evidence that LLMs are not always the right tool, and that tabular EHR-based models can outperform them for specific prediction tasks.
Chapters AI Fundamentals for Clinicians · Large Language Models in Clinical Practice · Infectious Diseases and Antimicrobial Stewardship
New evidence on AI-assisted clinical trial matching and enrollment
Research with AI adds that AI-assisted trial matching tools show improved enrollment but heterogeneous results (Lu et al., 2024), that such tools should complement rather than replace RCTs (Eichler et al., 2021), and that residual confounding remains a limitation even with increased enrollment diversity (Fewell et al., 2007).
Why Gives readers a calibrated view of AI's role in trial recruitment: helpful for enrollment, not a substitute for RCT rigor or human review.
Chapter Clinical Research with AI
Perincheri et al. on Paige Prostate's FDA clearance evidence
AI Tools for Physicians adds that Paige Prostate received FDA De Novo clearance in 2021 after the company demonstrated its prostate biopsy algorithm improves detection of high-grade cancer, describing it as an example of narrow scope with prospective validation.
Why Gives readers a primary source for a widely cited example of well-validated, narrowly scoped diagnostic AI.
Zeng et al. warn on DeepSeek's clinical risk from confident errors
Global Health Equity adds a JAMA perspective warning that DeepSeek tends to generate convincing but factually incorrect outputs, creating substantial clinical risk if deployed at scale without prospective validation.
Why Flags a named, widely used model's tendency toward confident errors as a specific clinical safety concern, not just a general LLM caveat.
Chapter AI and Global Health Equity
Byrd et al. and Stewart et al. on AI early warning systems
Internal Medicine adds that an AI early warning model showed acceptable discrimination (AUROC 0.76) for predicting cardiac arrest, ICU transfer, or death within 24 hours, and cites evidence for reduced severe hypoglycemia in the ICU with AI-assisted monitoring.
Why Gives readers quantified performance for a deterioration-prediction tool and a second use case for the same category of AI monitoring.
Thapa et al. on sleep-based prediction of dementia and Parkinson's
Neurology adds that foundation models trained on a single night of polysomnography data can predict dementia (C-index 0.85) and Parkinson's disease (C-index 0.89).
Why Gives readers an emerging, low-burden screening approach for neurodegenerative disease risk using data many patients already generate through sleep studies.
AI evidence updates for colonoscopy and prostate MRI in Surgical Subspecialties
Surgical Subspecialties cites a 44-RCT meta-analysis on AI-assisted colonoscopy (Soleymanjahi et al., 2024) and a trial finding higher adenoma detection with AI (54.8% versus 40.4%), and updates its PI-RADS Steering Committee AI requirements to the 2025 version, citing biparametric AI accuracy comparable to radiologists (88.4% versus 89.5%).
Why Gives readers updated, quantified evidence on where AI already changes screening outcomes in GI and urologic imaging.
Chapter Surgical Subspecialties
OpenEvidence adoption data and modern CDS evidence-synthesis tools
LLMs in Practice adds a section on modern LLM-based clinical decision support, reporting that OpenEvidence, launched in 2024, reached 45% daily use among U.S. physicians by late 2025 and handled over 18 million monthly clinical consultations (company-reported), and that UCSF integrated it as a SMART on FHIR app within Epic.
Why Gives readers usage figures for the fastest-growing clinical decision support tool, labeled as the company's own reporting rather than independent measurement.
New section on safety considerations for patient-facing AI
LLMs in Practice adds a section on patient-facing AI, covering evidence gaps, risk categories, a case example of ChatGPT used for health education, and guidance for physicians managing patients who bring AI-derived information to visits.
Why Gives physicians a framework for evaluating risk when patients use consumer AI tools for health information, separate from clinician-facing decision support.
New position statement on AI in surgical training
Surgical Subspecialties adds a 2026 position statement on AI in surgical training, noting the Society of Thoracic Surgeons offers educational content on AI/ML but no formal position.
Why Gives readers the first cited professional-society position on how AI should be used in surgical training.
Chapter Surgical Subspecialties
Global governance initiatives for AI in global health
Global Health Equity adds WHO, ITU, and WIPO's 2023 Global Initiative on AI for Health as an institutional framework for coordinating AI development, alongside a Lancet Global Health Commission synthesizing AI's economic and health impacts and Gates Foundation funding for AI equity and language inclusivity.
Why Gives readers the institutional bodies and funding sources shaping how AI is coordinated across global health systems.
Chapter AI and Global Health Equity
Open-weight LLMs for global health: benefits and validation risk
Global Health Equity adds evidence that open-weight models (DeepSeek, Llama 3, Mistral) match proprietary model performance at a fraction of the cost (Ong et al., 2026) and are deployed in 90 Chinese tertiary hospitals (Chen et al., 2025), alongside a JAMA warning that scaling without validation risks factually incorrect outputs (Wong et al., 2025).
Why Weighs the cost and access benefits of open-weight models against documented risks of deploying them clinically without validation.
Chapter AI and Global Health Equity
LLM-enhanced global health applications: DeepDR-LLM and malaria detection
Global Health Equity cites DeepDR-LLM as an LLM-enhanced global health application (Li et al., 2024) and transformer architectures applied to smartphone-based malaria detection from blood smears as a scalable alternative to conventional computer vision (Liu et al., 2023).
Why Gives readers primary sources for two applied uses of LLM-adjacent AI in resource-constrained clinical settings.
Chapter AI and Global Health Equity
AfriMed-QA dataset for African clinical AI evaluation
Global Health Equity adds AfriMed-QA, described as the first large-scale pan-African, multi-specialty medical question-answer dataset for evaluating LLM performance in African healthcare contexts, and publicly available for local researchers to evaluate and fine-tune models (Olatunji et al., 2025).
Why Addresses a gap: most LLM benchmarks lack large-scale evaluation data reflecting African clinical contexts.
Chapter AI and Global Health Equity
Diagnostic brittleness versus a successful early-warning AI system
Safety & Risk adds that diagnostic accuracy dropped 26 to 38 percentage points when familiar answer patterns were disrupted (Bedi et al., 2025), contrasted with a wearable AI system that predicted patient deterioration 8 to 24 hours before standard hospital alerts using continuous monitoring and prospective validation (Scheid et al., 2025).
Why Shows what separates working prediction systems from brittle ones: continuous monitoring and real-workflow validation rather than intermittent, standalone alerts.
OpenEvidence and Doximity GPT in AI Tools for Physicians
AI Tools for Physicians profiles OpenEvidence, a clinical decision support copilot the company reports is used daily by 40% of U.S. physicians with 18 million monthly consultations, and Doximity GPT, a free HIPAA-compliant documentation assistant with company-reported adoption among beta testers. A PMC-published study found OpenEvidence's recommendations aligned with physician treatment plans in primary care.
Why Adds two of the most widely used clinical AI tools, with vendor adoption figures explicitly labeled as company-reported rather than independently verified.
Resting-state fMRI for pre-surgical brain mapping
Neurology adds a section on resting-state fMRI brain mapping (Cirrus software, FDA 510(k) K251009), which maps functional networks without requiring task performance, reaching 87% usability versus 67% for task-based fMRI, and enabling mapping in pediatric, sedated, and non-English-speaking patients.
Why Addresses a documented gap: task-based fMRI fails in about a third of surgical candidates who cannot perform in-scanner tasks.
State of Clinical AI 2026 report: the jagged frontier in LLM diagnosis
LLMs in Practice, Safety & Risk, and Further Reading add the inaugural State of Clinical AI 2026 report (Stanford-Harvard ARISE Network): LLMs matched or outperformed physicians on fixed diagnostic cases but performed closer to medical students when reasoning under uncertainty (Johri et al., 2025; McCoy et al., 2025).
Why Adds a primary-source annual synthesis physicians can use to judge when benchmark claims of superhuman performance generalize to real uncertainty.
Chapters Further Reading and Resources · Large Language Models in Clinical Practice · Clinical AI Safety and Risk Management
State of Clinical AI 2026 report: an evaluation crisis
Evaluating AI Systems adds the report's review of more than 500 medical AI studies, finding nearly 50% tested models with exam-style questions, only 5% used real patient data, and few measured uncertainty recognition or bias.
Why Explains why benchmark performance may not predict real clinical utility, since most evaluations do not test the judgment calls physicians actually make.
State of Clinical AI 2026 report: human-AI collaboration evidence
Workflow Integration adds the report's synthesis of 2025 collaboration studies: AI-assisted radiologists detected more breast cancers without more false alarms, an RCT found AI-assisted physicians made better treatment decisions, and a Kenyan deployment cut errors across tens of thousands of visits. It also notes risks: clinicians following wrong AI advice performed worse, and low override rates may signal over-reliance.
Why Shows the report's central conclusion: outcomes depend on how AI is integrated into workflow, not just on model accuracy.
MedCalc-Bench: medical calculation errors among LLMs
LLMs in Practice and AI Tools for Physicians report that on MedCalc-Bench the best-performing model (GPT-4, one-shot) reached only 50.9% accuracy across 55 medical calculator tasks in 1,000+ scenarios, with errors from wrong equations, misextracted parameters, and faulty arithmetic. Claude's reported 98.1% accuracy on the same benchmark awaits independent validation.
Why Warns that LLMs can fail at routine clinical calculations that directly drive treatment decisions, such as anticoagulation and dosing thresholds.
Chapters Large Language Models in Clinical Practice · AI Tools Every Physician Should Know
AI agents in cancer research and precision oncology
Emerging Technologies and Oncology add a Nature Reviews Cancer primer on AI agents that plan and execute multi-step oncology workflows: drug design, treatment strategy, and clinical trial matching. Current agents reach 70-85% concordance with oncologist decisions, but prospective trials against standard care do not yet exist (Truhn et al., 2026).
Why Gives readers a primary-source review of agentic AI in oncology, distinguishing demonstrated concordance from still-unproven clinical benefit.
Chapters Emerging AI Technologies in Healthcare · Hematology-Oncology and Precision Medicine
Systematic review of LLM virtual patients for history-taking training
Medical Education adds a 2026 systematic review of 39 studies on LLM-powered virtual patients for history-taking, reporting top-k accuracy of 0.45 to 0.98 and low hallucination rates. The review found 95% of studies simulated only single disease types, with few covering multimorbidity.
Why Highlights a gap in AI training tools: virtual patients rarely represent the multimorbid patients common in real practice.
Chapter AI in Medical Education
Real-world RVU and revenue data for AI scribe adoption
Clinical Documentation adds findings from a study of 1.2 million encounters showing AI scribe adopters logged 1.81 more RVUs and 0.80 more encounters weekly than nonadopters, about $3,044 more annually per physician at 2025 Medicare rates, with no change in claim denials (Holmgren et al., 2026).
Why Answers whether ambient scribe adoption pays off financially, though it leaves open whether gains reflect more clinical work or coding changes.
EchoPrime multi-view echocardiography foundation model
Cardiology and Tools by Specialty describe EchoPrime, trained on 12,124,168 echocardiogram videos from 275,442 studies at Cedars-Sinai, validated across five health systems with mean AUC 0.85 to 0.92 on 17 classification tasks (Vukadinovic et al., 2025).
Why Gives readers a primary-source benchmark for a foundation model that synthesizes multiple echocardiogram views, trained on more than ten times the data of prior single-task models.
Chapters Cardiology and Cardiothoracic Surgery · AI Tools by Medical Specialty
Evidence base for clinical LLM prompting techniques
LLMs in Practice adds a new Part 5 comparing zero-shot, few-shot, chain-of-thought, and structured-template prompting for clinical tasks, citing studies on when each helps or degrades diagnostic reasoning, plus a practical R-C-T-C-F prompting framework.
Why Gives readers evidence-based guidance on which prompting techniques improve LLM accuracy in clinical tasks, and where chain-of-thought prompting can backfire.
Physician persona framing shifts LLM clinical recommendations
LLMs in Practice cites a Mount Sinai preprint testing 20 LLMs across 5 million clinical decisions, where assigned physician personas shifted affirmative action rates from 36.9% to 46.4% under identical clinical evidence.
Why Illustrates that system prompt design functions as a clinical policy choice with measurable effects on care recommendations, not a neutral setting.
The Physician AI Handbook registers a Zenodo DOI
References & How to Cite and Welcome add the handbook's Zenodo DOI (10.5281/zenodo.18251405), a DOI badge, and its CC BY 4.0 license terms.
Why Gives the handbook a permanent, citable identifier.
Chapters How to Cite This Handbook · Welcome
Memorization risks in healthcare foundation models
Privacy & HIPAA adds that healthcare foundation models trained on EHRs show memorization patterns that leave patients with rare conditions vulnerable to extraction attacks, and recommends combining differential privacy during training with pre-release evaluation, since no single technique is sufficient (Tonekaboni et al., 2025, preprint). The Glossary adds a Differential Privacy definition.
Why Gives readers primary evidence on a specific privacy risk unique to foundation models, plus practical mitigation guidance.
Chapters Privacy, HIPAA, and Patient Data Security · Glossary of Medical AI Terms
Penda Health LLM clinical copilot study added across three chapters
Primary Care, Workflow Integration, and LLMs in Practice add a preprint study of 39,849 patient visits at Penda Health, a Nairobi primary care provider, evaluating an LLM copilot (AI Consult) that flags concerns with a three-tier green, yellow, and red alert system for clinician review.
Why Gives readers early real-world evidence on LLM copilots functioning as a clinical safety net rather than just documentation support.
Chapters Primary Care, Family Medicine, and Preventive Medicine · Integration into Clinical Workflow · Large Language Models in Clinical Practice
Data leakage in MIMIC-based ICU prediction models
Evaluating AI Systems adds evidence that 40.2% of MIMIC-based same-admission prediction models used ICD codes as features despite those codes being finalized only after discharge, with ICD-only models reaching inflated AUROCs of 0.97 to 0.98 (Ramadan et al., 2025).
Why Warns readers about a specific, quantified form of data leakage that inflates reported model performance.
Intraoperative molecular imaging AI for lung nodule malignancy
Oncology adds a research-stage application combining machine learning with intraoperative molecular imaging to assess lung nodule malignancy during surgery in real time, citing a 2026 University of Pennsylvania cohort study (Azari et al., 2026).
Why Gives readers an early, research-stage example of AI-assisted intraoperative decision-making in thoracic oncology.
Google releases MedGemma open medical foundation models
Emerging Technologies adds MedGemma, Google's open-source medical AI model family released through the Health AI Developer Foundations program, available via Hugging Face and Vertex AI, alongside MedASR, a 105M-parameter speech-to-text model for clinical use (Google Research, January 2026).
Why Covers an open-weight alternative to proprietary medical foundation models like Med-PaLM.
Small factual refinements across foundations, policy, and specialty chapters
AI Fundamentals adds sector adoption rates for AI, noting 8.3% of healthcare firms use AI in producing goods or services compared to 11.6% in finance and 15.1% in education. Policy & Governance adds that 97% of AI devices are cleared through the FDA's 510(k) pathway (MedTech Dive, 2024).
Why Gives readers comparative context on healthcare AI adoption and regulatory pathway data cited elsewhere in the handbook.
Chapters AI Fundamentals for Clinicians · Healthcare Policy and AI Governance
AMIE diagnostic reasoning system evaluated against physicians
Emerging Technologies adds Google's AMIE diagnostic reasoning system, reporting that in a blinded evaluation of 302 challenging cases AMIE alone reached 59.1% top-10 diagnostic accuracy versus 33.6% for unassisted clinicians and 51.7% for AMIE-assisted clinicians, while noting the cases were curated for difficulty and the study was funded by Alphabet (McDuff et al., 2025).
Why Gives readers Nature-published evidence on LLM diagnostic performance, with the chapter's own caveats about generalizability and funding source.
AI telehealth triage matched physician guideline adherence
Primary Care adds a Cedars-Sinai study of 461 virtual urgent care visits in which AI recommendations matched or exceeded final physician recommendations on guideline adherence, though the AI was vendor-developed and concordance between initial AI and final physician recommendations was only 56.8% (Zeltzer et al., 2025).
Why Gives readers evidence on AI triage performance alongside the caveat that the tool has not been independently validated.
Chapter Primary Care, Family Medicine, and Preventive Medicine
Preprint on why patients turn to AI chatbots during mental health crises
Psychiatry adds a preprint study of 53 people with lived crisis experience and 16 mental health experts, finding people use AI chatbots to fill gaps in human support when professional help is unavailable, and recommending AI be designed as a bridge toward human connection rather than an endpoint (Ajmani et al., 2025, preprint).
Why Gives readers qualitative evidence on why patients use AI during crises and a design principle for crisis-response features.
Chapter Psychiatry and Behavioral Health
Anthropic launches Claude for Healthcare
AI Tools for Physicians adds Anthropic's Claude for Healthcare, an enterprise product with HIPAA support and connectors to CMS coverage data, ICD-10 codes, the NPI registry, and PubMed, reporting Banner Health's 22,000 clinicians and vendor-reported results (85% faster, with higher accuracy), alongside MedAgentBench context: 91.4% vendor-reported versus 69.67% in the independently published evaluation.
Why Covers a major vendor's entry into clinical AI tooling, labeling adoption and accuracy figures as vendor-reported and giving readers an independently published comparison point.
Practical guidance for clinical documentation
Clinical Documentation covers documentation burden and practical best practices: pilot testing before rollout, vendor due diligence, physician training requirements, quality monitoring, and patient transparency, plus recommended and not-recommended practice settings.
Why Cites updated evidence that AI scribe adoption reduces documentation time only modestly and that returns vary by baseline documentation burden.
FDA reviews AI mental health chatbots amid lawsuits
Psychiatry adds that the FDA's Digital Health Advisory Committee met in November 2025 to discuss generative AI mental health devices, raising concerns about confabulation and bias without issuing binding rules, alongside 2024 to 2025 lawsuits alleging AI chatbots contributed to a minor's suicide.
Why Shows regulatory and legal scrutiny of AI mental health chatbots is beginning, though the FDA committee reached no binding rules.
Chapter Psychiatry and Behavioral Health
Ambient documentation adoption and billing oversight concerns
Clinical Documentation adds that 31.5% of U.S. hospitals had deployed EHR-integrated generative AI by 2024, plus a new Emerging Policy Concerns section on ambient scribe vendors expanding into billing optimization; researchers propose a national AI-use registry, stronger Medicare claims auditing, and reimbursement adjustments similar to the existing 5.9% Medicare Advantage coding-intensity adjustment.
Why Flags a policy risk physicians should know as ambient scribes move from note-taking into billing optimization, ahead of any proposed safeguards being adopted.
Healthcare AI adoption lags other sectors
AI Fundamentals for Clinicians adds that 8.3% of healthcare firms used AI in producing goods or services as of 2025, versus 11.6% in finance, 15.1% in education, and 23.2% in information services; ambulatory care (8.7%) is adopting faster than nursing facilities (4.5%).
Why Gives physicians a firm-level adoption baseline far lower than self-reported personal use, clarifying that organizational deployment lags individual adoption.
Chapter AI Fundamentals for Clinicians
Chatbot responses to suicidal and psychiatric crises
Psychiatry adds evidence that AI chatbots often mishandle psychiatric crises: a Stanford/CMU study found LLMs responded appropriately only 45% of the time to delusions and 80% to suicidal ideation; a national survey found 13% of U.S. adolescents and young adults, about 5.4 million people, have used generative AI for mental health advice.
Why Gives psychiatrists quantified evidence, not just anecdote, that chatbots handle crisis scenarios inconsistently even as many young people already use them for mental health advice.
Chapter Psychiatry and Behavioral Health
LEADS: a fine-tuned LLM for systematic review screening
Research with AI adds LEADS, a fine-tuned Mistral-7B model for systematic reviews: in a study with 16 clinicians, human-AI collaboration improved recall (0.81 versus 0.78) with 20.8% time savings for screening, and improved extraction accuracy (0.85 versus 0.80) with 26.9% time savings; the model is not commercially available.
Why Demonstrates the potential of domain-specific LLM fine-tuning for literature screening, though the tool itself remains a research prototype.
Chapter Clinical Research with AI
Differential diagnosis generators: DXplain and Isabel Healthcare
AI Tools for Physicians adds evidence for DXplain, in clinical use since the 1980s, and Isabel Healthcare: a comparative study scored both 3.45 out of 5 for diagnostic accuracy, tied for the top-performing generators; Isabel's vendor reports 96% accuracy for including the correct diagnosis among its top suggestions.
Why Gives physicians comparative accuracy data for two long-running differential diagnosis tools, distinguishing independently studied results from vendor-reported figures.
Acute care AI: deterioration monitoring and stroke imaging
AI Tools for Physicians adds evidence for WAVE Clinical Platform (FDA-cleared 2018 patient surveillance, detecting instability 6.33 hours earlier and cutting critical events 58% in a UPMC study) and stroke and hemorrhage imaging tools Aidoc, Viz.ai, and RapidAI, citing validation studies with sensitivity above 90% for intracranial hemorrhage and a 39% reduction in door-to-groin time for stroke coordination.
Why Gives physicians FDA clearance status and outcome data for time-critical inpatient and stroke AI tools now widely deployed across hospitals.
Chest radiograph and mammography screening AI tools
AI Tools for Physicians adds evidence for chest radiograph triage systems Lunit INSIGHT CXR, Oxipit ChestLink, and qXR, and mammography systems iCAD ProFound AI, Lunit INSIGHT MMG, and Hologic Genius AI, citing validation studies reporting sensitivities from 92% to 100% and workload reductions of up to 36%.
Why Helps physicians compare independently published sensitivity and workload data across competing chest and breast imaging AI products.
Cardiac imaging AI: Arterys and HeartFlow FFR-CT
AI Tools for Physicians adds evidence for Arterys cardiac MRI and CT angiography analysis, validated in a multi-vendor comparison, and HeartFlow FFR-CT, citing the PRECISE trial (n=2,103, reduced unnecessary catheterization without increased death or MI) and the PACIFIC trial (87% sensitivity, 81% specificity, AUC 0.93).
Why Gives cardiologists trial-level evidence, not just vendor claims, for two widely used cardiac imaging AI platforms.
Pathology AI: Paige Prostate and PathAI's AIM-MASH
AI Tools for Physicians adds evidence for Paige Prostate, which received FDA De Novo clearance in 2021 for prostate biopsy cancer detection, and PathAI's AIM-MASH, which received FDA Drug Development Tool Qualification in December 2025, the first AI pathology tool qualified for use in MASH clinical trials.
Why Highlights two pathology AI tools with regulatory milestones, FDA clearance and the first FDA qualification of its kind, rather than only vendor marketing claims.
FTC health-data enforcement actions beyond HIPAA
Privacy & HIPAA adds an FTC Enforcement section on Section 5 and the Health Breach Notification Rule, with a table of actions against consumer health apps: GoodRx ($1.5 million, 2023), BetterHelp ($7.8 million consumer refund, 2023), Flo Health (consent order, 2021), and Premom ($100,000, 2023), all for sharing health data with advertisers without consent.
Why Shows physicians that consumer health apps they may recommend, such as period trackers or mental health platforms, typically fall outside HIPAA but remain subject to FTC enforcement.
FDA general wellness guidance and consumer wearables
Policy & Governance adds a section on FDA's General Wellness guidance (updated January 2026), which exempts low-risk products making only wellness claims from device regulation; the update addresses optical sensing (photoplethysmography) for blood pressure, oxygen saturation, glucose, and heart rate variability in wearables.
Why Helps physicians distinguish consumer wearables making general wellness claims from FDA-regulated medical devices, especially as optical sensors add new physiologic measurements.
Petri and Bloom: Anthropic's open-source LLM safety audit tools
AI Tools for Physicians adds Petri and Bloom, Anthropic's open-source tools for auditing LLM behavior such as sycophancy, deception, and crisis handling through simulated conversations; the chapter notes results may favor Claude models since Anthropic built both tools.
Why Gives physicians a primary source for two safety-auditing tools increasingly referenced when comparing LLM behavior across vendors.
First quantitative model of AI's effect on radiologist workforce
Radiology & Nuclear Medicine adds a task-based model projecting a 33% reduction in radiologist hours over 5 years (range 14 to 49%), driven mainly by AI-assisted report drafting and delegation of mammography and radiography studies; it projects workflow shifts rather than job loss because imaging volumes continue to grow.
Why Gives radiologists a primary, though not yet peer-reviewed, source for workforce impact estimates that are frequently discussed without quantitative grounding.
Platform crisis safeguards and age verification limits
Psychiatry adds detail on platform crisis safeguards: Anthropic's Claude uses a classifier to detect suicide or self-harm conversations and directs users to ThroughLine's network of over 170 crisis helplines (including 988 and Samaritans); OpenAI has similar interventions. Age verification for both platforms relies on self-attestation and can be circumvented.
Why Helps psychiatrists counsel patients on the limits of platform-level safeguards, including age verification that relies on self-attestation and is easily circumvented.
Chapter Psychiatry and Behavioral Health
OpenAI for Healthcare: enterprise platform with BAA support
AI Tools for Physicians adds an Enterprise AI Platforms section covering OpenAI for Healthcare (January 2026), including ChatGPT for Healthcare and API access with a Business Associate Agreement, citing a preprint evaluation of a deployment with Penda Health in Kenya. UCSF's chancellor confirmed a 9,000-user ChatGPT Enterprise deployment for early 2026.
Why Distinguishes HIPAA-compliant enterprise deployments from the consumer ChatGPT Health product, relevant as health systems evaluate BAA-backed AI options.
Chapters AI Tools Every Physician Should Know · Large Language Models in Clinical Practice
More detail on ChatGPT Health's data connections
AI Tools for Physicians adds detail on ChatGPT Health's data connections: medical records via b.well (U.S. only, covering Epic, Cerner, and Meditech systems), wellness apps (Apple Health, MyFitnessPal, Weight Watchers, AllTrails, Instacart, Peloton), and insurance navigation; OpenAI reports 1.6 to 1.9 million health insurance questions answered weekly.
Why Gives physicians specifics on which health records and apps ChatGPT Health can access, useful when patients raise the product during visits.
Utah approves first state AI prescription-renewal pilot
Policy & Governance adds a State Regulatory Sandboxes section describing Utah's approval of an AI prescribing pilot with Doctronic for routine chronic-condition refills, the first such state authorization; Case Studies adds a matching ongoing case study. Deseret News reports 99.2% agreement between the AI's treatment plans and physician decisions in testing.
Why State-level sandboxes are emerging as a path around federal AI regulation gaps, and the pilot raises questions the AMA has already flagged about physician oversight of automated prescribing.
Chapters Healthcare Policy and AI Governance · Clinical Case Study Library
SleepFM: a sleep-signal foundation model for disease prediction
Emerging Technologies adds SleepFM, a transformer trained on over 585,000 hours of polysomnography from about 65,000 participants, which predicts all-cause mortality (C-index 0.84), dementia (0.85), and Parkinson's disease (0.89) from a single night of sleep. The finding needs prospective validation.
Why Extends foundation-model methods beyond text and imaging to physiological signals, suggesting a low-cost early marker for neurodegenerative disease pending prospective testing.
Chapters Emerging AI Technologies in Healthcare · Neurology and Neurological Surgery
Validation gaps across 130 FDA-cleared AI devices
Evaluating AI Systems adds findings from an analysis of 130 FDA-approved AI devices (2015 to 2020): 97% were evaluated only with retrospective data, under 13% of summaries reported patient demographics, and externally tested algorithms performed about 10% worse.
Why Gives physicians primary evidence that FDA clearance does not guarantee prospective validation or demographic representativeness.
Trial Pathfinder study on cancer trial eligibility criteria
Research with AI adds Trial Pathfinder, an analysis of eligibility criteria using real-world data from 61,094 non-small-cell lung cancer patients, finding many exclusion criteria had minimal effect on treatment hazard ratios and that data-driven broadening more than doubled the eligible patient pool.
Why Shows how AI analysis of real-world data can identify overly restrictive trial criteria, relevant given only 3 to 5% of cancer patients currently enroll in trials.
Chapter Clinical Research with AI
EchoNet-Dynamic: randomized trial of echocardiogram AI
Cardiology adds EchoNet-Dynamic, an FDA-cleared, video-based deep learning system for beat-to-beat ejection fraction assessment, validated in a randomized trial and shown to trace left ventricular borders more accurately than human sonographers (mean absolute error 4.1%). The associated dataset of 10,030 labeled echocardiogram videos is openly available.
Why One of the first echo AI systems validated in a randomized trial, and its open dataset allows external validation that most proprietary systems lack.
HealthBench: OpenAI's health-specific evaluation framework
Evaluating AI Systems adds HealthBench, an open-source benchmark built with 262 physicians across 60 countries and 26 specialties, using conversation-specific rubrics rather than multiple-choice questions: 5,000 multi-turn conversations scored against 48,562 rubric criteria.
Why Gives physicians a primary source for a benchmark increasingly cited to support health AI product claims, rather than relying on multiple-choice exam scores.
OpenAI launches ChatGPT Health for consumer use
AI Tools for Physicians adds a Consumer Health AI section on OpenAI's ChatGPT Health, which connects users' medical records and wellness apps (Apple Health, MyFitnessPal, Peloton) for personalized health conversations; OpenAI reports over 230 million weekly health-related queries.
Why OpenAI's launch of a direct-to-consumer health product means physicians may increasingly encounter patients bringing its output into clinical encounters.
Chapters AI Tools Every Physician Should Know · Large Language Models in Clinical Practice · Emerging AI Technologies in Healthcare
History of AI in medicine: publication growth
History of AI in Medicine adds that AI-related medical publications increased 36-fold between 2000 and 2022, from 1,614 to over 58,000 annually, framing current AI activity in medicine within multiple earlier cycles of enthusiasm followed by underperformance over seven decades.
Why Grounds the chapter's historical framing in a quantified measure of how sharply AI-related medical research activity has accelerated since 2000.
Chapter History of AI in Medicine
Capacity-building frameworks for sustainable AI in LMICs
Global Health Equity adds that 80% of health AI pilots in low- and middle-income countries fail to survive the transition from donor funding to local operation, and describes the RAD-AID Three-Pronged Strategy and a Johns Hopkins consensus framework for implementing AI with, not on, local workforces.
Why Gives readers a structural explanation, and named frameworks, for why most AI pilots in low-resource settings fail to scale beyond initial funding.
Chapter AI and Global Health Equity
Validation gaps in machine learning for antimicrobial resistance prediction
Infectious Diseases adds a 2025 Lancet Infectious Diseases Series assessing machine learning approaches to antimicrobial resistance prediction across pathogens, finding substantial gaps between research promise and clinical readiness, and a companion framework explaining why outbreak prediction faces fundamental validation challenges.
Why Tempers expectations for resistance-prediction AI by documenting pathogen-by-pathogen validation status rather than treating the technology as clinically ready.
Surgical Subspecialties surveys AI guidance across GI and urology societies
Surgical Subspecialties adds that the AGA published its first guideline evaluating AI in GI care in 2025, making no recommendation for or against CADe-assisted colonoscopy due to very low certainty of evidence, while noting a modest adenoma detection rate improvement (44.8% versus 37.4%). The chapter also catalogs which surgical societies, including the AUA, have not issued formal AI position statements.
Why Gives surgeons the current, cautious state of professional-society guidance on AI-assisted colonoscopy, and which societies have not yet weighed in.
Chapter Surgical Subspecialties
New chapter: AI in Medical Education
A new Medical Education chapter covers undergraduate and continuing medical education curricula for AI, citing a 2025 scoping review that found 310 publications on AI in medical education, with 52% appearing after ChatGPT's November 2022 release, plus consensus frameworks defining AI competencies for physicians.
Why Gives medical educators an evidence-based curricular framework as schools develop AI teaching without an established evidence base for what to teach or when.
Chapter AI in Medical Education
Psychiatry: historical framing and evidence maturity gap for AI mental health tools
Psychiatry adds historical framing tracing AI-based mental health interventions back to foundational work in the 2010s, and notes that a 2025 systematic review found low evidence maturity among such tools.
Why Situates current AI mental health tools within a longer development history and flags that most such tools still lack mature evidence.
Chapter Psychiatry and Behavioral Health
VA/DoD and NICE guidelines now advise against suicide risk scales
Adds a section to the Psychiatry chapter on international consensus against suicide-prediction risk scales. The VA/DoD 2024 guideline states such algorithms are correct only about 1% of the time among those flagged at risk, and UK, Australian, and New Zealand guidelines advise against using risk-assessment scales to predict or allocate treatment for suicide.
Why Clinical guidelines increasingly reject suicide-prediction algorithms as unreliable, directly bearing on how physicians should weigh AI risk-scoring tools in practice.
Chapter Psychiatry and Behavioral Health
Sickle cell disease AI prediction models: new evidence added
Expands the Oncology chapter's sickle cell disease AI section with four new studies: organ failure prediction (96% sensitivity, 98% specificity at 6 hours), acute kidney injury prediction (AUROC 0.91 at 12 hours), a VOC biomarker model (90% sensitivity), and a readmission model outperforming the LACE index (C-statistic 0.77 vs 0.60).
Why Adds a limitations section alongside quantified performance data so readers can weigh clinical implementation challenges for each prediction model.
Chatbot mental health tools: efficacy testing gap noted in new reviews
Updates the Psychiatry chapter's chatbot therapy evidence, citing a World Psychiatry review finding only 16% of LLM chatbot studies included clinical efficacy testing, a Current Psychiatry Reports review concluding AI currently supports clinicians rather than replacing them, and a World Psychiatry review of digital mental health navigation.
Why Gives physicians current data on how much chatbot mental health research has actually been clinically validated before recommending these tools.
Chapter Psychiatry and Behavioral Health
2025 federal AI policy shift: executive orders and state preemption
Adds a section to the Policy and Governance chapter describing Executive Order 14179's rescission of the prior AI oversight framework, the July 2025 AI Action Plan's proposed regulatory sandboxes, and a December 2025 executive order establishing federal preemption of state AI laws.
Why Federal AI regulation shifted toward deregulation and state preemption in 2025, increasing the importance of institutional governance for physicians.
References, links, and wording were also kept current.
December brought the handbook's expansion of its specialty section from 12 to 19 chapters, a new Executive Summary chapter, and a relicensing to CC BY 4.0. Coverage included the Epic Sepsis Model's external validation (33% sensitivity, 12% PPV), PathAI's research-use-only AIM-HER2 algorithm, the American College of Physicians' AI position paper, WHO's 2024 large multi-modal model guidance, and DermaSensor's FDA clearance.
AI-induced psychosis: case report and Danish cohort study
Adds glossary definitions for AI-induced psychosis, chatbot psychosis, and sycophancy, and cites new evidence in the Safety and Risk and Psychiatry chapters. Sources include a UCSF case report of new-onset psychosis after intensive ChatGPT use, a Danish study of 38 psychiatric patients whose chatbot use worsened delusions, and OpenAI's reported 0.07% weekly prevalence of psychosis or mania signals.
Why New case-report and cohort evidence on chatbot-linked psychotic symptoms gave the handbook clinical grounding to define these terms for psychiatrists and other physicians.
Chapters Glossary of Medical AI Terms · Clinical AI Safety and Risk Management · Psychiatry and Behavioral Health
Pierson et al. on AI-detected racial bias in knee pain assessment
Adds a case study to the Ethics and Health Equity chapter describing an AI model trained on patient-reported pain that explained 4.7 times more of the racial disparity in knee osteoarthritis pain than standard radiographic grading. Adds a matching glossary term (Bedside to Bench).
Why Nature Medicine research on AI's ability to detect overlooked pain drivers gives the Ethics chapter a concrete example of AI reducing rather than amplifying diagnostic bias.
Chapters Medical Ethics, Bias, and Health Equity · Glossary of Medical AI Terms · Orthopedic Surgery and Physical Medicine
Clinical AI evaluation frameworks added to vendor toolkit
Adds five external evaluation frameworks (APPRAISE-AI, CLIX-M, ROBUST-ML, the APA AI Tool Evaluation Checklist, and CASoF) and the AMA STEPS Forward AI Governance Toolkit to the Vendor Evaluation appendix.
Why Points clinicians evaluating AI tools to established, peer-reviewed scoring instruments rather than the handbook's checklist alone.
Chapter AI Vendor Evaluation Framework
Research Data Platforms added to Further Reading
Adds a Research Data Platforms section to the Further Reading appendix, listing Nightingale Open Science (over 400,000 de-identified medical images linked to outcomes) and PhysioNet (physiological signal and critical care datasets including MIMIC-III/IV).
Why Gives readers direct links to the patient-outcome datasets behind bedside-to-bench research such as the knee pain study.
Chapter Further Reading and Resources
Generative AI and LLM use policy template added
Adds a ready-to-use generative AI and LLM use policy template to the Policy Templates appendix, covering tools such as ChatGPT, Claude, and Copilot in clinical settings.
Why Gives physicians and administrators a starting template for governing everyday generative AI tool use in clinical settings.
Chapter Clinical AI Policy Templates
FDA AI device counts refreshed for December 2025
Updates the Tools by Specialty appendix with the FDA's December 2025 tally: 1,356 authorized AI-enabled medical devices, 1,039 in radiology (77%), 97% cleared via 510(k), and only 2 CPT codes covering AI reimbursement.
Why Keeps the device count and reimbursement gap current as the FDA's authorized AI device list continues to grow.
Chapter AI Tools by Medical Specialty
WHO 2024 guidance grounds new international governance and liability sections
Policy and Governance adds an international governance section on multilateral coordination and current gaps (no AI treaty, voluntary commitments, regulatory arbitrage), and Liability and Compliance adds an international liability frameworks section covering presumption of causality, strict liability, and no-fault compensation funds, both citing WHO's 2024 guidance on large multi-modal models.
Why Gives readers the international regulatory context and alternative liability models beyond the U.S. malpractice framework.
Chapters Healthcare Policy and AI Governance · Physician AI Liability and Regulatory Compliance
Environmental and labor impact of AI foundation models added
Emerging Technologies adds a carbon and water footprint comparison for training and running foundation models (for example, training a large medical LLM at 300 to 600 tonnes of CO2 equivalent), and Global Health Equity adds a section on data-labeling labor conditions in the AI supply chain, including pay of $1 to $2 per hour in some outsourcing markets.
Why Extends the guidance's risk framework beyond clinical accuracy to the environmental and human costs of building these models.
Chapters Emerging AI Technologies in Healthcare · AI and Global Health Equity
WHO 2024 guidance risk concepts added: solutionism, de-skilling, peer review
Safety and Risk adds a section on technological solutionism illustrated by the Epic sepsis model's marketed 85% sensitivity versus its externally validated 33%, Workflow Integration adds WHO's cognitive de-skilling concern, LLMs in Practice explains next-word prediction over fact retrieval, and Research with AI adds a section on risks to peer review integrity, including fabricated citations.
Why Applies WHO's 2024 risk taxonomy for large multi-modal models to clinical workflow, documentation, and medical research settings.
Chapters Clinical AI Safety and Risk Management · Integration into Clinical Workflow · Large Language Models in Clinical Practice · Clinical Research with AI
Mammography AI shown non-inferior to double reading in RCT evidence
Radiology and Nuclear Medicine adds trial evidence that AI-assisted mammography reading is non-inferior to double-reading by radiologists with a 40 to 50% reduction in workload, including the Swedish ScreenTrustCAD trial that randomized 80,000 women to AI-supported versus standard double reading (Lång et al., 2023).
Why Gives radiologists randomized trial evidence for a specific workload-reduction strategy as double reading becomes harder to staff.
AI Fundamentals chapter adds terminology and bias concepts
AI Fundamentals for Clinicians adds sections translating AI jargon into clinical concepts, including prediction versus inference, a quick reference for AI types, translating vendor claims, and pitfalls of label leakage, temporal leakage, and automation bias.
Why Gives physicians a framework for evaluating AI vendor claims and recognizing common failure modes in model validation.
Chapter AI Fundamentals for Clinicians
Medical LLM training and terminology explained: RLHF, LLM vs LMM
History of AI in Medicine adds a four-step explanation of how medical-specific models improve through reinforcement learning from human feedback, and AI Fundamentals for Clinicians adds a terminology section distinguishing large language models from large multi-modal models.
Why Gives physicians the vocabulary and technical background to follow how medical AI models are trained and differentiated.
Chapters History of AI in Medicine · AI Fundamentals for Clinicians
Cardiology chapter cites TIMI and GRACE risk score studies
Cardiology adds citations for three validated risk scores: the TIMI Risk Score for STEMI 30-day mortality (Morrow et al., 2000), the TIMI Risk Score for UA/NSTEMI risk stratification (Antman et al., 2000), and the GRACE Score for in-hospital and 6-month mortality in acute coronary syndromes (Granger et al., 2003).
Why Gives the chapter primary sources for three widely used cardiac risk stratification tools.
Oncology chapter details FDA-cleared genomic profiling platforms
Oncology adds detail on FoundationOne CDx (FDA-approved companion diagnostic profiling 324 genes for 28 or more targeted therapies) and Tempus xT (a 648-gene panel for solid tumors, FDA 510(k) cleared), citing evidence on machine learning treatment recommendation platforms.
Why Grounds the chapter's genomic profiling platform descriptions in their FDA clearance records and outcome research.
Clinical evidence and limits across foundation and specialty chapters
Clinical Data Challenge, Critical Care, Emergency Medicine, Pediatrics, and Radiology and Nuclear Medicine cover central evidence and limitations, including the Kaiser Permanente Neonatal Sepsis Calculator's reported 48% reduction in unnecessary antibiotic exposure across over 600,000 births.
Why Gives readers chapter-level context on what evidence supports each specialty's AI applications before the detailed sections.
Chapters The Clinical Data Challenge · Critical Care and Pulmonary Medicine · Emergency Medicine · Pediatrics and Neonatology · Diagnostic Imaging, Radiology, and Nuclear Medicine
Autism screening device identified as Cognoa Canvas Dx, FDA cleared
Pediatrics identifies the autism screening tool as Cognoa Canvas Dx, citing its FDA De Novo clearance (DEN200069, June 2021) as the first authorized autism diagnosis aid, with pivotal study results of 425 patients showing 81% sensitivity and 98% specificity.
Why Grounds the autism screening tool's claims in its FDA regulatory clearance record and pivotal study data.
Chapter Pediatrics and Neonatology
Radiology chapter adds specific pneumothorax detection metrics
Radiology and Nuclear Medicine adds sensitivity of 94% and specificity of 92% across 4 US hospitals for pneumothorax detection, noting the results exceeded FDA benchmarks for triage devices, and adds that a 603,000 chest X-ray study found faster oxygen therapy initiation when CAD is paired with electronic alerts to clinicians.
Why Gives the chapter's pneumothorax detection claims precise, study-level figures instead of general ranges.
Ancker et al. on alert override rates and fatigue thresholds
Evaluating AI Systems adds a cited benchmark: alert override rates above 90% and more than 5 to 10 alerts per patient per shift are associated with alert fatigue, applied to a worked example comparing a hospital's clinical decision support alert volume against the benchmark.
Why Gives the chapter's alert fatigue discussion a primary source for the numeric thresholds it uses.
Over 950 FDA-cleared AI devices by late 2024; how physicians and AI work together
Future Perspectives describes how physicians and AI work together, noting that the FDA has cleared over 950 AI medical devices as of late 2024.
Why Connects physician-AI collaboration to the growing use of AI medical devices.
Neurology chapter adds BrainLab and Persyst tool specifics
Neurology adds detail on two FDA-cleared tools: BrainLab's automated brain tumor segmentation, which cuts radiation planning time from 6 to 8 hours to 30 to 60 minutes, and Persyst's automated seizure detection, reported at 92% sensitivity for generalized tonic-clonic seizures on long-term EEG monitoring.
Why Gives readers specific performance and workflow figures for two named neurology AI tools already in clinical use.
OB/GYN AI evidence, maternal health equity, and ethical considerations
OB/GYN separates obstetric and gynecologic applications with per-application accuracy figures (for example, preterm birth prediction AUC 0.70 to 0.80 but positive predictive value of only 10 to 20%) and adds sections on racial bias in maternal-fetal AI, ethical challenges, and future directions.
Why Expands the chapter with application-specific evidence and adds coverage of equity and ethical considerations in maternal-fetal AI.
Chapter Obstetrics and Gynecology
Study on AI-assisted colonoscopy skill decline
Adds a citation in the history chapter to a multicentre observational study finding that endoscopists' adenoma detection rate dropped from 28.4% to 22.4% when performing colonoscopies without AI assistance after months of AI-assisted practice.
Why Gives readers primary evidence that reliance on AI assistance may be linked to a decline in unaided skill.
Chapter History of AI in Medicine
History chapter opens with concrete MYCIN example
AI in Medicine: A Brief History opens with MYCIN, which reached 90.9% diagnostic accuracy yet was not used in clinical care. The chapter gives the FDA's AI device count as hundreds, with a link to the FDA's AI/ML database for the current figure.
Why Shows readers a system that reached high diagnostic accuracy yet never entered clinical care, and gives the FDA database as the source for the current device count.
Chapter History of AI in Medicine
CC BY 4.0 terms on the license, references, and citation pages
States Creative Commons Attribution 4.0 International (CC BY 4.0) terms on the license, references, and how-to-cite pages: share, adapt, and commercial use permitted with attribution.
Why Gives readers and institutions clearer, more permissive reuse rights for the handbook's content.
Chapters Welcome · How to Cite This Handbook · References
IOM's To Err Is Human linked to AI safety framing
Adds a new section, "AI Safety Within the Patient Safety Tradition," to Clinical AI Safety and Risk Management, plus context in the history, workflow, and failures-appendix chapters, framing AI failures as system failures using the Institute of Medicine's 1999 To Err Is Human report (44,000 to 98,000 annual hospital deaths from medical errors; Kohn et al., 2000).
Why Grounds AI safety in the established patient-safety tradition rather than treating it as a new problem.
Chapters History of AI in Medicine · Clinical AI Safety and Risk Management · Integration into Clinical Workflow · AI Failures in Medicine: Lessons Learned
NOTA test evidence on LLM benchmark pattern matching
Adds a section on LLM-specific evaluation to Evaluating AI Clinical Decision Support Systems and a case study to the failures appendix on the NOTA test, where GPT-4o, Claude-3.5 Sonnet, and Llama-3.3 accuracy dropped 26 to 38 percentage points when familiar answer patterns were disrupted (Bedi et al., 2025), and cites the same finding in five more chapters.
Why A 2025 JAMA Network Open study found benchmark accuracy can reflect pattern matching rather than clinical reasoning.
Chapters Evaluating AI Clinical Decision Support Systems · AI Failures in Medicine: Lessons Learned · AI Fundamentals for Clinicians · Clinical AI Safety and Risk Management · Large Language Models in Clinical Practice · Emerging AI Technologies in Healthcare
Clinical AI evaluation and common validation pitfalls
Evaluating AI Clinical Decision Support Systems covers the Epic sepsis model, whose 33% sensitivity meant missing 67% of cases, and the LLM benchmark gap. It sets out an evidence hierarchy and names common validation pitfalls: selection bias, temporal bias, site overfitting, label leakage, publication bias, and outcome shifts.
Why Gives the chapter a stronger evidentiary framework for separating vendor marketing from validated performance.
Surgical AI by specialty and regulatory pathway
Surgery, Anesthesiology, and Perioperative Care covers specialty-specific applications (general, orthopedic, neuro, cardiac, thoracic, vascular, plastic surgery) and a regulatory and medicolegal section covering FDA regulation, liability, ACS-recognized applications, and SAGES guidelines.
Why Groups surgical AI evidence by specialty and regulatory pathway rather than by generic use case.
NEJM analyses of AI liability and LLM reproducibility risk
Adds a dual liability framework, exposure for misusing AI and for not using it, to the glossary, radiology, ethics, and liability chapters, and a new section on LLM output reproducibility as a liability and documentation issue.
Why NEJM papers formalized this liability exposure and the reproducibility problem with LLM outputs, giving physicians primary legal-risk sources.
Chapters Glossary of Medical AI Terms · Diagnostic Imaging, Radiology, and Nuclear Medicine · Medical Ethics, Bias, and Health Equity · Physician AI Liability and Regulatory Compliance · Large Language Models in Clinical Practice · AI-Assisted Clinical Documentation
ACP position paper on AI in health care
Adds the American College of Physicians position paper "Artificial Intelligence in the Provision of Health Care" (Annals of Internal Medicine, June 2024) to Internal Medicine and the resources appendix, with a link to the ACP's AI Resource Hub.
Why Gives readers the primary source for the ACP's official recommendations on AI in patient care.
Chapters Internal Medicine and Hospital Medicine · Further Reading and Resources
Meta-analysis of AI-assisted colonoscopy for polyp detection
Adds a new section on AI-assisted colonoscopy and pathology workflow impact to Pathology and Laboratory Medicine, citing a systematic review and meta-analysis of AI polyp detection (Soleymanjahi et al., 2024, Annals of Internal Medicine).
Why Gives readers a primary evidence source for AI-assisted polyp detection performance.
DynaMedex AI clinical decision support tool listing
Adds DynaMedex with Dyna AI, a clinical decision support tool combining DynaMed and Micromedex with AI, to the physician toolkit chapter and the resources appendix.
Why Expands the toolkit with an AI-enabled point-of-care reference tool.
Chapters AI Tools Every Physician Should Know · Further Reading and Resources
Cognitive de-skilling and AI anchoring bias added across four chapters
Integration into Clinical Workflow adds a framework for optimal clinician-AI collaboration timing and describes cognitive de-skilling and anchoring-bias risks; AI in Medicine: A Brief History, Diagnostic Imaging and Radiology, and Clinical AI Safety and Risk Management add matching sections with historical skill-atrophy parallels and CAD evidence; the glossary adds definitions for anchoring bias, automation bias, and cognitive de-skilling.
Why Grounds a workflow risk, over-reliance on AI recommendations and erosion of independent diagnostic skill, in historical automation parallels and published evidence of skill atrophy.
Chapters History of AI in Medicine · Diagnostic Imaging, Radiology, and Nuclear Medicine · Clinical AI Safety and Risk Management · Integration into Clinical Workflow · Glossary of Medical AI Terms
Handbook relicensed under CC BY 4.0; DermaSensor evidence added
The handbook changes its license from CC BY-NC-ND 4.0 to CC BY 4.0, permitting commercial use, derivatives, and translations with attribution; Dermatology adds DermaSensor (2024), an elastic-scattering-spectroscopy device with FDA De Novo clearance and DERM-SUCCESS pivotal-trial data showing 96% sensitivity across skin cancers, and distinguishes FDA-cleared from FDA-exempt tools.
Why Opens the handbook to commercial and derivative use with attribution, and gives readers pivotal-trial evidence for the first AI device cleared for primary-care skin cancer evaluation.
Chapters Welcome · Dermatology
FDA-cleared skin-cancer devices added to appendix; Nevisense evidence expanded
The specialty tools appendix adds a table of FDA-cleared dermatology AI devices (DermaSensor, Nevisense) and consumer apps, noting sensitivity of 7 to 73% for smartphone melanoma apps; Dermatology expands Nevisense coverage with its 2017 PMA approval and Malvehy et al. (2014) pivotal-trial data showing 96.6% sensitivity and 34.4% specificity across 2,416 lesions.
Why Gives readers a single reference table for FDA clearance status across dermatology AI tools and primary trial data for an FDA-approved melanoma-detection device.
Chapters AI Tools by Medical Specialty · Dermatology
New clinical specialty chapters
Adds chapters for critical care and pulmonary medicine, psychiatry and behavioral health, orthopedic surgery, infectious diseases, allergy/immunology/genetics, and surgical subspecialties, with separate chapters for dermatology and ophthalmology.
Why Broadens coverage to more ACGME-recognized specialties.
Chapters Welcome · Critical Care and Pulmonary Medicine · Psychiatry and Behavioral Health · Dermatology · Ophthalmology · Infectious Diseases and Antimicrobial Stewardship
Psychiatry: VSAIL suicide-risk model and digital phenotyping evidence
Psychiatry and Behavioral Health adds Vanderbilt's VSAIL suicide-risk model prospective validation data with PPV tables and hybrid-approach results, systematic review evidence on chatbot therapy, and a revised digital phenotyping section describing what current research shows and why the claims warrant skepticism.
Why Gives readers primary validation data for a real-world suicide-risk model and evidence-based framing for digital phenotyping and chatbot therapy claims.
Chapter Psychiatry and Behavioral Health
Critical care AI: sepsis prediction and ventilation evidence
Critical Care and Pulmonary Medicine adds Wong et al. (2021) validation data showing the Epic Sepsis Model's 33% sensitivity and 88% false-positive rate, alternative sepsis-prediction approaches, and RCT evidence on closed-loop ventilation and CT triage.
Why Gives readers primary validation data on a widely deployed sepsis early-warning model's real-world performance and RCT evidence for ICU ventilation and imaging-triage tools.
Chapters Critical Care and Pulmonary Medicine · Neurology and Neurological Surgery
Evidence and equity expansion across five specialty chapters
Dermatology, Ophthalmology, Orthopedic Surgery, Infectious Diseases, and Allergy/Immunology/Genetics add expanded evidence sections: dermatology's skin-tone performance gap and FDA device landscape, ophthalmology's IDx-DR diabetic-retinopathy screening data, orthopedic fracture-detection systems, antimicrobial-stewardship decision support, and genetic-variant classification tools, citing 46 new sources; Neurology adds a prospective ICH-detection triage study.
Why Deepens newly created specialty chapters with primary literature, an FDA regulatory document, and IDSA/SHEA and ACMG professional-society guidance on AI tools' real-world performance and clinical adoption.
Chapters Neurology and Neurological Surgery · Dermatology · Ophthalmology · Orthopedic Surgery and Physical Medicine · Infectious Diseases and Antimicrobial Stewardship · Allergy, Immunology, and Medical Genetics
Colonoscopy AI trial and meta-analysis evidence
Surgical Subspecialties expands its coverage of colonoscopy AI with RCT and meta-analysis evidence, including a 10 to 20% relative improvement in adenoma detection rate and comparisons of real-world versus RCT performance, plus new evidence on prostate MRI and ENT applications.
Why Establishes colonoscopy AI as the chapter's strongest evidence base ahead of less-validated applications in prostate imaging and otolaryngology.
Chapter Surgical Subspecialties
Nuance DAX Copilot data added to LLM and documentation chapters
Adds reported figures for Nuance DAX Copilot ambient documentation: 50% reduction in documentation time, 77% physician satisfaction, deployment in 550+ health systems, and its status as a non-FDA-regulated, CDS-exempt tool, plus pricing in a new documentation-platform comparison table.
Why Gives readers reported performance, regulatory status, and pricing for an AI scribe tool used across many health systems, sourced in part from the vendor's pricing page.
Chapters Large Language Models in Clinical Practice · AI-Assisted Clinical Documentation
Pilot-abort checklist, data-rights checklist, and automation-bias guidance added
Adds an abort checklist for AI pilots, a data-rights and ownership checklist for vendor evaluation, a section on the automation-bias risk of AI-assisted documentation, and a section on the evolving standard of care and negligence liability for AI use.
Why Gives physicians and administrators practical criteria for stopping failing AI pilots, negotiating data rights with vendors, and recognizing legal and behavioral risks of relying on AI tools.
Chapters Evaluating AI Clinical Decision Support Systems · AI Vendor Evaluation Framework · Large Language Models in Clinical Practice · AI-Assisted Clinical Documentation · Healthcare Policy and AI Governance
Hematology AI content added; oncology chapter renamed
Renames the oncology chapter to Hematology, Oncology, and Precision Medicine and adds new sections on peripheral blood smear analysis, bone marrow aspirate analysis, flow cytometry, CAR-T cell therapy response prediction, sickle cell disease AI, and anticoagulation management.
Why Gives the chapter dedicated coverage of hematology applications alongside its existing oncology content.
Professional Society Guidelines sections added to eight specialty chapters
Adds Professional Society Guidelines sections to eight specialty chapters (radiology, internal medicine, surgery, emergency and critical care, cardiology, neurology, pathology, and dermatology and ophthalmology), citing position statements from ACR, ACP, ACS, ACEP, AHA, AAN, CAP, and AAD.
Why Grounds each specialty's AI guidance in its own professional society's published position, alongside existing peer-reviewed evidence.
Chapters Diagnostic Imaging, Radiology, and Nuclear Medicine · Internal Medicine and Hospital Medicine · Surgery, Anesthesiology, and Perioperative Care · Emergency Medicine · Cardiology and Cardiothoracic Surgery · Neurology and Neurological Surgery
LLM training-cutoff guidance generalized beyond one model
Replaces a GPT-4-specific training cutoff date with guidance that all large language models have training data cutoffs and may be unaware of drugs, guidelines, or treatments published afterward.
Why Keeps the guidance accurate as newer models are released, instead of naming one model's specific cutoff date.
Oncology chapter expanded with cancer screening and pathology AI sections
Expands the oncology chapter with new sections on lung, breast, and colorectal cancer screening AI and pathology applications including Gleason grading.
Why Covers AI applications across cancer screening and pathology.
Chapters Hematology-Oncology and Precision Medicine · Pathology and Laboratory Medicine
New Executive Summary chapter frames evidence-based approach and core principles
Adds a new Executive Summary chapter presenting key findings, five core principles for AI adoption, and recommendations for physicians, hospital administrators, and medical educators.
Why Gives readers a single overview of the handbook's evidence base, principles, and recommendations before working through individual chapters.
Chapters Executive Summary · Preface
Healthcare AI regulation and institutional governance
Healthcare Policy and AI Governance covers FDA regulatory pathways and its Predetermined Change Control Plan framework, the EU AI Act, WHO guidance on large multimodal models, reimbursement, and institutional governance, and adds a case study on the Epic Sepsis Model's real-world validation results.
Why Gives readers primary sources for the FDA, EU, and WHO regulatory frameworks and a documented example of a cleared model underperforming after deployment.
References, links, and wording were also kept current.
The handbook expanded internal medicine coverage of patient deterioration prediction and expanded cardiology, neurology, pathology, dermatology, and ophthalmology with named evidence, including Paige Prostate, IDx-DR, and EyeArt. It added vendor evaluation, governance, and career appendices, evidence-ranking tables across six chapters, and worked clinical scenarios across most specialty chapters. The license changed from MIT to Creative Commons Attribution-NonCommercial-NoDerivatives 4.0.
Evidence table added to Primary Care chapter
Adds a table ranking primary care AI applications (diabetic retinopathy screening, ambient documentation, symptom checkers, autonomous treatment) by evidence strength, plus five implementation principles on workflow fit and patient trust.
Why Gives primary care readers a quick view of which AI tools are ready for use and which remain unproven.
Chapter Primary Care, Family Medicine, and Preventive Medicine
Aidoc and Viz.ai: imaging AI cost and hospital adoption
The physician toolkit chapter describes imaging AI vendors in clinical context: Aidoc costs a reported $20-50K per scanner annually and is used across 1,000+ hospitals, and Viz.ai is deployed at 1,400+ hospitals.
Why Gives readers practical cost and adoption context, not just technical specifications, for choosing imaging AI tools.
Turing's test, the 1956 Dartmouth workshop, and pathology AI tools
AI in Medicine: A Brief History opens with Turing's practical test for machine intelligence and the 1956 Dartmouth workshop. Primary Care presents a symptom-checker scenario, and the physician toolkit chapter's Pathology AI section describes Paige Prostate, PathAI, Proscia, and 3Derm.
Why Gives the history chapter a clearer narrative frame and updates toolkit and primary care examples with more specific clinical detail.
Chapters History of AI in Medicine · Primary Care, Family Medicine, and Preventive Medicine · AI Tools Every Physician Should Know
Specialty chapters revised with added clinical detail and evidence
Expands narrative detail in Surgery, Pediatrics, OB/GYN, Oncology, and Neurology, adding MySurgeryRisk's validation across 400,000+ surgeries, CTG monitoring's lack of RCT evidence for newborn outcomes despite 85-95% sensitivity, and a Swedish trial where AI plus one radiologist matched two radiologists reading mammograms.
Why Sharpens specialty chapters with more specific validation numbers and clearer distinctions between proven and unproven AI applications.
Chapters Surgery, Anesthesiology, and Perioperative Care · Pediatrics and Neonatology · Obstetrics and Gynecology · Emergency Medicine · Hematology-Oncology and Precision Medicine · Neurology and Neurological Surgery
Evidence-ranking tables added across six chapters
Adds executive-summary tables ranking AI applications by evidence strength in the Clinical Data, Radiology, Surgery, Emergency/Critical Care, Liability, and Documentation chapters. Cites radiology's 500+ FDA-cleared devices and survey data that 62% of physicians report EHR documentation contributes to burnout.
Why Gives readers a quick, evidence-ranked view of which AI tools across six clinical areas are proven, problematic, or still emerging.
Chapters The Clinical Data Challenge · Diagnostic Imaging, Radiology, and Nuclear Medicine · Surgery, Anesthesiology, and Perioperative Care · Emergency Medicine · Physician AI Liability and Regulatory Compliance · AI-Assisted Clinical Documentation
Handbook license changes from MIT to CC BY-NC-ND 4.0
The handbook adopts a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International license in place of the MIT license, with an expanded copyright notice covering academic and educational use.
Why Clarifies permitted reuse of the handbook's content for medical education, CME, and citation while restricting commercial and derivative use.
Three new ethics case scenarios on algorithmic bias
Adds three case scenarios to Medical Ethics, Bias, and Health Equity on algorithmic racial bias in risk-stratification tools, plus new policy and career-guidance appendices and a References and How to Cite page.
Why Gives readers concrete scenarios for reasoning through algorithmic bias, plus practical governance templates, career guidance, and a citation reference.
Chapters Medical Ethics, Bias, and Health Equity · Clinical AI Policy Templates · Medical AI Career Guide · How to Cite This Handbook
Clinical scenario exercises added to primary care chapter
Primary Care and Family Medicine adds three worked clinical scenarios, including a symptom-checker case involving delayed recognition of sepsis and an ambient documentation tool (Nuance DAX Copilot) that missed a medication change.
Why Extends the scenario-based learning format to the specialty with the broadest patient contact.
Chapter Primary Care, Family Medicine, and Preventive Medicine
Clinical scenario exercises added to AI evaluation chapter
Evaluating AI Clinical Decision Support Systems adds worked scenarios, including a case where a vendor's reported 85% sensitivity and 90% specificity sepsis algorithm was deployed without local validation despite FDA clearance.
Why Illustrates the gap between vendor-reported performance and safe local deployment that the chapter's evaluation framework addresses.
Vendor evaluation, governance, and career appendices added
Appendices include the Clinical AI Vendor Evaluation Toolkit, the Hospital AI Governance Policy Template, and Career Pathways in Physician AI. Large Language Models in Clinical Practice centers on real cases of confidently incorrect chatbot output, Healthcare Policy and AI Governance contrasts the Epic Sepsis Model with the EU AI Act, and the AI failure catalog adds five documented cases.
Why Gives readers practical tools for vendor and institutional evaluation, alongside documented failure cases to calibrate expectations before adoption.
Chapters AI Vendor Evaluation Framework · AI Failures in Medicine: Lessons Learned · Large Language Models in Clinical Practice · Healthcare Policy and AI Governance
Clinical scenario exercises added to four specialty chapters
Diagnostic Imaging and Radiology, Surgery and Perioperative Care, Pediatrics and Neonatology, and Obstetrics and Gynecology add worked clinical scenarios, including a missed intracranial hemorrhage after AI triage, an AI risk calculator overestimating surgical risk, an adolescent AI suicide risk algorithm, and a noninvasive prenatal test false positive.
Why Gives readers concrete decision-making practice grounded in documented failure patterns.
Chapters Diagnostic Imaging, Radiology, and Nuclear Medicine · Surgery, Anesthesiology, and Perioperative Care · Pediatrics and Neonatology · Obstetrics and Gynecology
Clinical scenario exercises added to emergency and oncology chapters
Emergency Medicine and Critical Care and Oncology and Precision Medicine add worked clinical scenarios, including a sepsis alert fatigue case and a genomic classifier misinterpretation that led to overtreatment.
Why Extends the scenario-based learning format introduced elsewhere in the handbook to two more specialty chapters.
Chapters Emergency Medicine · Hematology-Oncology and Precision Medicine
Cardiology chapter expanded with named model evidence
Cardiology and Cardiovascular Medicine adds detailed coverage of ECG interpretation AI, including a Mayo Clinic model for detecting low ejection fraction and AI detection of hyperkalemia from ECG, plus cardiac imaging AI and an account of why accurate algorithms still miss STEMIs in practice.
Why Expands cardiology coverage from a brief overview into implementation-level detail and named model performance.
Neurology chapter expanded with stroke AI detail
Neurology and Neuropsychiatry adds detailed coverage of large vessel occlusion and intracranial hemorrhage detection tools, framed as stroke AI's clearest success story, plus neuroimaging applications beyond stroke and cases where accurate stroke AI still fails in practice.
Why Expands neurology coverage from a brief overview into named-tool evidence and implementation detail.
Pathology chapter expanded with FDA-cleared tool detail
Pathology and Laboratory Medicine adds detailed coverage of histopathology AI, including Paige Prostate (FDA-cleared in 2021) for cancer grading and Google/Verily LYNA for breast cancer lymph node metastasis detection, plus HER2 scoring and cytopathology applications.
Why Expands pathology coverage from a brief overview into named, FDA-status-specific tool evidence.
Dermatology and ophthalmology chapter expanded
Dermatology and Ophthalmology adds detailed coverage of IDx-DR, the first FDA-cleared autonomous diagnostic AI from 2018, and EyeArt for diabetic retinopathy screening, alongside an account of why melanoma detection AI has fallen short of the promise shown in early studies.
Why Expands the chapter from a brief overview into evidence-specific detail distinguishing proven from unproven applications.
Hospital medicine AI and implementation
Internal Medicine and Hospital Medicine covers patient deterioration prediction, including the Epic Deterioration Index, readmission risk, chronic disease management, medication safety, diagnostic decision support, and hospital-specific implementation challenges.
Why Expands hospital medicine coverage with implementation-specific detail.
Chapters Internal Medicine and Hospital Medicine · Preface
References, links, and wording were also kept current.
The Physician AI Handbook published its first edition, launching with 30 chapters across five parts: foundations, specialties, implementation, practical tools, and future directions. Within days it completed its ethics, privacy, safety, and workflow chapters and added a glossary, case studies, a specialty tools appendix, and a catalog of notable AI failures including IBM Watson for Oncology, the Epic Sepsis Model, and Google Flu Trends.
Medical ethics chapter expanded with bioethical framework
Medical Ethics, Bias, and Health Equity adds bioethical principles for medical AI, covering autonomy and informed consent, beneficence, non-maleficence, and justice and equity, plus a section cataloging documented algorithmic bias cases.
Why Documented algorithmic bias cases, read against the principles of autonomy, beneficence, non-maleficence, and justice, give clinicians a principled basis for judging whether a medical AI tool serves patients equitably.
Chapters Medical Ethics, Bias, and Health Equity · Surgery, Anesthesiology, and Perioperative Care
Privacy and HIPAA chapter expanded with AI-specific risks
Privacy, HIPAA, and Patient Data Security adds HIPAA's regulatory foundation and its limits for AI, plus five categories of privacy risk: re-identification, model inversion and membership inference attacks, data breaches, secondary use without consent, and algorithmic inference of sensitive information.
Why HIPAA compliance alone does not address AI-specific risks such as re-identification and model inversion, so clinicians handling patient data need to know where the law's protections end.
AI safety and workflow integration chapters completed
Clinical AI Safety and Risk Management adds FDA's Software as a Medical Device (SaMD) classification and AI/ML Action Plan, a failure mode and effects analysis framework, and common AI failure patterns. Integration into Clinical Workflow adds workflow mapping, human factors engineering, and common integration challenges.
Why Failure mode and effects analysis, workflow mapping, and human factors engineering give clinicians and health systems structured ways to anticipate how an AI tool can fail or disrupt care before patients are exposed.
Chapters Clinical AI Safety and Risk Management · Integration into Clinical Workflow
Liability, documentation, and research chapters completed
Medical Liability and Regulatory Compliance adds liability allocation among physicians, hospitals, and vendors and FDA's regulatory framework. AI-Assisted Clinical Documentation adds the evidence base for ambient AI documentation tools. Clinical Research with AI adds applications across literature synthesis, trial recruitment, and real-world evidence.
Why Clinicians need to know how liability is allocated among physicians, hospitals, and vendors, and what evidence supports ambient documentation tools, before relying on AI in practice.
Chapters Physician AI Liability and Regulatory Compliance · AI-Assisted Clinical Documentation · Clinical Research with AI
Emerging technology, global health, and policy chapters completed
Emerging AI Technologies in Healthcare adds foundation models, multimodal AI, and federated learning. AI and Global Health Equity adds AI applications and equity risks in low-resource settings. Healthcare Policy and AI Governance adds FDA, EU AI Act, and WHO regulatory approaches.
Why Physicians need to understand foundation models, multimodal AI, and federated learning, the differing FDA, EU AI Act, and WHO regulatory approaches, and the equity risks AI carries in low-resource settings.
Chapters Emerging AI Technologies in Healthcare · AI and Global Health Equity · Healthcare Policy and AI Governance
Medical AI glossary, case studies, and two future-directions chapters completed
The Glossary of Medical AI Terms adds more than 100 term definitions, and Case Studies in Medical AI adds five case studies, including diabetic retinopathy screening and stroke detection. Medical Misinformation and AI and The Physician-AI Partnership: Future Perspectives gain full content.
Why Definitions for more than 100 terms and worked case studies such as diabetic retinopathy screening and stroke detection help clinicians read AI evidence and relate it to their own practice.
Chapters Glossary of Medical AI Terms · Clinical Case Study Library · Medical Misinformation and AI · The Physician-AI Partnership: Future Perspectives
Medical AI tools by specialty appendix added
Medical AI Tools by Specialty catalogs FDA-cleared and widely used AI decision support tools across more than 15 specialties, with evidence levels and FDA status noted for each.
Why Gives readers a specialty-organized reference for evaluating available AI tools before adoption.
Chapter AI Tools by Medical Specialty
AI failure catalog and further resources appendices added
Notable AI Failures in Medicine catalogs cases including IBM Watson for Oncology, the Epic Sepsis Model, Google Flu Trends, and Theranos. Resources for Further Learning curates journals, websites, and courses.
Why Gives readers documented failure cases and a curated path for continued learning.
Chapters AI Failures in Medicine: Lessons Learned · Further Reading and Resources
First edition of The Physician AI Handbook published
Launches with Welcome, Preface, and 30 chapters across five parts: foundations, specialties, implementation, practical tools, and future directions, plus six appendices including a chapter summary compilation.
Why Establishes the handbook as a specialty-organized, evidence-based reference for physicians evaluating AI tools.
Chapters Welcome · Preface · History of AI in Medicine · Diagnostic Imaging, Radiology, and Nuclear Medicine · Primary Care, Family Medicine, and Preventive Medicine · Evaluating AI Clinical Decision Support Systems
Specialty AI from surgery to ophthalmology, and ethics, privacy, safety, and liability
The first edition has full chapters on AI in surgery, pediatrics, obstetrics and gynecology, oncology, cardiology, neurology, pathology, and dermatology and ophthalmology, and on the implementation topics of ethics, privacy, safety, workflow, liability, documentation, and research, plus five further chapters.
Why Completes the initial chapter set so every specialty and implementation topic launches with substantive content.
Chapters Surgery, Anesthesiology, and Perioperative Care · Obstetrics and Gynecology · Medical Ethics, Bias, and Health Equity · Privacy, HIPAA, and Patient Data Security · Clinical AI Safety and Risk Management · Physician AI Liability and Regulatory Compliance