Emerging AI Technologies in Healthcare

Foundation models now score highly on medical question-answering benchmarks, multimodal systems combine several clinical data types, and generative models can create synthetic training examples. These capabilities are real, but none automatically establishes safe clinical use. A benchmark result, regulatory authorization, workflow study, and patient-outcome trial answer different questions. The intended use, population, comparator, study design, endpoint, and system version must remain attached to every claim.

Learning Objectives

After reading this chapter, clinicians will be able to:

• Understand foundation models and their healthcare applications
• Evaluate multimodal AI systems integrating imaging, text, and genomics
• Assess generative AI capabilities and limitations in clinical settings
• Recognize edge AI, federated learning, and privacy-preserving technologies
• Anticipate emerging trends: AI agents, real-time diagnostics, digital twins
• Navigate the hype cycle, distinguishing genuine advances from overpromises
• Evaluate time-horizon claims without treating forecasts as measured facts

Chapter Summary (TL;DR)

Clinical Context

Emerging AI technologies promise capabilities far beyond current narrow applications: foundation models that reason across medical knowledge, multimodal systems that integrate imaging + genomics + clinical data, generative AI creating synthetic training data, and digital twins simulating patient-specific treatment responses. Physicians must distinguish genuine advances from hype.

Key Applications

Demonstrated in bounded settings: - Med-PaLM 2 scored 86.5% on the MedQA question-answering benchmark, not a clinical outcome study (Singhal et al., 2025) - Randomized and observational ambient-scribe studies report product- and workflow-specific effects; ambient documentation is not categorically FDA-cleared - Caption Guidance has prospective evidence for novice acquisition of specified diagnostic-quality cardiac ultrasound views, not autonomous interpretation - Federated-learning studies demonstrate distributed training, while privacy and generalizability depend on the protocol and participating sites - Wearable studies demonstrate selected detection endpoints, not universal screening accuracy or improved treatment outcomes

Not established: - Autonomous agents making unrestricted high-stakes clinical decisions - Broad digital twins that reliably simulate an individual’s treatment response - Unverified LLM medication dosing or auto-sent clinical instructions - A universal accuracy range for generative clinical notes - Synthetic data as a substitute for representative real-patient validation

What’s Uncertain: - Multimodal AI for diagnosis (promising pilots, limited external validation) - Real-time sepsis prediction from wearables (high false positive rates in practice) - Humanoid caregiving robots (strong workforce need, but published evidence remains concentrated in short, small studies on social interaction, cognitive support, and monitoring rather than autonomous hands-on care) - Privacy-preserving AI (differential privacy trade-offs with model performance) - EHR foundation models for trajectory prediction (CLMBR, JEPA approaches: research-stage, no clinical validation)

Critical Insights

• Evidence boundary: Technical benchmarks, external validation, workflow effects, comparative clinical outcomes, authorization, and postdeployment monitoring are distinct evidence levels
• Local implementation: Cost, staffing, infrastructure, monitoring cadence, and deployment time must be measured or estimated from the actual contract and workflow
• Failure specificity: Hallucination and citation-error rates vary across model, prompt, retrieval, task, and evaluation design; no universal rate should be assumed
• Care Robotics Evidence Gap: Workforce data justify serious evaluation of caregiving robots, but current studies do not yet support replacing human staff for feeding, transfers, toileting, bathing, or medication administration.
• Adoption and trust: Survey results depend on how “AI use” is defined, sampled, and dated. Adoption, confidence, and willingness to delegate different tasks should be reported separately rather than collapsed into one percentage

Clinical Bottom Line

Near-term adoption is most defensible where the task, evidence, oversight, and failure pathway are narrow. Broader multimodal, agentic, digital-twin, and embodied systems should be described as demonstrated, theoretical, or beyond current capabilities for each exact use. Physicians should demand evidence, participate in bounded pilots, and shape governance without treating a calendar forecast as fact.

Medico-Legal Considerations

• Documentation: Signing a note can carry attestation consequences, but liability is fact- and jurisdiction-specific rather than automatically assigned by a universal AI rule
• Regulatory scope: Device status depends on the exact intended use and function; not every foundation-model use requires FDA authorization
• Patient notice and consent: Requirements depend on jurisdiction, recording, data use, clinical function, and institutional policy
• Data and intellectual property: Federated projects require written agreements covering data rights, model ownership, publication, withdrawal, security, and commercialization
• Clinical harm: A fabricated protocol creates a foreseeable verification risk, but negligence and causation require case-specific legal analysis

Essential Reading

• Singhal et al., 2023, Med-PaLM medical question answering
• Singhal et al., 2025, Med-PaLM 2 benchmark and physician-rater evaluation
• Moor et al., 2023, generalist medical AI framework
• Rieke et al., 2020, federated learning in digital health
• Wornow et al., 2023, EHRSHOT benchmark

An Evidence Ladder for Emerging Systems

The same term, such as clinical agent, can describe a research prototype, a supervised drafting tool, or a system authorized to initiate actions. Evidence should be categorized before conclusions are drawn.

Evidence level	What it can establish	What it cannot establish
Technical benchmark	Performance on specified tasks and test data	Clinical utility, safety, or generalizability
Retrospective external validation	Transportability to an independent dataset	Workflow effect or patient benefit
Prospective workflow study	Feasibility, usability, access, and operational failure modes	Patient benefit unless outcomes are measured and design supports inference
Comparative clinical study	Effect relative to a comparator under study conditions	Universal effectiveness outside the studied setting
Regulatory authorization	Permission to market for the specified intended use	Local utility, superiority, or unrestricted use
Postdeployment monitoring	Performance and harms during actual use	Causal benefit without an appropriate comparator

The FDA list of AI-enabled medical devices can help locate authorized products. FDA notes that the list is not comprehensive, so the exact authorization record remains the controlling source. No model should be promoted from benchmark success to patient-affecting autonomy without evidence and controls at the level of the intended action.

---

Part 1: Deployment Lessons from Diabetic Retinopathy Screening

Taught us: Validation in controlled datasets does not guarantee successful access, workflow, or follow-up in practice. Deployment evidence can reveal constraints that diagnostic accuracy alone cannot measure.

The Development Evidence (2016)

Gulshan et al. developed a deep-learning system for diabetic-retinopathy screening using 128,175 retinal images and evaluated it on the EyePACS-1 and Messidor-2 datasets (Gulshan et al., 2016). At a high-sensitivity operating point, the system reported 97.5% sensitivity and 93.4% specificity for referable diabetic retinopathy in EyePACS-1. Those endpoints established diagnostic performance on specified image datasets. They did not establish whether primary-care staff could acquire gradable images, obtain results within the clinic workflow, arrange referral, or achieve treatment and vision outcomes.

The distinction matters because a screening intervention includes more than a classifier. It includes patient identification, image acquisition, image-quality handling, connectivity, result communication, referral capacity, attendance, confirmatory examination, and treatment. A model can perform as designed while the complete pathway fails to deliver timely care.

Qualitative Workflow Evidence (2020)

Beede et al. studied deployment of a deep-learning diabetic-retinopathy system in Thai clinics and documented how image acquisition, network reliability, staff expectations, patient flow, and referral processes shaped use (Beede et al., 2020). The study was qualitative. It is useful evidence about sociotechnical failure modes, but it does not support the precise rejection rates, processing times, referral-adherence percentages, cost estimates, patient quotations, or project-discontinuation claims that were previously attributed to it in this chapter.

The correct lesson is not that the Thailand program “failed dramatically.” The evidence shows that workflow and infrastructure constraints remained consequential even when diagnostic performance was strong.

Prospective National-Screening Evidence (2022)

Ruamviboonsuk et al. then reported a prospective interventional cohort study across nine Thai primary-care sites, eight rural and one urban (Ruamviboonsuk et al., 2022). Between December 2018 and March 2020, 7,940 people were screened and 7,651 were eligible for analysis. For vision-threatening diabetic retinopathy, the system reported 94.7% accuracy, 91.4% sensitivity, and 95.4% specificity against an adjudicated reference standard. Regional retina specialists over-read images as a safety mechanism.

The study also reported that 2,412 of 7,651 eligible participants were referred for diabetic retinopathy, diabetic macular oedema, ungradable images, or low visual acuity. That referral denominator should not be interpreted as an AI false-positive rate because several referral indications were included. The study measured real-world diagnostic performance and feasibility, not a randomized effect on vision loss, treatment completion, mortality, or cost-effectiveness.

What Each Study Establishes

Evidence source	What it establishes	What it does not establish
Gulshan et al., 2016	Dataset-specific diagnostic performance	Clinic feasibility, referral completion, or patient outcomes
Beede et al., 2020	Qualitative workflow and implementation constraints	Quantitative effectiveness, costs, or causal outcome effects
Ruamviboonsuk et al., 2022	Prospective multisite diagnostic performance and feasibility	Randomized patient benefit or universal transportability

The Lesson for Physicians

1. The intervention is the pathway, not only the model. Image acquisition, connectivity, staff roles, referral capacity, and patient follow-up belong in the evaluation.

2. Different designs answer different questions. A high diagnostic sensitivity does not contradict a qualitative report of workflow difficulty. Both can be true.

3. Safety mechanisms change the intervention. Specialist over-reading can reduce risk, but it also changes staffing, turnaround time, cost, and the degree of autonomy.

4. Abstention is a clinical event. An ungradable or rejected image needs an immediate, feasible fallback. Returning later may impose unequal burdens on rural patients, hourly workers, caregivers, and people with limited transport.

5. Follow-up capacity determines value. Detecting referable disease creates benefit only when confirmatory care and treatment can occur.

Deployment questions:

• What proportion of the intended population can complete image acquisition?
• What happens when the image is ungradable, the network fails, or the model abstains?
• Who explains the result, arranges referral, and closes the loop?
• What are the baseline and postdeployment times from screening to specialist evaluation and treatment?
• Do access, failure, and follow-up differ across sites or patient groups?
• Which endpoint would justify continuation, restriction, or withdrawal?

---

Part 2: A Bounded Success in Ultrasound Acquisition Guidance

Taught us: Emerging technologies are easiest to evaluate when they address a defined task, user, endpoint, and fallback. Caption Guidance illustrates acquisition assistance, not autonomous echocardiography.

The Problem

Point-of-care ultrasound (POCUS) is clinically valuable for rapid assessment of cardiac function, fluid status, pneumothorax, and procedural guidance, but requires significant training and operator expertise.

Barriers to POCUS adoption: - Competence in basic critical-care echocardiography requires defined image-acquisition and interpretation skills, while advanced applications require more extensive training and experience (Mayo et al., 2009) - Image acquisition is operator-dependent - Limited examinations do not replace comprehensive echocardiography when the latter is indicated - Training, credentialing, quality assurance, interpretation, and referral remain part of the clinical pathway

The opportunity: AI guidance that helps novice operators acquire diagnostic-quality images and interpret findings.

The Technology

Caption Health, formerly Bay Labs, developed an AI-guided cardiac-ultrasound acquisition system. Product configurations have included guidance and measurement functions, but evidence and authorization should remain connected to the exact function.

1. AI-guided image acquisition: Real-time feedback helps an operator position the probe to acquire specified standard cardiac views.
2. Separate analysis functions: Measurement or interpretation functions require their own product, record, and evidence review.

How it works: - The software analyzes the live image and supplies positioning and quality cues - The operator remains responsible for patient positioning, probe movement, acquisition, and escalation - The intended limited examination, compatible equipment, labeling, and required training should be verified in the current record

The Validation Journey

Phase 1: Algorithm development

Training-corpus size and retrospective measurement performance can explain how a product was built, but they do not establish whether a novice can acquire a clinically usable examination. Those claims should be drawn from the device record or connected development paper rather than reconstructed from marketing summaries.

Phase 2: Novice Operator Study (2019) - 8 nurses with no prior echocardiography experience, each scanning 30 patients (240 patients total) at two academic hospitals - Paired design: each patient scanned by a nurse using the AI software and by a sonographer without it - Results (Narang et al., 2021): - Nurse-acquired studies judged of diagnostic quality for left ventricular size and function in 98.8% (237/240) of cases, and for right ventricular size in 92.5% (222/240) - All four prespecified endpoints exceeded the 80% performance goal - Nurse and sonographer scans did not differ significantly for most secondary parameters

Phase 3: FDA De Novo authorization (2020) - Caption Guidance, DEN190040, received De Novo authorization for its specified acquisition-guidance function - De Novo authorization is not a 510(k) clearance and does not establish that every later Caption or GE HealthCare ultrasound function shares the same intended use

Phase 4: Postauthorization implementation

The product was acquired by GE HealthCare and has been integrated into commercial ultrasound offerings. Deployment counts, compatible devices, contract prices, and billing arrangements are dynamic market facts. They do not establish diagnostic accuracy, patient benefit, coverage, or payment and should be verified at procurement.

The Evidence

Prospective multicenter diagnostic study (n=240 patients) (Narang et al., 2021). Eight nurses with no prior echocardiography experience each scanned 30 patients using the deep-learning guidance software, and five level-3 echocardiographers independently judged whether the studies were of diagnostic quality:

Diagnostic target	Studies of diagnostic quality	95% CI
Left ventricular size	98.8% (237/240)	96.7-100
Global left ventricular function	98.8% (237/240)	96.7-100
Right ventricular size	92.5% (222/240)	88.1-96.9
Nontrivial pericardial effusion	98.8% (237/240)	96.7-100

All four endpoints exceeded the prespecified 80% performance goal. The study measured whether novice-acquired images were interpretable, not whether novice-acquired studies match expert studies on diagnostic accuracy.

Economic and clinical-impact boundary:

The prospective acquisition study did not measure avoided formal echocardiograms, time to disposition, earlier detection of decompensation, hospital utilization, patient outcomes, or cost-effectiveness. Those outcomes require separate comparative evidence. A local value analysis should measure training, device and software cost, image-acquisition success, interpretation, confirmatory imaging, downstream testing, time, and patient outcomes rather than assume that a limited examination replaces a formal study.

Why Caption Health Succeeded Where Google Health Failed

Factor	Caption Health	Google Health India
Evaluation feature	Caption Guidance evidence	Thailand retinal-screening evidence
—	—	—
Intended task	Assist acquisition of specified ultrasound views	Detect vision-threatening diabetic retinopathy in a screening pathway
Prospective endpoint	Diagnostic quality of selected novice-acquired views	Diagnostic accuracy and feasibility at nine primary-care sites
Human role	Operator acquires images; expert readers assessed quality	Staff acquire images; specialists provided safety over-read
Remaining questions	Interpretation, downstream decisions, outcomes, and local value	Referral completion, treatment, vision outcomes, cost, and transportability

This comparison does not establish one product as a universal success and the other as a failure. It shows how two prospective programs tested different tasks, endpoints, and safety pathways.

The Lesson for Physicians

Emerging technologies succeed when:

1. Clear unmet need: Caption addressed genuine barrier (POCUS training) limiting adoption of valuable technology
2. Augmentation, not replacement: The system assists novice acquisition; it does not establish expert-level interpretation or comprehensive examination
3. Defined fallback: An inadequate or uncertain examination needs a prespecified path to repeat acquisition, expert review, or formal echocardiography
4. Workflow-integrated: Fits into existing ED/ICU workflows without major disruption
5. Phased validation: Technical development, prospective acquisition evaluation, authorization, implementation study, and patient outcomes remain separate stages
6. Appropriate deployment environment: Designed for settings with reliable infrastructure

Questions to ask when evaluating emerging technologies:

What specific clinical problem does this solve? - Good answer: “Assists a defined operator in acquiring specified limited cardiac-ultrasound views, with an explicit interpretation and escalation pathway” - Bad answer: “Revolutionizes cardiac imaging” (vague, no specific unmet need)

What happens when AI is wrong or uncertain? - Good answer: “The protocol defines image-quality failure, repeat attempts, expert review, and criteria for formal echocardiography” - Bad answer: “AI rejects image, patient sent home” (unsafe failure mode)

What’s the validation pathway? - Good answer: “Prospective multicenter study (n=240) in which novice operators scanned patients under AI guidance and blinded experts judged diagnostic quality, published in a peer-reviewed journal, and authorized by FDA through the De Novo pathway” - Bad answer: “Validated on 10,000 images” (retrospective only, no operator study, no FDA clearance)

Does it augment or replace clinical judgment? - Good answer: “The acquisition, measurement, interpretation, and clinical-action functions are evaluated and governed separately” - Bad answer: “AI makes diagnosis, clinician documents AI recommendation”

Current market context: GE HealthCare acquired Caption Health in 2023. Current product names, integrations, indications, compatible hardware, availability, and contract terms should be verified on official records. Acquisition and adoption do not establish reimbursement or patient benefit.

---

Part 3: Foundation Models and Medical Language Tasks

Foundation models are large neural networks trained on vast, diverse datasets (text, images, code), then fine-tuned for specific tasks. Examples: GPT-4 (OpenAI), PaLM 2 (Google), Claude (Anthropic), LLaMA (Meta). In healthcare, foundation models are trained on medical literature, clinical notes, guidelines, and patient data.

Med-PaLM: Medical Question-Answering at Expert Level

Med-PaLM and Med-PaLM 2 (Google Research): Med-PaLM was evaluated in Nature in 2023, while the peer-reviewed Med-PaLM 2 evaluation appeared in Nature Medicine in 2025 (Singhal et al., 2023; Singhal et al., 2025).

Evaluated or proposed uses: - Answering medical benchmark questions - Generating candidate differentials in controlled evaluations - Drafting explanations at requested reading levels - Translating medical terminology - Summarizing supplied clinical or research sources

Performance in one use should not be transferred to another. A benchmark model can produce an answer without reliably retrieving current evidence, recognizing missing context, or acting safely in a live workflow.

Performance (Med-PaLM 2, 2023):

Benchmark	Med-PaLM 2 result	Appropriate interpretation
MedQA (USMLE-style)	86.5%	Question-answering accuracy on this benchmark
MedMCQA	72.3%	Performance on a separate exam-style dataset
PubMedQA	81.8%	Performance on the evaluated research-question dataset
MMLU clinical topics	Reported in the paper’s benchmark suite	Not a clinical-workflow comparison

Physician evaluation study (blinded comparison): - 1,066 clinical questions answered by Med-PaLM 2 and physicians - Blinded physician raters evaluated responses on 9 criteria (accuracy, comprehensiveness, harm potential, bias, etc.) - Results: Physician raters preferred Med-PaLM 2 answers over comparison physician answers on eight of nine axes in the study’s pairwise evaluation - Boundary: The 1,066 items were consumer medical questions and the outcome was rater preference, not diagnostic accuracy, independent practice, or patient benefit

Critical caveat: Benchmark accuracy may overstate reasoning capability. When familiar answer patterns are disrupted (replacing correct answers with “None of the Other Answers”), LLM accuracy drops 26-38%, suggesting pattern matching rather than genuine clinical reasoning (Bedi et al., 2025). This has implications for novel clinical presentations.

AMIE: Diagnostic Reasoning with LLMs

AMIE (Articulate Medical Intelligence Explorer) is Google’s LLM-based diagnostic system designed for clinical history-taking and differential diagnosis (McDuff et al., 2025).

Key findings (Nature, 2025):

In a blinded evaluation of 302 challenging published case reports:

Condition	AMIE Standalone	Unassisted Clinicians	AMIE-Assisted Clinicians
Top-10 diagnostic accuracy	59.1%	33.6%	51.7%
Search-assisted clinicians	-	44.4%	-

Interpretation: AMIE’s standalone differential and AMIE-assisted clinicians performed better than specified comparison conditions at the reported endpoints. Published case reports are curated, comparatively complete, and already known to have a diagnosis; they are not equivalent to undifferentiated live patients.

Limitations:

• Cases selected for diagnostic difficulty (not representative of routine practice)
• No prospective validation in clinical workflows
• Funded by Alphabet (potential conflict of interest)
• Does not address workflow integration, liability, or patient acceptance

Prospective feasibility data (2026): A 100-patient preprint described AMIE-supported previsit history taking in ambulatory primary care. The system included the final diagnosis in its differential in 90% of cases and in its top three in 75%; real-time safety monitors did not intervene, while primary-care physicians received higher ratings for practicality and cost-effectiveness of management plans (Brodeur et al., preprint, 2026). The study supports feasibility questions, not randomized patient benefit, and it should remain labeled a preprint.

Multimodal diagnostic conversation (2026): A Nature Medicine study extended AMIE from text-only diagnostic dialogue to multimodal patient conversations involving images, documents, and ECG data. In 105 simulated telehealth consultations, specialist physicians rated multimodal AMIE higher than primary care physicians on diagnostic accuracy and most evaluation axes, including several measures of multimodal reasoning (Saab et al., 2026). The study was an exploratory OSCE-style evaluation, not a randomized clinical trial with patient outcomes.

Clinical bottom line: AMIE has shown strong performance on published cases, simulated consultations, and early feasibility work. These comparisons are task- and endpoint-specific. They support further evaluation, not a general claim that an LLM matches or exceeds physicians or is ready for autonomous clinical use.

MedGemma: Open Medical Foundation Models

MedGemma is Google’s open-weight medical AI model family, released through the Health AI Developer Foundations program for developer evaluation and adaptation (Google Research, January 2026).

Unlike the proprietary Med-PaLM, model weights are available through Hugging Face and Vertex AI under current model-specific terms. Open-weight availability does not make a downstream system open source, clinically validated, or authorized.

MedGemma 1.5 4B capabilities (January 2026):

• High-dimensional imaging: CT volumes, MRI volumes, whole-slide histopathology
• Longitudinal imaging: Chest X-ray time series comparison
• Anatomical localization: Anatomical feature detection in chest X-rays
• Document extraction: Structured data from medical lab reports
• Medical reasoning: Text-based clinical question-answering

Reported performance improvements (Google benchmarks):

Task	MedGemma 1.5	MedGemma 1	Improvement
MedQA (medical reasoning)	69%	64%	+5%
EHRQA (EHR question-answering)	90%	68%	+22%
CT findings classification	61%	58%	+3%
MRI findings classification	65%	51%	+14%
Lab report extraction (F1)	78%	60%	+18%

Performance metrics are vendor-reported on internal benchmarks; independent validation pending.

MedASR: Medical Speech Recognition

Released alongside MedGemma 1.5, MedASR is a 105M-parameter Conformer-based model fine-tuned for medical dictation. Reported word error rates:

Task	MedASR	Whisper v3 Large	Reduction
Chest X-ray dictation	5.2%	12.5%	58% fewer errors
General medical dictation	5.2%	28.2%	82% fewer errors

Word error rates are vendor-reported; comparative performance on diverse clinical audio requires independent validation.

Intended use: MedGemma models are intended as starting points for developers, not deployment-ready clinical tools. Google explicitly notes: “HAI-DEF models are not intended to be used without appropriate validation, adaptation and/or making meaningful modification by developers for their specific use case.”

Vendor-reported implementations: Google has described use in Malaysia’s askCPG guideline-navigation work and by Taiwan’s National Health Insurance Administration for analysis of pathology reports. These are vendor-reported implementation examples, not peer-reviewed comparative outcome studies.

GPT-4 in Clinical Settings

OpenAI GPT-4 (non-medical-specific foundation model) demonstrates surprising medical capabilities without specialized fine-tuning (Nori et al., 2023, preprint).

Performance: - USMLE Step 1: 84% (passing: 60%) - USMLE Step 2: 81% - USMLE Step 3: 76% - Physician benchmarks exceeded in many domains

Candidate uses that require task-specific validation:

1. Clinical decision support - Differential diagnosis generation from chief complaint + HPI - Medication interaction checking - Evidence-based treatment recommendations

2. Medical documentation - Draft clinic notes from conversation transcripts - Summarize hospital course for discharge summaries - Generate patient-friendly explanations of diagnoses

3. Medical education - Tutoring medical students (personalized explanations, practice questions) - Simulated patient encounters - Literature search and synthesis

The Hallucination Problem

Critical limitation: LLMs confidently generate plausible but factually incorrect information (“hallucinations”).

Hallucination rates in medical contexts: - Citation fabrication: of 115 references in ChatGPT-generated medical papers, 47% were fabricated, 46% were authentic but inaccurate, and only 7% were both authentic and accurate (Bhattacharyya et al., 2023); fabricated references also appear in ChatGPT-assisted case writing (Alkaissi & McFarlane, 2023) - Diagnostic errors: on complex case-derived vignettes, the correct diagnosis was absent from ChatGPT-4’s top-10 differential in 17% of cases, from its top-5 in 19%, and from its top-1 position in 40% (Hirosawa et al., 2023)

Hypothetical failure exercises:

The following scenarios are pedagogical composites, not reported patient cases. The clinical details are intentionally abbreviated and must not be used as protocols.

Case 1: Fabricated chemotherapy protocol - Physician asked GPT-4 for pediatric ALL consolidation protocol - LLM generated detailed protocol with drug names, dosages, timing - Error pattern: The model supplies a plausible dose from the wrong protocol or treatment phase - Potential consequence: A protocol mismatch can cause treatment failure or toxicity; the current protocol, order set, label, and pharmacy review remain controlling

Case 2: Nonexistent clinical trial citation - Physician asked for evidence on specific surgical technique - The model supplies a plausible author, journal, year, and page range that do not resolve to a real paper - Verification: Open the DOI or primary record and confirm that author, journal, design, and claim form a connected record - Potential consequence: A fabricated citation can contaminate clinical documentation, research, or teaching

Case 3: Outdated guideline adherence - The model supplies a recognizable but superseded recommendation without identifying its date or jurisdiction - Verification: Open the current professional-society guideline and confirm population, recommendation class, exceptions, and update status - Potential consequence: Outdated guidance can produce advice that no longer matches current evidence or the patient’s context

Mitigation Strategies

For physicians using LLMs clinically:

1. Never use LLMs for medication dosing without verification against authoritative sources (UpToDate, Micromedex, package insert)
2. Verify all citations before citing in clinical documentation or research
3. Cross-check diagnoses/recommendations with clinical guidelines
4. Use LLMs for ideation/drafts, not final clinical decisions
5. Document AI use in clinical notes: “Differential diagnosis generated with AI assistance, verified by physician”

Technical mitigations (vendor responsibility):

• Retrieval-augmented generation (RAG): LLM retrieves information from trusted databases (UpToDate, clinical guidelines) before answering
• Confidence scores: LLM flags low-confidence responses for human review
• Citation verification: Automated checking that cited sources actually exist
• Human-in-the-loop: All LLM outputs reviewed by physician before clinical use

Consumer Health AI: The Direct-to-Patient Expansion

Major AI companies are expanding beyond clinician-facing tools to direct-to-consumer health products. In January 2026, OpenAI launched ChatGPT Health, with optional medical-record and wellness-app connections in supported configurations. OpenAI has reported high health-query volume, but vendor usage estimates are not evidence of safety, accuracy, or clinical benefit.

This represents a significant shift: LLMs moving from physician-controlled clinical tools to patient-accessed health companions. The development included collaboration with 262 physicians across 60 countries and the creation of HealthBench, an open-source evaluation framework with 48,562 physician-developed rubric criteria.

Implications for physicians:

• Patients will increasingly present with AI-interpreted lab results and health recommendations
• HIPAA applicability depends on the parties, data, and function; direct consumer use may fall outside HIPAA, which is different from a universal product-level compliance conclusion
• Risk communication may differ from clinical conventions (see AI Tools Every Physician Should Know for details)
• Physicians should ask patients about AI health tool usage and be prepared to contextualize AI-generated insights

Multi-Agent AI Systems in Telehealth

Beyond single-model approaches, some vendors are deploying multi-agent architectures where specialized LLM agents handle distinct clinical tasks: one agent for history-taking, another for differential diagnosis, another for treatment planning, and another for documentation. This mirrors the division of labor in clinical teams.

Examples: Doctronic has described a multi-agent framework for the Utah prescription-renewal demonstration. The claim that its architecture coordinates more than 100 specialized agents is company-reported. The public policy demonstration, company architecture, and evidence for diagnostic or prescribing performance are separate claims.

Oncology applications: AI agents in cancer research represent one of the most advanced deployments of multi-agent systems in medicine. A Nature Reviews Cancer primer (Truhn et al., 2026) describes agents that orchestrate complex workflows:

• Drug design agents: Screen millions of compounds, predict binding affinity, assess toxicity, propose structural modifications, iterate autonomously until candidate identified
• Treatment strategy agents: Integrate patient genomics + pathology + imaging, search literature for similar cases, propose personalized regimens with evidence-based rationale
• Clinical trial matching agents: Analyze eligibility across thousands of protocols, flag exclusions from unstructured EHR data, rank by likelihood of benefit

These agents go beyond single predictions to handle multi-step problems requiring planning, tool use, and iterative reasoning. Ferber et al. evaluated an autonomous oncology agent on 20 multimodal cases with entirely fictional patient profiles paired with selected images. The system reached correct conclusions in 91.0% of the cases and accurately cited relevant guidelines in 75.5% (Ferber et al., 2025). This was a small benchmark of constructed cases, not a clinical case series, prospective trial, or patient-outcome study.

Potential advantages:

• Specialization: Each agent optimized for specific task (e.g., medication dosing, diagnostic reasoning, drug-target matching)
• Modularity: Individual agents can be updated without retraining entire system
• Traceability opportunity: Tool calls and handoffs can be logged, but generated explanations are not automatically faithful reasoning records
• Complex problem-solving: Multi-agent coordination handles workflows beyond single-model capabilities

Critical limitations:

AI agents inherit hallucination risks from base LLMs. Agents can fabricate drug combinations, cite nonexistent trials, or misinterpret genomic variants. No agent should autonomously make clinical decisions without physician verification.

A 28 August 2026 Lancet Digital Health Comment sharpens the vocabulary the multi-agent section still lacks. Qin et al. distinguish LLM chatbots (text or image outputs only), traditional agentic systems (task-specific, fragmented workflows), and autonomous agentic AI systems (AAAS) that add cross-task autonomy, persistent memory across usage histories, and hooks into messaging platforms such as WhatsApp, Slack, or Telegram, naming OpenClaw, ZeroClaw, Hermes, and MIRA as examples, not evaluated products (Qin et al., 2026). Because AAAS can execute actions and touch EHRs or tools, prompt injection, malicious plugins, or unintended writes can propagate across infrastructure in ways suggestion-only chatbots cannot. The authors sketch governance across technical safeguards (system-level red-teaming, RBAC, sandboxes, audit logs, skill validation), personnel rules (human-in-the-loop for high-risk record changes; AAAS as assistive collaborators, not independent decision makers), and regulatory oversight (autonomy scope, liability, re-certification when skills or capabilities change). Comment taxonomy ≠ clinical safety evidence.

Current evidence status:

• Peer-reviewed prospective validation comparing multi-agent vs. single-model clinical performance is not yet available
• Vendor-reported concordance metrics (e.g., “99% agreement with physicians,” “70-85% oncologist agreement”) reflect internal testing; independent evaluation is pending
• Whether coordination between agents introduces new failure modes or improves reliability remains an open question
• Regulatory status must be checked for the exact software function, intended use, and authority; an architectural label such as “agent” is not itself an FDA category
• Peer-reviewed multi-agent biomedical discovery systems have reported wet-lab validation. Co-Scientist reported multi-agent hypothesis generation with in vitro validation in acute myeloid leukemia, and Robin reported a lab-in-the-loop workflow that identified and validated candidate therapeutics for dry age-related macular degeneration (Gottweis et al., 2026; Ghareeb et al., 2026). This evidence should not be treated as support for unsupervised clinical decision-making.
• A medRxiv preprint analyzed 2,379 telemedicine encounters and reported 91.3% top-one diagnostic concordance with the treating clinician, rising to 96.3% in a confidence-selected subset. It also reported disposition concordance endpoints (Saenz et al., preprint, 2026). Treating-clinician agreement is not a gold-standard diagnostic-accuracy endpoint, confidence selection changes the evaluated population, and the preprint does not establish autonomous patient benefit.

What physicians should know: Multi-agent AI is an emerging architectural pattern with peer-reviewed evidence in constructed oncology cases and wet-lab discovery workflows. The Utah demonstration may generate implementation data, but its future results should not be forecast as evidence. For foundational concepts on agents, see AI Fundamentals: AI Agents. For oncology-specific agent applications, see Oncology: AI Agents in Cancer Research. For policy details on the Utah demonstration, see Case Study 6.

Key citation: Truhn, D., Azizi, S., Zou, J., et al. (2026). Artificial intelligence agents in cancer research and oncology. Nature Reviews Cancer. DOI: 10.1038/s41568-025-00900-0

Ambient Clinical Documentation with LLMs

Problem: Documentation and EHR work contribute to clinician burden, with time and after-hours patterns varying by specialty and workflow (Shanafelt et al., 2016).

AI solution: LLMs listen to patient encounters, generate draft notes for physician review.

Leading vendors: - Nuance DAX (Microsoft): Ambient documentation product with EHR integrations; documentation functionality should not be described categorically as FDA-cleared - Abridge: Specialty-focused (cardiology, oncology), mobile app - Suki: Voice-first AI assistant, commands + documentation

Evidence for Nuance DAX:

Stanford Health Care pilot (2024, n=48 physicians) (Shah et al., JAMIA, 2025):

• Burnout: Significant reduction (-1.94 points, p<.001)
• Task load: Large reduction (-24.42 points, p<.001)
• Usability: Improved perceptions (+10.9 points, p<.001)
• Physicians reported improved efficiency and documentation quality

Documentation time impact (Ma et al., JAMIA, 2025):

• Time per note reduced by 0.57 minutes (statistically significant)
• Daily documentation time: -6.89 minutes/day
• After-hours EHR time: -5.17 minutes/day
• Total EHR time: -19.95 minutes/day

Following the pilot, Stanford Health Care deployed DAX Copilot enterprise-wide in March 2024.

Head-to-head RCT (NEJM AI, 2025): The first randomized comparison of commercial ambient scribes randomly assigned 238 outpatient physicians across 14 specialties at UCLA to DAX Copilot, Nabla, or usual care (Lukac et al., NEJM AI, 2025):

Arm	Documentation time change vs. control	Statistical significance
Nabla	−9.5% (95% CI: −17.2% to −1.8%)	p = 0.02
DAX Copilot	−1.7% (95% CI: −9.4% to +5.9%)	p = 0.66 (not significant)

Both scribes improved burnout and task load on secondary endpoints. The contrast between Nabla’s significant primary endpoint effect and DAX’s non-significant result in this rigorous design warrants attention when evaluating real-world deployment decisions.

Limitations:

1. Review burden: Physicians need a reliable verification workflow. Time savings depend on whether review and correction are faster than the displaced documentation work
2. Unsupported content: Omission, addition, attribution, and carry-forward errors require product- and setting-specific measurement; no universal hallucination rate applies
3. Note length: Generated notes can be longer or differently structured, so review time, signal-to-noise, and downstream use should be measured locally
4. Privacy: Recording and processing depend on the parties, data flow, safeguards, contract, applicable law, and institutional policy; neither a BAA nor a vendor label resolves every use
5. Liability: Signing and attestation have consequences, but responsibility and civil liability remain fact- and jurisdiction-specific
6. Modality gap: In simulated medication histories, evaluated audio-only systems achieved 81% accuracy versus 98% for evaluated vision-enabled systems, with more omission errors (Menz et al., 2026). This does not establish that every current scribe is audio-only or that the comparison transfers across products. See AI-Assisted Documentation for details.

Medico-legal considerations:

• Malpractice risk: Omission of a critical detail can create clinical and legal risk; duty, breach, causation, and allocation depend on the facts and jurisdiction
• False documentation: A fabricated statement can produce an inaccurate record and downstream harm, regardless of whether software or a person drafted it
• Copy-forward errors: AI may propagate errors from previous notes
• Billing integrity risk: Coding must reflect the documented, supported service rather than note length or an unverified automated suggestion

Physician recommendations:

Use ambient AI for draft notes, not final notes Use a verification process appropriate to the note and clinical risk; a signature should not be reduced to a cursory skim Compare AI note to your mental note before signing Document AI use: “Note drafted with AI assistance, reviewed and edited by physician” Patient notice and consent: Follow applicable recording, privacy, institutional, and professional requirements for the exact workflow

Foundation Model Costs and Sustainability

Hidden cost: Foundation-model use creates contract, integration, review, monitoring, and computing costs that are often distributed across budgets.

Legacy price illustration, not a current quote:

Cost input	What must be measured
Model or platform use	Current written contract, included volume, overage, model, region, and retention terms
Encounter size	Input, output, retrieval, audio, image, tool, and retry volume
Human review	Time to verify and correct outputs, including downstream clarification
Integration	Interfaces, identity, security, testing, upgrades, and downtime support
Monitoring	Audit, incident response, subgroup analysis, and version-change evaluation

Illustrative health-system arithmetic:

The following legacy scenario is hypothetical: a health system has 1 million encounters, applies a tool to half, assigns an assumed per-encounter cost, and assigns a dollar value to time. The arithmetic can illustrate sensitivity to utilization, but it is not a measured price, adoption rate, time saving, cash saving, or return on investment.

Repurposed clinician time is not automatically cash savings. A defensible value analysis separates measured time, usable capacity, avoided expenditure, revenue, quality, safety, and patient outcomes.

Cost-effectiveness depends on the comparator, perspective, time horizon, implementation cost, uncertainty, and outcomes. It cannot be inferred from vendor token prices or a hypothetical hourly rate.

Environmental Sustainability of Foundation Models

WHO’s 2025 guidance raises often-overlooked environmental concerns about large AI models (WHO, 2025). Training and deploying foundation models requires substantial energy and water resources.

Environmental measurement questions:

Activity	Information needed
Model training or fine-tuning	Hardware, location, duration, energy source, utilization, and allocation across users
Inference	Exact model, input and output volume, batching, caching, quantization, retries, and data-center efficiency
Water use	Cooling design, climate, location, water source, season, and whether reported use is direct or supply-chain
Avoided work	Credible comparator, rebound effects, and whether digital activity replaces or adds to existing activity

Company-wide water or energy reporting cannot be assigned to one medical model without an allocation method. Likewise, a historical estimate for a model family should not be transferred to a current deployment with different hardware, location, workload, and energy mix.

Healthcare implications:

Healthcare organizations adopting foundation models should consider:

• Request deployment-specific environmental data from vendors: Ask for boundaries, methodology, location, time period, and uncertainty
• Benchmark efficient alternatives: Smaller or task-specific models can reduce compute in some tasks, but comparative performance and total system use must be measured
• Compare edge and cloud deployment: Local inference can reduce network dependence but may shift energy use to devices; the lower-impact design depends on workload and infrastructure
• Sustainability reporting: Include AI environmental impact in institutional sustainability metrics

These concerns support measured comparison rather than a universal argument for or against adoption. Environmental cost should be reported beside clinical or operational benefit using explicit boundaries and uncertainty, not a hypothetical lives-saved comparison.

Imaging Foundation Models: From Single-Condition to Multi-Condition Triage

While text-based foundation models dominate headlines, the same paradigm is transforming medical imaging. Historically, many FDA-authorized imaging AI devices addressed one condition per algorithm (one model for intracranial hemorrhage, another for pulmonary embolism). Foundation models trained on large imaging datasets are being evaluated for multiple findings from a single scan.

• Aidoc CARE: FDA cleared the device under K252970 in January 2026 for the labeled triage-and-notification functions involving specified abdominal CT findings. The device record, not the broader phrase “foundation model,” defines its authorized use. See the Radiology chapter for the evidence boundary.

Why this matters: Multi-condition triage can consolidate some technical functions and support urgency-based worklists. Whether it reduces integration burden, false alerts, turnaround time, or patient harm requires product- and site-specific comparison. A vendor-reported reduction should not be presented as a general consequence of the architecture.

Caveat: Do not infer market-wide uniqueness from a periodically updated FDA list or vendor category. Compare exact device records, architectures, indications, and versions at the time of review.

Biosignal Foundation Models: Beyond Text and Images

Foundation models are not limited to language and vision. Researchers have begun applying the same self-supervised pretraining paradigm to physiological signals: ECG, EEG, respiratory waveforms, and polysomnography (PSG).

SleepFM (Stanford, 2026): A multimodal foundation model trained on over 585,000 hours of PSG recordings from approximately 65,000 participants (Thapa et al., 2026).

Architecture: Channel-agnostic transformer that integrates EEG, ECG, EMG, and respiratory signals using contrastive learning. The model learns sleep representations without requiring labeled data, then fine-tunes for downstream tasks.

Disease prediction from one night of sleep:

Condition	C-Index	Clinical Relevance
All-cause mortality	0.84	Comparable to established risk scores
Dementia	0.85	Sleep disturbances precede cognitive decline by years
Heart failure	0.80	Sleep apnea and cardiac function interlinked
Parkinson’s disease	0.89	REM sleep behavior disorder is prodromal marker
Chronic kidney disease	0.79	Sleep disorders common in CKD
Stroke	0.78	Sleep apnea increases stroke risk
Atrial fibrillation	0.78	Nocturnal arrhythmias detectable in PSG

The model achieved C-Index ≥0.75 for 130 conditions (Bonferroni-corrected P < 0.01), suggesting sleep physiology contains rich information about systemic health.

Why this matters for clinical practice:

1. PSG is underutilized for prognosis. Sleep studies are ordered for sleep disorders but rarely mined for cardiovascular, neurological, or mortality risk. Foundation models could extract this latent information.

2. Transfer learning works. SleepFM showed strong performance on the Sleep Heart Health Study dataset despite exclusion from pretraining, suggesting generalization across clinical sites.

3. Multimodal integration outperforms single signals. Combining EEG, ECG, EMG, and respiratory data improved predictions over any single modality, consistent with sleep’s role as a window into multiple organ systems.

Limitations:

• Selection bias: Training cohort consisted of patients referred for sleep studies, not the general population
• Deployment gap unknown: No prospective validation in clinical workflows yet
• Actionability unclear: Predicting dementia risk from sleep data raises the same ethical questions as Alzheimer’s neuroimaging (see Neurology chapter)

Clinical bottom line: Biosignal foundation models represent an emerging application of self-supervised learning to physiological data. SleepFM demonstrates that overnight PSG contains prognostic information far beyond sleep disorders. Clinical utility will depend on prospective validation and integration into care pathways where early risk identification changes management.

EHR Foundation Models: From Documents to Dynamical Systems

Foundation models trained on structured electronic health records represent an emerging paradigm distinct from clinical language models trained on medical text. While text-based models (Med-PaLM, GPT-4) process clinical notes as documents, EHR foundation models operate on sequences of medical codes: diagnoses, procedures, medications, lab values, and vitals arranged chronologically.

CLMBR and the EHRSHOT Benchmark (Stanford, 2023)

The Clinical Language Model for Biomedical Research (CLMBR) introduced autoregressive pretraining over medical code sequences, demonstrating that representations from EHR foundation models transfer effectively to downstream clinical prediction tasks (Steinberg et al., 2021).

EHRSHOT benchmark characteristics:

• 6,739 patients from Stanford Medicine
• 41.6 million clinical events across 921,499 encounters
• 2.3x more clinical events per patient than MIMIC-IV
• Includes outpatient encounters (not just ICU/ED patients)
• 15 downstream prediction tasks across operational, anticipatory, and assignment categories

Performance: The 141M-parameter CLMBR-T-base model, pretrained on 2.57 million patient records, outperformed count-based gradient boosted models on most tasks when given ≥64 training examples, demonstrating that self-supervised pretraining on EHR data learns transferable clinical representations (Wornow et al., 2023).

The Reconstruction vs. Simulation Distinction

A conceptual tension exists in how foundation models represent patients:

Paradigm	Objective	Learns	Limitation
Reconstruction (standard LLMs)	Predict the next token	What the patient record says	Can defer trajectory reasoning to decode time
Simulation (world models)	Predict future states in latent space	Where the patient is going	Requires encoding dynamics before observing future

Standard autoregressive models optimized for next-token prediction model distributional regularities in clinical documentation but are not explicitly incentivized to encode how a patient’s disease state will evolve over time. The encoder can rely on the decoder to translate current representations into future tokens without encoding trajectory dynamics in the representation itself.

Emerging Approach: Joint-Embedding Predictive Architectures (JEPA)

Researchers have begun applying world model architectures from robotics and video understanding to clinical trajectories. Joint-Embedding Predictive Architectures (JEPA) predict masked content in representation space rather than input space, forcing the encoder to capture abstract dynamics that reconstruction objectives can ignore (LeCun, 2022).

SMB-Structure (Standard Model Biomedicine, 2026) combines supervised fine-tuning with JEPA to model patient trajectories across oncology and pulmonary embolism cohorts (Adam et al., 2026, preprint):

• SFT grounds the model in clinical semantics (reconstruction)
• JEPA forces trajectory dynamics into representations (simulation)
• Curriculum training (SFT first, then JEPA) outperformed simultaneous optimization
• Evaluated on 40,000+ patients (MSK oncology, Stanford INSPECT)

Reported findings:

• JEPA provided larger gains for long-horizon predictions (365-day mortality) than short-term tasks (30-day readmission)
• Cross-disease training helped JEPA models more than SFT-only models (“trajectory regularization”)
• Curriculum learning resolved objective interference between reconstruction and dynamics objectives

Evidence hierarchy: This is a preprint (not peer-reviewed) authored by the company developing the product. Validation is limited to linear probes on retrospective cohorts. No prospective clinical validation or deployment exists.

Why this matters for physicians:

1. Fundamental change: EHR foundation models may evolve from “summarizing patient records” to “simulating patient trajectories,” with implications for how AI generates predictions vs. how it could eventually reason about disease progression.

2. Evaluation sophistication: Linear probe evaluation at multiple points along disease trajectories (as used in these studies) may become standard for assessing whether models capture disease dynamics vs. static risk factors.

3. Limitation awareness: Models optimized for document reconstruction may struggle with trajectory prediction tasks where dynamics matter more than current state description.

Current status (2026): EHR foundation models remain research-stage. CLMBR-T-base weights are publicly available (Hugging Face). World model approaches like JEPA for EHR are emerging but require independent validation and prospective studies before clinical conclusions can be drawn.

---

Part 4: Multimodal AI Integrating Imaging, Text, and Genomics

Most current medical AI focuses on single modalities: images (radiology), text (NLP), or time-series (ECG). Real clinical reasoning integrates multiple data types: symptoms + imaging + labs + patient history + social determinants.

Multimodal AI systems combine imaging, text, genomics, wearables, and other data sources to generate holistic assessments.

Vision-Language Models for Medical Imaging

Problem: Radiology reports provide critical context missing from images alone (prior studies, clinical history, findings correlation).

Solution: AI models that understand both images and text.

Example: BiomedCLIP (Microsoft Research, 2023) (Zhang et al., 2023)

Architecture: Contrastive learning. AI learns to match medical images with corresponding text descriptions (radiology reports, pathology notes).

Capabilities: - Zero-shot classification (classify images into diagnostic categories without task-specific training) - Image-text retrieval (find similar cases based on text query) - Visual question answering (answer questions about medical images)

Reported preprint benchmark performance: - Chest X-ray diagnosis: 83% accuracy (14 pathologies) vs. 77% (prior vision-only models) - Pathology slide classification: 89% accuracy vs. 82% (vision-only) - Dermoscopy: 76% accuracy vs. 71% (vision-only)

Illustrative clinical-use concept: Radiology differential diagnosis

The following scenario and probabilities are hypothetical. They illustrate the information a multimodal system might combine, not a validated BiomedCLIP output or clinical recommendation.

Scenario: 45-year-old smoker with cough, weight loss. A chest radiograph shows a right upper-lobe nodule.

Vision-only AI: Detects nodule, flags for radiologist review (no differential)

Vision-language AI (multimodal): - Integrates image + clinical history (“45yo smoker, weight loss”) - Differential diagnosis: Lung cancer (65% probability), granulomatous disease (20%), pneumonia (10%), other (5%) - Recommendations: CT chest with contrast, consider PET-CT, pulmonology referral

Potential value: Contextualization may improve relevance, but the differential, probabilities, recommendations, and downstream effects require prospective validation. The cited BiomedCLIP preprint does not validate this scenario.

Genomic-Clinical Integration

Problem: Genetic test results interpreted in isolation miss clinical context (medications, comorbidities, family history).

Solution: AI integrating genomics + EHR data.

Illustrative pharmacogenomic decision-support scenario

Scenario: Patient prescribed clopidogrel for acute coronary syndrome. CYP2C19 genotyping shows 2/2 (poor metabolizer).

Genomics-only alert: “CYP2C19 poor metabolizer detected. Consider alternative antiplatelet therapy.”

Hypothetical multimodal output (genomics + EHR): - Integrates CYP2C19 genotype + current medications (omeprazole, a CYP2C19 inhibitor) + recent PCI + stroke history - Recommendation: “High-risk patient (recent PCI, stroke history) with CYP2C19 poor metabolizer genotype AND taking omeprazole (further reduces clopidogrel efficacy). Recommendation: Switch to prasugrel or ticagrelor, discontinue omeprazole or switch to H2 blocker.”

Evidence boundary: TAILOR-PCI enrolled 5,302 patients undergoing PCI. In the prespecified primary analysis of CYP2C19 loss-of-function carriers, the composite endpoint occurred in 4.0% with genotype-guided therapy and 5.9% with conventional therapy (hazard ratio 0.66, 95% CI 0.43–1.02, P=.06). The 34% relative estimate was not statistically significant at the prespecified primary endpoint and should not be stated as a demonstrated reduction. The trial tested a genotype-guided antiplatelet strategy, not the hypothetical multimodal AI above (Pereira et al., 2020).

Barriers to adoption: - Cost: Genotyping $200-500 (not always reimbursed) - Turnaround time: 1-7 days (too slow for acute decisions) - EHR integration: Most pharmacogenomic AI tools not integrated into Epic/Cerner

Generalist Biological AI: From Single Genes to Whole Systems

Pharmacogenomic AI operates on individual gene-drug pairs. A newer class of models works across the full information cascade from DNA sequence to RNA expression to protein structure to cellular behavior simultaneously.

What GBAI models do

A 2026 Nature Biotechnology review defines Generalist Biological AI (GBAI) as foundation models trained to process and generate across biological domains (DNA, RNA, proteins, and cellular systems) within a single framework (Rao et al., Nature Biotechnology, 2026, review). Prior AI tools solved one task at a time: AlphaFold predicts protein structure, Evo models genomic sequences, pharmacogenomic CDS flags individual variants. GBAI integrates these tasks.

Clinical implications physicians need to understand

Capability	Current evidence boundary	What would change the assessment
Joint DNA, RNA, and protein modeling	Active research	Independent task validation and experimental confirmation
Automated therapeutic-target identification	Discovery and preclinical workflows	Reproducible wet-lab validation and prospective development evidence
Drug design informed by patient genomic data	Preclinical concept	Validated treatment-selection studies with patient-relevant endpoints
Virtual-cell treatment-response simulation	Research concept	Prospective evidence that simulation improves decisions and outcomes

What already affects patients now

The genomic-clinical integration tools described above (pharmacogenomics, multimodal cancer profiling) are early, narrow GBAI applications already entering clinical use. The broader GBAI vision (models that can simulate a patient’s cellular response to a drug before prescribing it) remains research-stage.

Three unresolved problems the review flags

1. Data gaps. Most available datasets are from high-resource settings. Models trained on non-representative genomic data will perform differently across patient populations, a direct patient safety concern.

2. Experimental validation lag. AI predictions about molecular function often outpace wet-lab confirmation. A model predicting a therapeutic target does not mean that target is safe or actionable.

3. Accountability gap. Responsibility depends on the system, intended use, parties, workflow, representations, contracts, and jurisdiction. FDA regulates device functions rather than the label “biological foundation model,” and change management depends on the applicable regulatory record and policy.

What physicians should watch

GBAI may enter clinical practice incrementally, but the order and timing cannot be stated as fact. The evaluation framework is the same as for any clinical AI tool: what evidence supports this specific application, what is the failure mode, and who can detect and stop harm?

Key reference: Rao VM, Topol EJ, Rajpurkar P. “Generalist biological artificial intelligence in modeling the language of life.” Nature Biotechnology, 2026 (review). DOI: 10.1038/s41587-026-03064-w

Wearable Data + EHR Integration for Early Warning

Problem: Wearables generate continuous physiological data (heart rate, activity, sleep), but siloed from EHR.

Solution: AI integrating wearable data + clinical context for early warning.

Example: Apple Heart Study (2019, n=419,297) (Perez et al., 2019)

Single modality (wearable only): - Apple Watch detects irregular rhythm - Notification: “Irregular rhythm detected. Contact your doctor.” - Positive predictive value for AFib (confirmed by ECG patch): 84%

Multimodal (wearable + EHR): - Apple Watch detects irregular rhythm - AI retrieves EHR data: Age 72, CHA₂DS₂-VASc score 4 (high stroke risk), no current anticoagulation - Alert: “AFib detected in high-risk patient (CHA₂DS₂-VASc=4) not on anticoagulation. Urgent: Consider cardiology referral for anticoagulation.”

Clinical impact: In a randomized trial, Apple Watch monitoring detected substantially more new atrial fibrillation than usual care (9.6% vs 2.3%, HR 4.40, 95% CI 1.66-11.66), and 57.1% of detected episodes were asymptomatic (van Steijn et al., 2026). Whether earlier detection shortens time to anticoagulation has not been established.

Barriers: - Wearable data ownership (patient owns device, data not in EHR) - Interoperability (Apple HealthKit, Google Fit, Fitbit do not share standard format) - False positives (wearable AFib detection generates 10-15 false positives per true positive)

Pathology-Radiology-Genomics Integration in Oncology

Problem: Cancer treatment decisions require integrating tumor morphology (pathology), tumor burden (imaging), and molecular profile (genomics).

Illustrative hypothetical: Multimodal AI for glioblastoma treatment planning

The input and output below are not findings from a UCSF pilot. Chen et al. evaluated multimodal pan-cancer survival prediction using histopathology and genomic data across cancer types; it did not validate glioblastoma treatment recommendations or treatment-response probabilities (Chen et al., 2022).

Input data: - MRI (tumor location, size, contrast enhancement, edema) - Histopathology (H&E slides, tumor grade, necrosis, vascular proliferation) - Genomics (IDH mutation, MGMT methylation, EGFR amplification) - Clinical data (age, KPS, prior treatments)

Hypothetical AI prediction: - Overall survival (median, 95% CI) - Treatment response probability (TMZ chemotherapy, bevacizumab, immunotherapy) - Personalized treatment recommendation

Performance: - Survival prediction: C-index 0.78 (multimodal) vs. 0.65 (clinical data only), 0.71 (genomics only) - Treatment response: AUROC 0.82 (multimodal) vs. 0.68 (genomics only)

Potential research value: A validated system might support prognosis or trial discussion. The hypothetical output does not establish treatment benefit, and prognostic discrimination is not treatment-effect prediction.

Limitations: - Requires access to all specified modalities; the local missing-data rate should be measured rather than assumed - Not prospectively validated (retrospective analysis only) - No FDA clearance (research use only)

Challenges for Multimodal AI

1. Data heterogeneity: Different formats (DICOM images, HL7 text, VCF genomics), resolutions, timestamps
2. Missing data: Not all patients have all modalities; missingness may be informative and setting-specific
3. Temporal alignment: Imaging from last week, labs from today, genomics from 2 years ago. How to weight temporally discordant data?
4. Explainability: Which modality drove prediction? Hard to interpret complex multi-source models
5. Validation: Requires representative datasets with the relevant modalities, acquisition conditions, missingness patterns, reference standards, and outcomes
6. Computational cost: Processing several modalities can add storage, transfer, preprocessing, inference, and review costs; measure them for the actual architecture

Current status: Multimodal AI has strong research activity and limited bounded deployment. Adoption timing depends on the exact use, evidence, infrastructure, regulatory pathway, and economics; a universal 3–10 year forecast is not evidence.

---

Part 5: Edge AI and Point-of-Care Diagnostics

Edge AI: Running AI models locally on devices (smartphones, ultrasound machines, wearables) rather than cloud servers.

Advantages: - Latency: Local processing can reduce network-dependent delay for some workloads - Privacy: Keeping selected data local can reduce transmission, but local storage, telemetry, updates, backups, and access still require controls - Accessibility: A fully local function may continue when connectivity is limited, if acquisition, power, storage, and updates also work offline - Cost: Local processing can shift cost from cloud inference to hardware, battery, support, and device replacement

Disadvantages: - Limited computational resources: Complex models may be too large for devices - Battery constraints: Continuous processing can reduce battery life; the effect depends on hardware, model, duty cycle, and workload - Model updates: Difficult to update on-device models (vs. cloud models updated centrally)

Smartphone-Based Diagnostics

1. Diabetic Retinopathy Screening

EyeArt (Eyenuk): FDA-authorized autonomous diabetic-retinopathy screening software used with specified retinal-camera workflows. It should not be described generically as a smartphone-based or on-device system.

How it works: - A compatible retinal-imaging workflow captures fundus photographs - The exact software version analyzes images according to its labeled process - Image-quality failure, output categories, referral, and specialist follow-up are part of the pathway

Evidence: EyeArt has product-specific prospective evidence, while smartphone retinal adapters and other autonomous systems have separate acquisition and performance evidence. Metrics should not be transferred among these products. See the Ophthalmology chapter for record-level appraisal.

Use case: Primary care screening (avoid ophthalmology referral for low-risk patients)

Cost-effectiveness: Hardware, software, staff time, dilation, repeat images, referrals, payer coverage, and treatment capacity determine local value. A smartphone-adapter price or ophthalmology-visit charge cannot be used as a universal break-even calculation.

Limitations: - Performance and imageability depend on the exact camera, software, operator, dilation protocol, population, and reference standard - The labeled intended use defines which disease targets and patient groups are covered - A negative screen does not replace evaluation for symptoms or conditions outside the intended use

2. Melanoma Detection

SkinVision, MoleMapper, and similar apps: Smartphone photo + AI → melanoma risk score.

Performance (independent evaluation of commercially available apps against 15 histologically proven melanomas and 20 benign lesions) (Sun et al., 2022): - Mean sensitivity: 0.28 (95% CI 0.17-0.39) - Mean specificity: 0.81 (95% CI 0.71-0.91); mean accuracy 0.59 (95% CI 0.55-0.62) - Eight of the evaluated apps identified no melanoma at all in their top-ranked output - Comparison boundary: The app results should be compared only with clinician performance measured on the same cases, images, instructions, and reference standard

Problems: - High false negative rate (most melanomas were missed in the evaluation above) - Variable performance based on image quality (lighting, distance, angle) - No FDA clearance for most apps (marketed as “wellness tools” not medical devices)

3Derm regulatory boundary: The company has reported FDA Breakthrough Device Designation for an autonomous skin-cancer function. Breakthrough designation is not marketing authorization. Current FDA status, intended use, acquisition system, and product-specific performance require primary-record verification.

Clinical bottom line: Independent evidence for consumer melanoma apps has shown serious performance limitations. Regulatory status and evidence are product-specific; neither absence from an FDA list nor a company designation substitutes for review of the exact function and record.

Point-of-Care Ultrasound AI (Revisited)

Caption Guidance (see Part 2): De Novo-authorized for a specified ultrasound-acquisition guidance function. Adoption does not establish patient benefit.

Emerging competitors:

1. Koios DS (breast/thyroid ultrasound): - AI-assisted BI-RADS scoring for breast lesions - Product-specific diagnostic studies and device records should be reviewed by indication, population, comparator, and operating point; a single uncited sensitivity and specificity pair is not a stable class claim

2. Aidoc (multi-condition triage): - Aidoc markets multiple radiology devices and workflow products; each indication and version has its own evidence and authorization record - K252970 covers specified abdominal CT triage-and-notification functions and does not authorize the entire portfolio - Deployment counts and vendor timing claims are adoption or market measures, not evidence of improved patient outcomes

Wearable AI: Continuous Monitoring

Apple Watch AFib Detection:

Algorithm: Photoplethysmography (PPG) sensor detects irregular pulse → AI classifies as AFib.

Validation (Apple Heart Study, n=419,297) (Perez et al., 2019): - Irregular pulse notification rate: 0.52% of participants over a median 117 days of monitoring - Among 450 notified participants who returned analyzable ECG patches, AFib was present in 34% - Positive predictive value: 84% (95% CI 0.76-0.92) for AFib on ECG concurrent with a subsequent irregular pulse notification - The study did not measure sensitivity: participants without notifications were not monitored, so missed AFib could not be counted

Clinical implications: - The study demonstrated that a large pragmatic wearable-screening workflow could identify some previously undiagnosed atrial fibrillation - Subsequent evaluation and treatment remained clinician decisions, and the study did not establish sensitivity or improved stroke outcomes - Cost-effectiveness depends on device access, confirmatory testing, downstream care, false alerts, and outcomes

Limitations: - False positives: roughly 1 in 6 notifications was not concordant with AFib on simultaneous ECG, and only 34% of notified participants who returned a patch had AFib on the patch overall (generates unnecessary anxiety, cardiology visits) - Sensitivity was never established, so the proportion of AFib missed, particularly brief episodes, is unknown - Not suitable for patients with known AFib (designed for screening, not monitoring)

Other wearable AI applications:

1. Fall detection (consumer wearables): - Accelerometer and gyroscope signals can trigger fall-detection and emergency-contact workflows - Sensitivity, false alarms, user response, and emergency escalation are device-, population-, and setting-specific

2. Sleep-apnea screening: - Consumer devices may combine oxygen saturation and other signals to estimate sleep-disordered breathing - The exact device record, endpoint, reference polysomnography, intended population, and current authorization must be verified rather than inferred from a brand name

3. Non-AI comparator: CardioMEMS pulmonary-artery pressure monitoring: - CardioMEMS is pressure-guided remote monitoring, not established machine-learning evidence. The patient must actively take a daily reading; transmission then occurs through the system - CHAMPION reported a 30% reduction in heart-failure hospitalizations at six months and 39% over mean 15-month follow-up in the pressure-guided group (Abraham et al., 2011) - This comparator shows that a monitored signal changes care only through an acquisition, transmission, review, and treatment-response pathway

4. Ring-form wearables: multi-biomarker continuous monitoring (Oura Ring):

Ring-form wearables combine PPG, continuous distal skin temperature, and accelerometry in a form factor that enables overnight temperature capture that wrist-worn devices replicate poorly. The Oura Ring Gen 3 has been the primary device in several large peer-reviewed studies.

Early illness detection:

• In TemPredict, 63,153 participants enrolled, but model development and evaluation used a much smaller set of 73 PCR-confirmed cases meeting the study’s high-quality-data criteria. The reported 2.75-day lead time, 82% sensitivity, 63% specificity, and AUC of 0.819 belong to that selected case analysis, not the full enrollment denominator (Mason et al., 2022)
• Methodological challenges and opportunities in wearable-based early illness detection are reviewed, including baseline correction and label refinement as candidate approaches for improving pre-symptomatic detection accuracy (Phipps et al., NPJ Digital Med, 2024)

Metabolic disease detection:

• Sleep and temperature features from a 21-night window achieved 0.88 AUROC for diabetes detection across 11,200+ participants (389 with diabetes, 10,820 without); combined sleep and temperature features improved AUROC from 0.808 to 0.880 over demographic-only models (Viswanath et al., Commun Med, 2026)

Limitations:

• Heart rate accuracy is moderate: MAE 9-14 bpm (MAPE 11-16%, CCC 0.45-0.66) vs. gold standard in a 45-adult validation study; Fitbit Charge 6 performed substantially better (MAE 4.5 bpm); intermittent walking substantially increased ring errors (Gielen et al., JMIR Form Res, 2026)
• The evaluated ring and research algorithm should not be treated as an authorized diagnostic claim. Verify current regulatory status for the exact product and function (see Consumer Wearables and FDA Regulation)
• Population-scale illness detection depends on aggregated signals; individual-level alerts carry high false positive rates

Embodied AI for Caregiving: Humanoid and Socially Assistive Robots

Why the field matters: The care-capacity problem is measurable. The U.S. Bureau of Labor Statistics projects employment of home health and personal care aides to grow 17% from 2024 to 2034, with about 765,800 openings per year (BLS Occupational Outlook Handbook, 2025). AARP estimates that 59 million adult caregivers provided 49.5 billion hours of unpaid care in 2024, valued at $1.01 trillion (Houser et al., AARP Public Policy Institute, 2026). Nursing-home providers still report difficult recruitment despite post-pandemic workforce recovery (AHCA/NCAL, 2026). For physicians, this is not a distant robotics question. It affects discharge planning, dementia care, fall prevention, caregiver burnout, and whether older adults can safely remain at home or in lower-acuity settings.

Current product direction: Twolabs, a Y Combinator Spring 2026 company, describes Tobi as a semi-humanoid robot for nursing homes, senior living communities, and care facilities. The company site emphasizes night-shift restocking, item delivery, low-touch assists, meal service, and companionship, while its YC profile lists broader intended roles including feeding, dressing, medication reminders, object retrieval, mobility support, and wayfinding (Twolabs, 2026; Y Combinator, 2026). The clinical implication is not that these tasks are solved. It is that the product category is moving from social presence toward activities of daily living, which raises the evidence bar: a robot that reminds a resident to take medication is a different safety problem than a robot physically assisting feeding, transfers, or dressing.

Evidence boundary: A 2026 JMIR Aging systematic scoping review of humanoid robot-assisted support for older adults identified 59 studies and 25 humanoid robot models, but found that applications were concentrated in social interaction, cognitive stimulation, and physical activity. The authors noted short interventions, small samples, few randomized trials, and limited coverage of daily-living needs such as fall prevention, feeding, toileting, bathing, and position changing (Cui et al., 2026). Earlier assisted-living robotics work found 73 publications from 69 studies with mixed findings, limited attention to care context, and rare longitudinal, large-sample designs (Trainum et al., 2023). A 2026 systematic review of social robots and caregivers included 11 studies and found mixed workload effects: some studies reported reduced physical labor or improved interaction, while others found added training, supervision, and quality-of-life burdens (Fan et al., 2026). Home-based care robots remain even less mature; a 2025 integrative review of community-dwelling older adults found 40 studies and concluded that practical support for independent daily living still requires broader functions, cost support, and collaboration across public agencies, academia, and health enterprises (Lee et al., 2025).

Clinically plausible near-term roles: - Medication and hydration reminders, with clear escalation when reminders fail - Wayfinding inside memory care, senior living, and rehabilitation facilities - Social engagement, cognitive exercise, and activity prompting - Fall detection, nighttime monitoring, and staff notification - Object retrieval and low-risk environmental assistance under staff supervision

Not ready for unsupervised deployment: - Feeding, dressing, toileting, transfers, or bathing without human supervision - Medication administration decisions - Mobility support for residents at high fall risk without validated safety testing - Dementia companionship that is used to replace human contact rather than supplement it - Any workflow where a robot failure leaves the patient alone, immobilized, overmedicated, unfed, or unable to call for help

Evaluation questions before pilot deployment:

1. Does the robot reduce staff time per resident without adding hidden supervision, cleaning, charging, troubleshooting, or documentation work?
2. Which outcomes improve: falls, pressure injuries, medication adherence, loneliness, agitation, staff injury, caregiver burnout, hospital transfers, or discharge disposition?
3. What is the safe default when the robot is uncertain, loses balance, loses network connectivity, misunderstands speech, or encounters an obstructed environment?
4. How are consent, privacy, dignity, and resident refusal handled, especially for dementia care?
5. Does the device have the appropriate regulatory status for its intended clinical function, or is it being marketed as a nonmedical support tool?

Clinical bottom line: Humanoid caregiving robots address a real care-capacity problem, but the current evidence supports bounded assistive pilots, not staff replacement. The strongest near-term case is not autonomous caregiving. It is supervised augmentation: reminders, wayfinding, companionship, monitoring, and low-risk task support with measurable outcomes and clear escalation to human caregivers.

---

Part 6: Federated Learning and Privacy-Preserving AI

Problem: Multisite development can improve representation and external evaluation, but sharing health data is constrained by law, governance, contracts, security, consent, purpose, and institutional capacity.

Solution: Federated learning trains AI models across multiple institutions without centralizing data.

How Federated Learning Works

1. Central server distributes initial model to participating institutions (e.g., 10 hospitals)
2. Each hospital trains model on local data (data never leaves hospital)
3. Model information is exchanged according to the protocol; gradients or weights can still leak information under some attacks
4. Central server aggregates updates to improve global model
5. Repeat according to the training and stopping protocol; no universal number of rounds applies

Analogy: Like 10 students studying separately for exam, then sharing only their study notes (not the actual textbook pages they read).

Privacy boundary: - Raw records can remain at the institution under a federated design - Model updates are not automatically anonymous or outside privacy and security obligations - Differential privacy, secure aggregation, access control, audit, and contractual governance address different threat models

NVIDIA FLARE: Federated Learning Platform

NVIDIA FLARE (Federated Learning Application Runtime Environment): Open-source platform for healthcare federated learning.

Multisite research evaluation: Breast-density classification (Roth et al., 2020)

Participants: 7 clinical institutions worldwide

Task: BI-RADS breast density classification from mammograms across diverse populations, equipment, and protocols

Federated training results: - Federated models performed 6.3% better on average than models trained on single-institution data alone - 45.8% relative improvement in generalizability when evaluated on other sites’ data - Key insight: Federated model generalized better to new sites despite substantial inter-site differences in mammography systems, class distributions, and dataset sizes

Governance questions, not reported study facts: - Who owns the resulting model, updates, improvements, and deployment rights? - How are scientific credit and authorship determined under contribution standards rather than data volume alone? - Can a site withdraw, and if so, what happens to learned parameters, checkpoints, publications, and deployed versions?

Challenges for Federated Learning

1. Technical Complexity - Requires standardized data formats across institutions (DICOM, FHIR) - Communication overhead depends on model size, update frequency, compression, security, and network capacity - Computational cost (each institution runs training)

2. Data Heterogeneity - Institutions have different patient populations, imaging equipment, labeling standards - Model may overfit to largest contributor (institution with 200K patients dominates vs. institution with 5K) - Solution: Weighted aggregation (balance contribution by data size and quality)

3. Security Risks - Model inversion attacks: Adversary reconstructs training data from model updates - Membership inference: Adversary determines if specific patient in training set - Mitigation: Differential privacy (adds noise to updates), secure aggregation (encrypts updates)

4. Governance and Trust - Who owns the model? (Intellectual property disputes) - What if institution provides low-quality data? (Degrades global model) - How to credit contributions fairly? (Co-authorship, licensing revenue)

5. Regulatory and Legal Analysis - Federated training across sites is not the same as external validation unless the design includes an independent evaluation that supports the claim - The responsible manufacturer, intended use, device record, and change protocol determine FDA obligations - Whether an update is protected health information or creates another regulated disclosure requires analysis of the actual information, attack surface, parties, and legal framework; it cannot be resolved by calling it a gradient

Differential Privacy: Mathematical Privacy Guarantees

Problem: Even without sharing raw data, adversaries might infer sensitive information from model outputs.

Example: Model trained on hospital data achieves 99% accuracy on diabetes prediction. Adversary adds/removes one patient’s data, re-trains model, observes accuracy change → infers patient diabetes status.

Solution: Differential privacy adds calibrated noise to model outputs/updates such that presence/absence of any single patient does not significantly affect results.

Formal definition: Algorithm is ε-differentially private if probability of any output changes by at most e^ε when one patient’s data is added/removed.

Practical interpretation: Smaller epsilon often represents a stronger guarantee within a specified differential-privacy mechanism, but epsilon cannot be graded by a universal strong, moderate, or weak table. Composition, delta, clipping, sampling, threat model, and release context matter. NIST SP 800-226 provides a framework for evaluating differential-privacy guarantees.

Trade-off: Stronger privacy → more noise → lower model accuracy.

Illustrative example (hypothetical magnitudes; the cited practical guide describes this privacy-utility trade-off but does not report these figures) (Ponomareva et al., 2023): - Non-private model: AUROC ~0.89 - ε=1 (strong privacy): AUROC ~0.84 (approximately −5 percentage points) - ε=5 (moderate privacy): AUROC ~0.87 (approximately −2 percentage points)

Physician interpretation: Differential privacy provides a formal guarantee defined by the mechanism and parameters. Utility loss is empirical and task-specific, not a universal 2%–5%. Screening and high-stakes diagnosis also cannot be assigned a generic tolerance without defining the consequences of false positives, false negatives, and unequal error.

---

Part 7: Generative AI for Synthesis, Simulation, and Creation

Generative AI creates new content: images, text, audio, video. In healthcare: synthetic medical images, personalized patient education, clinical scenario simulations.

Synthetic Medical Data for Training

Problem: AI training requires massive labeled datasets. Medical data is scarce, expensive to label, privacy-restricted.

Solution: Generative AI creates synthetic medical images that resemble real images but do not correspond to actual patients.

Methods:

1. Generative Adversarial Networks (GANs): - Generator creates fake images - Discriminator tries to distinguish real vs. fake - Both improve iteratively until fake images indistinguishable from real

2. Diffusion Models (Stable Diffusion, DALL-E): - Gradually denoise random noise into realistic images - Conditioning: “Generate chest X-ray showing pneumonia in right lower lobe”

Real-world example: Synthetic Chest X-Ray Generation (Bluethgen et al., Nat Biomed Eng, 2024)

A vision-language foundation model for realistic synthetic CXR generation demonstrated that augmenting real training data with synthetic images consistently improves downstream classification AUROC compared to training on real data alone. The benefit was present but modest (approximately 2% AUROC improvement), and fine-tuned models trained only on real data achieved comparable performance when real data volume was sufficient.

Interpretation: Synthetic data augmentation improves performance when real data is scarce, but does not replace real-world validation.

Limitations:

1. Domain shift: Synthetic and real images can differ in clinically or technically relevant ways; distinguishability depends on generator, detector, dataset, and task
2. Artifact propagation: If training data contains artifacts (e.g., pacemaker wires), synthetic images over-represent artifacts
3. Rare pathology: Generative models struggle to create realistic rare diseases (insufficient training examples)
4. Validation requirement: AI trained on synthetic data must be validated on real patients before clinical use

AI-Generated Patient Education

Problem: Physicians lack time to create personalized, literacy-appropriate patient education materials.

Solution: LLMs generate explanations tailored to patient health literacy, language, cultural context.

Example: GPT-4 for Post-Discharge Instructions

The medication and warning text below is an illustrative draft, not a validated discharge instruction set or a recommendation for an individual patient. A clinical version would require reconciliation against the actual diagnosis, medication list, dose, renal function, follow-up plan, health literacy, language quality, and institutional instructions.

Input to LLM: - Diagnosis: Acute heart failure exacerbation - Medications: Furosemide 40 mg daily, metoprolol 50 mg twice daily - Patient health literacy: 6th grade reading level - Language: Spanish

LLM output (translated back to English for illustration):

“You were in the hospital because your heart wasn’t pumping blood well. Your body filled with extra water, making it hard to breathe.

Your medications: - Water pill (furosemide): Take one pill every morning. This helps your body get rid of extra water. You’ll urinate more. This is normal. - Heart pill (metoprolol): Take one pill in the morning and one at night. This helps your heart beat slower and stronger.

Warning signs to call your doctor: - You gain 3+ pounds in one day or 5+ pounds in one week (weigh yourself every morning) - You’re short of breath when lying flat - Your ankles are swollen

What to do: Call [clinic phone number] during daytime. If nighttime or weekend, go to emergency room.”

Evaluation: The closest published benchmark is not discharge instructions but patient questions posted to a public forum: blinded licensed health care professionals preferred chatbot responses to physician responses in 78.6% (95% CI, 75.0%-81.8%) of 585 evaluations, and rated them higher on both quality and empathy (Ayers et al., 2023).

Voice-cloned patient education: A three-arm RCT of 174 participants compared traditional education with AI-generated content delivered using physician-voice or patient-self-voice cloning. The study reported compliance of 73.2%, 86.7%, and 92.5%, respectively, with one-month adherence and psychosocial outcomes also reported (Sun et al., 2026). These results belong to the studied intervention, population, follow-up, and outcomes; they do not establish that voice cloning generally improves medical understanding or safety.

Challenges:

1. Accuracy: AI can provide incorrect or incomplete medical information; error rates vary by model, prompt, source grounding, task, language, and evaluation
2. Liability: Who is responsible if patient harmed by AI-generated advice?
3. Cultural sensitivity: AI may miss cultural nuances (dietary restrictions, health beliefs)
4. Oversimplification: Lowering reading level may omit critical details

Physician recommendations:

Use AI to draft patient education, physician reviews and edits Verify all medical information against authoritative sources Test materials with patients for comprehension Do NOT auto-send AI-generated instructions without physician review

Digital Twins: Personalized Simulation Models

Digital twin: Virtual replica of a patient, a computational model simulating physiology, disease progression, treatment responses.

Concept: Integrate patient’s imaging, labs, genomics, medical history into predictive model. Run simulations to forecast disease trajectories, test treatments virtually before real-world administration.

Cardiovascular Digital Twins

Example: HeartFlow FFR-CT (FDA-cleared 2014, updated 2024)

How it works: - Input: Coronary CT angiography (CCTA) - AI creates patient-specific 3D model of coronary arteries - Computational fluid dynamics simulates blood flow - Output: Fractional flow reserve (FFR), a measure of stenosis hemodynamic significance

Clinical value: Non-invasive FFR (vs. invasive catheterization)

Diagnostic accuracy (NXT trial): FFR-CT achieves sensitivity 86%, specificity 79% vs. invasive FFR as the reference standard (Nørgaard et al., JACC, 2014).

Real-world outcomes (ADVANCE registry, n=5,083) (Patel et al., JACC Cardiovasc Imaging, 2019): - Clinical impact: FFR-CT-guided care substantially reduced invasive catheterizations that found no obstructive CAD; among patients referred based on FFR-CT findings, the majority had obstructive disease confirmed at invasive angiography

Economic evidence: Contract price, imaging cost, invasive angiography, downstream testing, payer rules, and patient outcomes vary. A historical charge difference is not a universal net saving. Economic conclusions should come from a defined comparison and perspective.

Limitations: - Requires CCTA that meets the product’s acquisition and quality requirements; the local nonevaluable rate should be measured - Turnaround time depends on workflow, version, transfer, analysis, and service operations - Accuracy degrades in heavily calcified arteries, stents

Oncology Digital Twins (Research Stage)

Concept: Tumor model from imaging + genomics → predict growth, treatment response.

Current research direction: Multimodal pan-cancer prognosis

The nearest clinical analog is multimodal deep learning for pan-cancer survival prediction, exemplified by approaches integrating whole-slide histopathology images with genomic data across 14 cancer types (Chen et al., Cancer Cell, 2022). These systems predict survival more accurately than genomics or pathology alone but are prognostic models rather than simulation platforms. They do not simulate tumor growth trajectories or predict responses to specific treatment sequences.

True glioblastoma digital twins (simulating individual tumor growth under different treatment regimens) remain entirely research-stage. No peer-reviewed prospective validation or FDA clearance exists for treatment-response simulation in GBM or other solid tumors as of mid-2026.

Limitations of oncology digital twin research: - No prospective clinical validation in any cancer type - High computational cost and multi-modality data requirements - Individual tumor biology produces high variance that limits simulation accuracy - Regulatory treatment depends on the exact intended use, risk, manufacturer, and function; the phrase “digital twin” does not create or eliminate a pathway

Current status: Digital twins for broad disease-trajectory and treatment-response prediction remain largely research-stage. Bounded cardiovascular modeling and procedure-planning products have regulatory records, including HeartFlow functions, inHEART segmentation, and FEops HEARTguide under DEN200030. Each should be described by its labeled function rather than used to validate the entire digital-twin category. No reliable calendar predicts broad oncology adoption.

---

Part 8: Evaluating Emerging Technologies

When vendors pitch emerging technologies (“our AI uses foundation models and multimodal learning!”), ask these questions:

1. What Specific Clinical Problem Does This Solve?

Good answer: “Assists the intended operator in acquiring specified limited echocardiographic views; the prospective study defines diagnostic-quality endpoints and the local pilot will measure acquisition success, time, escalation, and downstream imaging”.

Bad answer: “Our AI revolutionizes healthcare” (vague, no specific unmet need)

2. What’s the Evidence?

Hierarchy of evidence (strongest → weakest):

1. Prospective RCT published in high-impact journal (NEJM, JAMA, Lancet)
2. Prospective observational study at ≥3 external sites
3. FDA clearance/approval with published validation
4. Retrospective single-site validation
5. Vendor white paper (not peer-reviewed)
6. “Validated on 100,000 patients” with no publication
7. Media coverage only (no peer-reviewed publication)

Red flags: - “Validated internally” (no external validation) - “Deployed in 150+ hospitals” (deployment ≠ effectiveness) - “AI approved by hospital IT” (IT approval ≠ clinical validation)

3. What Are the Failure Modes?

Ask: “What happens when AI is wrong? How often is it wrong? Can errors be caught before patient harm?”

Good answer: “At the prespecified threshold, the study reports sensitivity, specificity, uncertainty, and abstention with confidence intervals. The workflow defines who reviews each output and what confirmation is required.”

Bad answer: “Our AI is 98% accurate” (does not address what happens in the 2% when it’s wrong)

4. What’s Required for Deployment?

Checklist:

• Device analysis completed for the exact function, with the applicable FDA record verified if authorization is required
• EHR integration? (Epic, Cerner native vs. third-party add-on)
• Workflow changes? (Physicians must change practice patterns = adoption barrier)
• Training required? (How many hours? Who provides?)
• Reimbursement? (CPT codes? Payer coverage policies?)
• Privacy, security, legal-role, contract, data-flow, and BAA analysis completed for the exact service

Red flags: - “Just install our app” (ignores workflow integration complexity) - “No training needed. AI is intuitive” (physicians need training for any new tool) - “Reimbursement coming soon” (no established CPT codes = no revenue)

5. What’s the Cost-Benefit Analysis?

Ask for total cost of ownership. The legacy ranges below are illustrative placeholders, not market benchmarks:

Cost Category	Illustrative planning placeholder
Software license	$10K-500K (institution-wide)
Cloud inference costs	$0.10-5.00 per patient encounter
EHR integration	$50K-250K (one-time)
Training	$20K-100K (physician time)
Ongoing support	$10K-50K (annual)
Total (first year)	$90K-905K

Ask for value proposition: - Time savings: Hours per physician per week - Avoided costs: Fewer unnecessary tests, shorter LOS - Revenue: New billable services, improved quality metrics - Quality: Improved patient outcomes, safety

Hypothetical break-even calculation: - A spreadsheet may assume two hours saved per physician per week, a $150 hourly rate, 100 physicians, and 52 weeks, producing $1.56 million in attributed time value - If an assumed annual cost is $200,000, the arithmetic produces 7.8 times attributed value, but it is not a measured ROI. Time value becomes a financial return only if the organization can use or monetize the released capacity and if review, implementation, and downstream costs are included

Red flags: - Vendor cannot provide TCO estimate (hidden costs likely) - Value proposition based on one unvalidated measure; physician experience can be measured, but it should not substitute for the endpoint the deployment is intended to change - A break-even calculation depends on a very high adoption assumption without testing uptake, review time, capacity use, or uncertainty

6. What’s the Deployment Timeline?

Illustrative planning sequence:

Phase	Local duration to estimate	Activities
Vendor selection	Based on procurement and risk	RFP, demonstrations, evidence review, contract negotiation
Legal, privacy, and security review	Based on data and function	Roles, data flow, safeguards, contract, BAA where applicable
EHR integration	Based on interfaces and testing	Build, identity, interface, validation, downtime testing
Bounded pilot	Based on event volume and endpoint	Silent, shadow, or active mode with prespecified criteria
Evaluation	Based on sample size and risk	Analyze results, failures, subgroups, workflow, and uncertainty
Scale, restrict, or stop	Based on pilot evidence	Training, rollout, monitoring, remediation, or withdrawal

Red flags: - A fixed rapid timeline that omits evidence, integration, governance, training, and safety work - “Skip pilot, go straight to institution-wide” (unsafe) - Vendor pressures rapid decision (“limited-time offer expires Friday”)

---

Part 9: Human-AI Interaction Design in Clinical Settings

The success of clinical AI depends not just on algorithm accuracy but on how physicians interact with AI systems. A systematic review found that AI development has “disproportionately focused on technological advancements, often overlooking the alignment between algorithmic outputs and human expectations” (Gomez et al., 2024).

Interaction Patterns in Clinical AI

Research identifies distinct patterns in how AI presents information to clinicians:

The numeric probabilities in the examples below are interface illustrations, not calibrated clinical estimates.

Pattern	Description	Clinical Example
Display-Information	AI presents supplementary evidence supporting recommendations	SHAP values showing which features drove a sepsis prediction
Display-Uncertainty	AI communicates confidence levels and prediction reliability	“High confidence (92%)” vs. “Low confidence (61%)”
Display-Alternatives	AI shows multiple possible diagnoses with probabilities	Differential diagnosis ranking with likelihood scores
Request-Input	AI asks for additional information to refine predictions	“Order lactate to improve prediction accuracy”
Explain-Counterfactual	AI shows how outcomes would change with different inputs	“If potassium normalizes, risk drops from 78% to 34%”

Communicating Uncertainty

Overconfident AI is dangerous AI. Deep learning models tend toward overconfidence when faced with unfamiliar conditions. A 2025 study on epistemic uncertainty in clinical decision support emphasized that “inappropriate extrapolation beyond well-supported scenarios may have dire consequences” (Lindenmeyer et al., 2025).

Effective uncertainty communication should:

1. Distinguish types of uncertainty:
   • Aleatoric (inherent randomness in the data)
   • Epistemic (model’s lack of knowledge about this case)
   • Clinical (uncertainty in the underlying diagnosis)
2. Use calibrated confidence displays:
   • Avoid false precision (“87.3% probability”)
   • Use categories when appropriate: “High,” “Moderate,” “Low” confidence
   • Display confidence intervals, not just point estimates
3. Flag out-of-distribution cases:
   • Alert when patient differs significantly from training population
   • “This patient’s presentation is unusual. Exercise additional caution with AI recommendations.”

Sequential vs. Simultaneous AI Integration

Research on human-AI teaming shows that teaming mode affects outcomes (Liu et al., 2025):

Sequential mode: Clinician forms an initial impression, then sees the AI recommendation - Preserves independent clinical judgment - May lead to anchoring on initial impression - Effects can differ by clinician expertise and task

Simultaneous mode: Clinician sees AI recommendation at the same time as clinical data - Greater average augmentation in the included studies analyzed by Liu et al. - Risk of automation bias (over-reliance on AI) - Effects can differ for junior and senior clinicians

No single interaction mode is optimal for all situations. Consider expertise level, decision stakes, and time pressure when designing AI interfaces.

Expertise Effects in Human-AI Collaboration

The 52-study reliability analysis reported that expertise moderated teaming performance in the included clinical studies (Liu et al., 2025):

• Junior clinicians: Benefit more from AI assistance but are more susceptible to automation bias
• Senior clinicians: Smaller performance gains but better at recognizing AI errors
• Implication: Training programs should teach critical evaluation of AI outputs, not just how to use AI tools

Design Principles for Clinical AI Interfaces

1. Preserve clinical agency: - Always allow override of AI recommendations - Do not require justification for overrides (creates implicit pressure to comply) - Make it easy to document clinical reasoning when differing from AI

2. Reduce cognitive load: - Integrate AI into existing workflow (do not require separate applications) - Minimize alert frequency (quality over quantity) - Use progressive disclosure (summary first, details on demand)

3. Support decision-making, not just decision display: - Show what actions would change the prediction - Highlight modifiable risk factors - Provide actionable next steps, not just risk scores

4. Enable calibration learning: - Allow clinicians to track their agreement/disagreement with AI over time - Provide feedback on when AI vs. clinician was correct - Build appropriate trust through experience

The SepsisLab Paradigm

The SepsisLab study at CHI 2024 demonstrated a promising interaction model (Zhang et al., 2024):

• AI ranks the top 5 lab tests that would reduce prediction uncertainty
• Clinicians can simulate “what if we ordered this test?” before actually ordering
• Counterfactual predictions show how each result might change the diagnosis
• Users reported this provided “a better human-AI team experience” compared to AI simply displaying predictions

This represents a shift from “AI tells you the answer” to “AI helps you find the answer.”

Warning: Human-AI Teaming Rarely Achieves Full Complementarity

Despite the promise, a reliability analysis of 52 empirical studies found that medical AI can augment clinicians while full complementarity was uncommon in the included studies (Liu et al., 2025). Potential contributors include:

• Clinicians over-rely on AI when it’s wrong (automation bias)
• Clinicians ignore AI when it’s right (algorithm aversion)
• Interface design does not facilitate appropriate reliance
• Training on AI interpretation is inadequate

The goal: Design and evaluate interactions for appropriate reliance, including whether clinicians accept useful outputs, detect errors, preserve independent skill, and act safely.

---

Check Your Understanding

Scenario 1: LLM Medication Dosing Error

Explicitly hypothetical: A hospitalist asks a general-purpose LLM to dose an antimicrobial for a patient with a serious infection, reduced kidney function, and a local resistant-organism risk. The model supplies a plausible regimen but selects the wrong disease context and does not account for the current institutional protocol, infusion strategy, susceptibility, renal-dosing method, or pharmacy policy. The clinician copies the answer into an order without opening an authoritative source or obtaining pharmacy review.

No dose in this exercise should be used for patient care. Product labeling, current guidelines, local susceptibility data, institutional protocols, and pharmacy expertise determine the applicable regimen.

Answer 1: What was the error?

The model selected information from a superficially similar indication while failing to bind the answer to the actual syndrome, organism, renal function, infusion protocol, and local guidance. The error is not merely arithmetic. It is a context-selection and verification failure.

Answer 2: Why did GPT-4 make this error?

Possible mechanisms include retrieval of the wrong protocol, outdated guidance, missing patient data, overconfident completion, and failure to surface uncertainty. The exercise does not establish which internal mechanism caused a particular model output.

Answer 3: What are the legal and governance issues?

Liability cannot be decided from the vignette alone. Duty, breach, causation, damages, institutional policy, software representations, warnings, contracts, and jurisdiction all matter. Clinically, the workflow should require:

1. Verify AI recommendations against authoritative sources (UpToDate, Micromedex, Sanford Guide)
2. Use the complete patient and treatment context: Indication, organ function, allergies, interactions, route, infusion strategy, microbiology, and local policy
3. Escalate when needed: Pharmacy or specialty consultation for high-risk, unfamiliar, or discordant recommendations

Documentation should record the clinical decision and verification that matter to care, not create a formulaic AI disclaimer that substitutes for safe practice.

Lesson: Never use an unverified general-purpose LLM output as a medication order. Use current authoritative references, local guidance, and pharmacy review appropriate to the risk.

Scenario 2: Federated Learning Model Ownership Dispute

Explicitly hypothetical: A CMIO’s academic medical center joins a 20-institution federated-learning project for breast-image analysis. The institutions use a distributed-learning platform, but the scenario is not a report about NVIDIA, FLARE, or an actual consortium.

Project details: - One hospital supplies a fictional 25% share of the study records - The resulting model has a fictional AUROC of 0.89 in the consortium’s evaluation - No formal agreement on model ownership/IP rights

Later: A lead institution seeks to become the manufacturer or sponsor and commercialize the system. The original agreement does not clearly address model ownership, regulatory responsibility, commercialization, learned parameters, or withdrawal.

Hospital leadership asks whether a data contribution creates intellectual-property, control, revenue, or credit rights.

Answer 1: Who owns the AI model?

The answer depends on contracts, intellectual-property law, employment and invention policies, funding terms, licenses, database rights, data-use permissions, and the actual contributions. Data volume alone does not determine ownership, authorship, regulatory sponsorship, or revenue allocation.

Lead institution argument: - We coordinated project, provided infrastructure (NVIDIA FLARE servers) - We wrote grant, secured funding - Our researchers designed model architecture - We’re filing FDA application (regulatory sponsor)

Your hospital argument: - We contributed 25% of training data (second-largest contributor) - Model performance depends on multi-site data (would not achieve 0.89 AUROC on lead institution’s data alone) - We invested physician time (labeling images, validating outputs)

Possible outcomes include negotiation, amendment, abandonment, dispute resolution, or litigation. The vignette cannot predict which would occur.

Answer 2: What should have been done differently?

Federated-learning agreement topics before project start:

1. Intellectual property: Define ownership and licenses for code, architecture, weights, updates, improvements, documentation, and deployed versions
2. Commercialization: Define whether revenue sharing exists and how contributions are valued; proportional sharing is a negotiated option, not a default rule
3. Publication and credit: Apply authorship criteria and contributor statements rather than promising authorship for data access alone
4. Regulatory roles: Identify the manufacturer, sponsor, quality-system duties, validation responsibilities, and postmarket obligations
5. Withdrawal and retention: Define whether records, updates, checkpoints, derived models, and published analyses can be removed or must be retained

Answer 3: How should hospitals approach federated learning projects?

Before joining federated learning project, CMIO should:

1. Review governance agreement (IP, revenue, authorship, opt-out)
2. Consult hospital legal/tech transfer office
3. Determine legal roles, permitted data use, security controls, and whether a BAA or another agreement is required
4. Define data contribution scope (how many patients, what data elements)
5. Establish data quality standards (labeling accuracy, completeness)
6. Determine whether the activity is research, quality improvement, product development, or clinical operations and obtain the applicable ethics and institutional review

Lesson: Federated learning can support multisite development, but it does not resolve governance. Do not contribute data, compute, personnel time, or model updates without a written agreement that addresses the intended project and downstream uses.

Scenario 3: Digital Twin Treatment Simulation Failure

Explicitly hypothetical: An oncologist is treating a patient with newly diagnosed glioblastoma after multidisciplinary review.

Fictitious vendor pitch: “GlioPredict AI” is an invented name used only for this exercise. The vendor claims that uploaded MRI, pathology, genomics, and clinical data can simulate survival under different treatments.

You input patient data: - MRI: 4.2 cm right frontal tumor, significant edema, mass effect - Pathology: Grade IV glioblastoma, high mitotic index, necrosis - Genomics: IDH wild-type, MGMT unmethylated (poor prognosis), EGFR amplified

AI output: - Standard therapy (surgery + radiation + TMZ): Predicted median survival 11 months - Experimental therapy through a fictitious trial identifier: Predicted median survival 18 months (+7 months)

AI recommendation: “Enroll in the experimental immunotherapy trial.”

You discuss with patient: “AI predicts you’ll live 7 months longer with immunotherapy trial.” Patient enrolls.

6 months later: Patient dies from progressive disease. Tumor grew rapidly despite immunotherapy. Survival: 6 months (vs. AI prediction: 18 months).

Family: “Doctor said AI predicted 18 months. Why did she die in 6 months?”

Answer 1: What went wrong?

Digital twin limitations:

1. Evidence mismatch: A prognostic model may be presented as though it estimates an individual treatment effect
2. Retrospective bias: Selection, missingness, treatment assignment, follow-up, and confounding can produce misleading associations
3. Uncertainty not communicated: A point estimate is presented without calibration, interval uncertainty, competing outcomes, or transportability analysis
4. Individual variation: A population model cannot guarantee an individual disease course

Answer 2: What legal and ethical questions arise?

The vignette does not establish malpractice. Legal analysis would examine duty, standard of care, informed consent, representations, evidence available at the time, reliance, causation, damages, institutional governance, product warnings, and jurisdiction. Ethical questions include whether uncertainty and alternatives were explained accurately.

Informed consent: Did patient understand AI prediction was estimate (not guarantee)?

Clinical appropriateness: Was trial eligibility verified, and did discussion accurately distinguish established care, research participation, prognosis, and uncertain treatment effect?

AI reliance: Did physician over-rely on unvalidated AI recommendation?

Plaintiff argument: - Physician presented AI prediction (18 months) as fact, not estimate - Patient chose immunotherapy trial based on AI prediction (would not have enrolled if knew uncertainty) - AI was unvalidated, research-stage tool (not FDA-cleared) - A different treatment would have produced a better outcome, which would require evidence rather than hindsight alone

Defense argument: - Trial enrollment can be clinically appropriate when eligibility, alternatives, evidence, and consent are handled correctly - AI was one input to shared decision-making (not sole basis) - Physician discussed uncertainty (“AI is estimate, not guarantee”) - Informed consent documented (“experimental therapy, no survival guarantee”)

Lesson:

1. Verify the exact system: “Digital twin” can describe different functions; determine intended use, evidence, validation, and regulatory status
2. Communicate uncertainty: Explain the difference among prognosis, treatment-effect estimation, simulation, and a guaranteed outcome
3. Document the real decision: Record alternatives, evidence, uncertainty, patient preferences, and research consent rather than inventing a precision interval the model has not validated
4. Do not over-rely on an unvalidated simulation: Multidisciplinary recommendations and trial consent should rest on evidence and patient values, not a fictitious individual survival promise

---

Questions About Emerging Clinical AI

Frequently Asked Questions

What are foundation models in healthcare?

Foundation models are pretrained on broad datasets and adapted to multiple downstream tasks. Med-PaLM 2 scored 86.5% on the MedQA benchmark, which demonstrates question-answering performance under that evaluation, not independent clinical practice. MedGemma is an open-weight model family intended for developer adaptation and validation.

What is multimodal AI in medicine?

Multimodal AI processes more than one data type, such as imaging, text, waveforms, laboratory results, or genomics. Performance must be evaluated for the exact input combination, acquisition conditions, missing-data pattern, task, comparator, and downstream action.

Are AI agents ready for clinical use?

Readiness depends on the exact task and authority granted to the system. Agents have completed simulated EHR and oncology tasks, but benchmark success does not establish safe autonomous patient care. Patient-affecting actions require least-privilege access, explicit approval boundaries, audit logs, rollback, and evidence matched to the intended use.

What are digital twins in healthcare?

Digital twin is used for several kinds of patient-specific computational model, from anatomy and blood-flow simulation to proposed disease-trajectory and treatment-response models. Some bounded procedural or physiological applications have regulatory records, while broad individualized treatment simulation remains largely research-stage.

What is edge AI in medicine?

Edge AI performs some or all computation on a local device rather than relying entirely on a remote cloud service. Local processing can reduce latency and network dependence, but it does not by itself guarantee privacy, diagnostic validity, offline operation, low energy use, or regulatory authorization.

What are EHR foundation models?

EHR foundation models are pretrained on longitudinal health-record data such as codes, medications, laboratory results, events, or text, then adapted to downstream tasks. Their retrospective transfer performance does not establish clinical utility or transportability to another health system.

What is CLMBR in medical AI?

CLMBR is a family of representations learned from longitudinal electronic health records using autoregressive pretraining. A later CLMBR-T model was pretrained on 2.57 million patient records and evaluated on EHRSHOT tasks. These are research benchmarks, not authorization for clinical deployment.

What is the difference between reconstruction and simulation in clinical AI?

Reconstruction and next-event objectives learn to reproduce or predict observed record content. Proposed simulation or world-model objectives aim to represent future states or trajectories. This is a useful conceptual distinction, not proof that a latent model faithfully simulates an individual patient’s disease course.

---

Key Takeaways

1. Match evidence to the claim: A benchmark, external validation, workflow study, authorization, and patient-outcome trial establish different things.

2. Deployment is a care pathway: The Thailand retinal-screening program combined strong prospective diagnostic performance with important workflow and infrastructure constraints.

3. Bounded use is easier to validate: Caption Guidance supports a specified acquisition task; it does not establish autonomous interpretation or comprehensive echocardiography.

4. Failure modes matter: Identify what happens when the system is wrong, uncertain, unavailable, out of distribution, or updated.

5. Governance follows the function: Federated learning, digital twins, foundation models, and embodied systems require agreements and controls matched to data, authority, risk, and jurisdiction.

6. LLM error is task-specific: Citation fabrication, omission, and clinical errors have been measured in defined studies; no universal 8%–15% rate applies. Medication outputs require authoritative verification.

7. Cost and time are local: Current contracts, integration, review, monitoring, volume, and usable capacity determine value and deployment time.

8. Physician role: Clinicians can define intended use, endpoints, failure pathways, escalation, and monitoring without forecasting inevitable replacement or partnership.