Appendix E — Medical AI Tools by Specialty
Medical AI directories are useful only when they distinguish a product name from its exact regulatory record, intended use, and clinical evidence. This specialty index preserves a broad set of searchable tools while marking which entries have a primary FDA record on this page and which remain discovery leads that require verification before procurement or clinical use.
After reviewing this directory, readers should be able to:
- Distinguish FDA clearance, De Novo authorization, and PMA approval from evidence of clinical utility.
- Route specialty-specific product questions to the handbook chapters that evaluate the relevant evidence.
- Verify an exact device, submission number, intended use, version, and decision date before relying on a regulatory claim.
- Separate authorized devices from research systems, non-device clinical decision support, consumer wellness products, and non-AI comparators.
- Build a local evaluation plan that matches the device’s proposed use and the consequences of error.
The directory is a routing resource, not an endorsement list. A product name alone does not establish that a particular version is FDA-authorized, effective in a local population, integrated safely, or commercially available.
- FDA list: Periodically updated and not comprehensive, because FDA identifies entries primarily from AI-related terms in public authorization records (FDA AI-Enabled Medical Devices)
- Radiology: The largest specialty represented on FDA’s public AI-enabled device list; product-specific evidence and intended uses still differ substantially.
- Cardiology: The category includes ECG interpretation, echocardiographic acquisition and measurement, CT-derived physiology, and research models. These are not interchangeable use cases.
- Pathology: Paige Prostate has a primary De Novo record for adjunctive detection on digitized prostate-biopsy slides; that record does not authorize tumor grading or every Paige product.
- Ophthalmology: IDx-DR, now marketed as LumineticsCore, has a De Novo record for autonomous detection of more-than-mild diabetic retinopathy within its labeled population and imaging configuration.
- Primary care/internal medicine: KidneyIntelX.dkd is a De Novo Class II prognostic test for CKD progression risk assessment in adults with type 2 diabetes and existing CKD, not a screening test
- Dermatology: DermaSensor is an adjunctive referral aid for specified suspicious lesions in patients aged 40 years or older, not a screening or standalone diagnostic device.
Regulatory pathway: A 510(k) clearance is based on substantial equivalence, De Novo establishes a class I or II device type when no legally marketed predicate exists, and PMA is the premarket approval pathway for class III devices (FDA 510(k) overview; FDA De Novo overview; FDA PMA overview).
Critical reminder: FDA authorization answers a defined regulatory question for a labeled device and intended use. It does not by itself establish patient-outcome benefit, local transportability, workflow safety, reimbursement, or superiority to standard care.
Introduction
This directory organizes FDA-authorized devices, research systems, non-device decision support, consumer tools, and non-AI comparators by specialty. FDA’s AI-enabled medical device list is updated periodically and is not comprehensive; the agency identifies entries from public authorization records, and newly published decision summaries may appear in later updates (FDA AI-Enabled Medical Devices).
Rows without a direct primary regulatory record are intentionally retained as discovery leads because readers search for these products by name. An unlinked status in a vendor site, review article, conference presentation, or older handbook snapshot must not be treated as a current FDA determination. Before procurement, verify the exact trade name, software version, submission number, indication, limitations, decision date, and whether a later supplement, recall, or labeling change applies.
Legend: - Primary-record verified: an exact FDA database or decision-summary link appears in the row. - Discovery lead: the product and clinical area are retained for search and routing, but this page does not independently assert current authorization or performance. - Research: a model, dataset, or investigational system, not a marketing-authorization claim. - Non-AI comparator: a digital or molecular tool whose evidence may inform workflow design but does not establish AI efficacy.
Verification Rules for Every Product
- Match the record to the exact product. Company-level authorization is not a blanket authorization for every model, module, or future version.
- Read the indication and limitations. Detection, triage, quantification, prognosis, acquisition guidance, and autonomous diagnosis are different functions.
- Match the evidence to the claim. Standalone accuracy supports an accuracy claim; reader studies support assisted-reader claims; workflow studies support process claims; comparative trials with patient endpoints are needed for outcome-benefit claims.
- Confirm the deployment context. Population, prevalence, equipment, acquisition protocol, clinical role, alert destination, and fallback process affect performance and safety.
- Check current status. Authorization does not establish current commercial availability, institutional approval, payer coverage, or compatibility with a local EHR, PACS, or laboratory system.
Radiology
Radiology contains many AI-enabled devices, but authorization, clinical function, and evidence must be checked at the level of the exact module. The names below preserve useful routes into chest imaging, stroke, mammography, and cardiac imaging. The detailed evidence review is in the Radiology chapter, including randomized, prospective, and observational study boundaries. See AI in Radiology.
Chest X-Ray
| Tool | Vendor | Directory role | Principal use to verify | Evidence route |
|---|---|---|---|---|
| Annalise CXR | Annalise.ai | Discovery lead | Multi-finding chest-radiograph assistance | Verify the exact module and FDA record; see Radiology |
| ClearRead | Riverain Technologies | Discovery lead | Chest-radiograph processing and nodule-related workflows | Verify the exact product name and intended use |
| Lunit INSIGHT CXR | Lunit | Discovery lead | Chest-radiograph detection and triage functions | Do not transfer a metric from one finding or dataset to another |
| qXR | Qure.ai | Discovery lead | Chest-radiograph interpretation, including tuberculosis workflows | Check product, jurisdiction, and setting-specific evidence |
| RedDot | Behold.ai | Discovery lead | Abnormal chest-radiograph triage | Verify current regulatory and commercial status |
Chest CT
| Tool | Vendor | Directory role | Principal use to verify | Evidence route |
|---|---|---|---|---|
| Aidoc Pulmonary Embolism | Aidoc | Discovery lead | Pulmonary-embolism detection or triage | Confirm the exact submission and whether the output is prioritization or diagnosis |
| Arterys Pulmonary Workstation | Arterys | Discovery lead | Lung-nodule or pulmonary-image analysis | Verify the current trade name, owner, and record |
| Lung AI | Thirona | Discovery lead | Quantitative lung analysis | Product name is not specific enough for a regulatory claim |
| AVIEW LCS | Coreline Soft | Discovery lead | Lung-cancer-screening and nodule workflows | Verify exact version and intended use |
| Veolity | MeVis | Discovery lead | Lung-nodule volumetry and longitudinal comparison | Verify current product and authorization record |
Head CT and Stroke
| Tool | Vendor | Directory role | Principal use to verify | Evidence route |
|---|---|---|---|---|
| Aidoc ICH | Aidoc | Discovery lead | Intracranial-hemorrhage detection or triage | Confirm exact submission, image type, and output |
| Brainomix e-ASPECTS / e-Stroke | Brainomix | Discovery lead | ASPECTS and stroke-imaging workflow support | Do not infer randomized evidence from another vendor |
| RapidAI | iSchemaView | Discovery lead | Perfusion, ASPECTS, and large-vessel-occlusion workflows | Product-specific workflow evidence is principally observational in the cited handbook record |
| Viz.ai LVO | Viz.ai | Discovery lead with product-specific trial evidence | Large-vessel-occlusion triage and team notification | Randomized and observational workflow evidence discussed in Radiology and Emergency Medicine applies to the evaluated Viz.ai workflow, not every stroke platform |
| Avicenna cHeadCT | Avicenna.AI | Discovery lead | Head-CT triage modules | Verify each module separately |
Mammography
| Tool | Vendor | Directory role | Principal use to verify | Evidence route |
|---|---|---|---|---|
| Transpara | ScreenPoint Medical | Discovery lead | Mammographic cancer detection or scoring | Match study version, reader role, and screening program |
| Genius AI | Hologic | Discovery lead | Tomosynthesis computer-aided detection | Verify the exact software and compatible acquisition platform |
| ProFound AI | iCAD | Discovery lead | Mammographic or tomosynthesis decision support | Verify exact product and record |
| Mia, second-reader workflow | Kheiron Medical | Non-U.S. deployment lead | Breast-screening second-reader workflow | A CE marking or overseas study is not an FDA authorization claim |
| Volpara | Volpara Health | Discovery lead | Breast-density and imaging-quality functions | Verify which module and claim are under review |
Cardiac Imaging
| Tool | Vendor | Directory role | Principal use to verify | Evidence route |
|---|---|---|---|---|
| Arterys Cardio AI | Arterys | Discovery lead | Cardiac MRI quantification | Verify current ownership, exact version, and record |
| Circle CVI | Circle Cardiovascular Imaging | Discovery lead | Cardiac MRI and CT post-processing | A software suite can contain functions with different regulatory statuses |
| ClearRead and bone-suppression products | Riverain | Discovery lead | Image processing | Verify the exact product and claimed cardiac use |
| HeartFlow FFR-CT | HeartFlow | Device and evidence lead | CT-derived coronary physiology | Confirm current PMA record, labeling, payer policy, and comparative evidence before adoption |
Musculoskeletal Radiography
OsteoDetect is a primary-record-verified example. FDA granted De Novo DEN180005 on May 24, 2018 for machine-learning-assisted identification and highlighting of distal-radius fractures on adult posterior-anterior and lateral wrist radiographs. That authorization does not establish the status or performance of other fracture-detection products. See AI in Orthopedics.
Ophthalmology
| Tool | Vendor | Directory role | Principal use to verify | Evidence boundary |
|---|---|---|---|---|
| IDx-DR, now marketed as LumineticsCore | Digital Diagnostics | Primary-record verified | Autonomous detection of more-than-mild diabetic retinopathy in adults with diabetes who have not previously been diagnosed with diabetic retinopathy | FDA granted De Novo DEN180001 on April 11, 2018; the pivotal evaluation was prospective and single-arm, not an RCT |
| EyeArt | Eyenuk | Discovery lead | Autonomous diabetic-retinopathy detection | Verify exact FDA submission, cameras, population, and current labeling |
| RETFound | Academic research | Research model | Foundation-model representations for retinal imaging | Research performance does not establish authorization or clinical utility |
| Altris AI | Altris AI | Discovery lead | OCT decision support | Verify the exact product, jurisdiction, and record before assigning an FDA status |
Autonomous diabetic-retinopathy systems have narrow labeled populations, equipment requirements, image-quality rules, and referral outputs. Those boundaries are part of the intervention. See AI in Ophthalmology.
Dermatology
See Dermatology for detailed clinical evaluation of these tools, including skin tone bias concerns.
Other Tools
| Tool | Vendor | FDA Status | Use Case | Evidence Level | Notes |
|---|---|---|---|---|---|
| DermAssist | Research or informational product lead | Skin-condition information | No FDA authorization claim is made here | ||
| SkinVision | SkinVision | Non-U.S. consumer-product lead | Lesion-risk information | Performance varies by study and use; do not use to rule out malignancy | |
| MoleScope | MetaOptima | Imaging-device lead | Dermoscopy image capture and management | An imaging-device record does not automatically authorize an AI diagnostic function | |
| VisualDx | VisualDx | Clinical decision-support lead | Differential-diagnosis support | Regulatory status depends on the exact function; no exemption claim is made here |
Patient-facing tools must not be used to rule out malignancy or defer clinically indicated assessment. See Dermatology for the evidence on consumer informational tools and teledermatology.
Cardiology
ECG Analysis
| Tool | Vendor | Directory role | Principal use to verify | Evidence boundary |
|---|---|---|---|---|
| Apple Watch ECG and rhythm-notification functions | Apple | Consumer-device lead | User-initiated ECG and rhythm notifications | Different Apple functions have different records, populations, outputs, and evidence; the Apple Heart Study did not validate every later feature |
| KardiaMobile | AliveCor | Device lead | Consumer or ambulatory ECG acquisition and algorithmic rhythm classification | Verify the exact hardware, algorithm, and submission |
| Cardiologs | Cardiologs | Device lead | Cloud-based ECG interpretation | Verify exact version and intended use |
| Eko AI-enabled stethoscope functions | Eko Health | Device lead | Murmur, rhythm, or cardiac-function support, depending on module | Verify each module separately; a company-level label is insufficient |
Echocardiography
| Tool | Vendor | Directory role | Principal use to verify | Evidence boundary |
|---|---|---|---|---|
| Caption Guidance | Caption Health / GE HealthCare | Device lead | Ultrasound acquisition guidance | Verify current trade name, compatible system, operator population, and exact record |
| EchoNet-Dynamic | Academic research | Research model and dataset | Video-based ejection-fraction estimation | The research model and its randomized reader-workflow evidence do not create a blanket 510(k) authorization claim |
| EchoPrime | Academic research | Research foundation model | Multi-view echocardiographic representation and prediction | Retrospective research evidence only; no FDA authorization claim is made here |
| Us2.ai | Us2.ai | Non-U.S. and device lead | Automated echocardiographic measurement | Verify jurisdiction, exact module, and record |
| TTE.ai / Bay Labs lineage | Product-history lead | Echocardiographic acquisition or measurement | Product ownership and names have changed; verify the currently marketed device |
See AI in Cardiology for the distinction between standalone model performance, assisted interpretation, randomized workflow studies, and clinical outcomes.
Pathology
| Tool | Vendor | Directory role | Principal use to verify | Evidence boundary |
|---|---|---|---|---|
| Paige Prostate | Paige | Primary-record verified | Adjunctive detection of suspicious areas on digitized prostate-biopsy whole-slide images | FDA granted De Novo DEN200080 on September 21, 2021; the record does not authorize Gleason grading or every Paige product |
| PathAI products and research models | PathAI | Discovery and research leads | Digital-pathology analysis varies by product and study | Verify the exact trade name, intended use, and record |
| Prosigna | Veracyte | Non-AI molecular comparator | Breast-cancer gene-expression prognosis | Molecular-assay evidence is not evidence for an image AI system |
| Ibex Galen / Galen Second Read | Ibex | Device lead | Whole-slide-image detection support | Verify the exact trade name, jurisdiction, and submission; do not transfer evidence across organ-specific modules |
Neurology
| Tool | Vendor | Directory role | Principal use to verify | Evidence boundary |
|---|---|---|---|---|
| Viz.ai LVO | Viz.ai | Stroke-workflow lead | Large-vessel-occlusion triage and notification | Product-specific randomized and observational workflow evidence must not be transferred to RapidAI or Brainomix |
| RapidAI | iSchemaView | Stroke-imaging lead | Perfusion, ASPECTS, or LVO functions | Verify exact module; the evidence summarized in this handbook is principally observational for product-specific workflow claims |
| Brainomix e-Stroke | Brainomix | Stroke-imaging lead | Stroke-image analysis and workflow support | Verify exact module and evidence; no Viz.ai trial transfer |
| icobrain | Icometrix | Quantitative-imaging lead | Brain volumetrics and lesion quantification | Verify jurisdiction, product, population, and record |
| Cortechs.ai products | Cortechs Labs | Quantitative-imaging lead | Brain MRI volumetrics | Verify exact product and intended use |
Primary Care / Internal Medicine
| Tool | Vendor | Directory role | Principal use to verify | Evidence boundary |
|---|---|---|---|---|
| Epic Sepsis Model | Epic Systems | EHR model and implementation-history lead | Sepsis-risk prediction | External validation reported 33% sensitivity and 12% positive predictive value at one health system; that study did not evaluate causal patient benefit (Wong et al., 2021) |
| KidneyIntelX.dkd | Renalytix | Primary-record-verified De Novo device | CKD progression risk assessment in adults with type 2 diabetes and existing CKD | DEN200052; derivation and independent validation are described by Nadkarni et al., 2023. It is prognostic, not a screening or standalone diagnostic test |
| Ambient clinical documentation | Nuance DAX, Abridge, Suki, and others | Documentation-product class | Draft clinical notes and workflow support | Evidence and data handling are product-, version-, setting-, and endpoint-specific; see AI for Clinical Documentation |
| UpToDate AI-enabled functions | Wolters Kluwer | Knowledge-support lead | Evidence retrieval and clinical decision support | Verify the exact function and current product documentation; do not describe it as an autonomous diagnostic device |
Gastroenterology
| Tool | Vendor | Directory role | Principal use to verify | Evidence boundary |
|---|---|---|---|---|
| GI Genius | Medtronic | Device and evidence lead | Real-time computer-aided polyp detection | Randomized evidence addresses evaluated CADe-assisted colonoscopy workflows; verify current labeling and hardware compatibility |
| EndoBRAIN | Olympus / Cybernet | Non-U.S. or research lead | Polyp characterization | Verify jurisdiction, product version, and whether the claim concerns detection or optical diagnosis |
| CAD EYE | Fujifilm | Non-U.S. or device lead | Polyp detection or characterization | Verify the exact module and regulatory record |
The Surgical Subspecialties chapter reviews colonoscopy AI evidence and separates adenoma-detection outcomes from serrated-lesion and advanced-adenoma claims.
Oncology
| Tool | Vendor | Directory role | Principal use to verify | Evidence boundary |
|---|---|---|---|---|
| Oncotype DX | Exact Sciences | Non-AI molecular comparator | Breast-cancer recurrence scoring and treatment-decision support | Assay-specific clinical evidence does not establish AI efficacy; verify the exact assay indication and guideline context |
| Tempus products | Tempus | Genomic and computational-product leads | Molecular profiling, data analysis, and trial matching vary by product | Verify the exact assay or software, regulatory status, and evidence; company-level language is insufficient |
| Foundation Medicine assays | Foundation Medicine / Roche | Molecular comparator and device leads | Tumor genomic profiling | Verify the exact assay and PMA or supplement; do not label the entire company portfolio as AI |
| Paige Prostate | Paige | Primary-record-verified pathology AI | Adjunctive detection on digitized prostate-biopsy slides | De Novo DEN200080, not a 510(k), and not an authorization for grading |
See AI in Oncology for trial-level genomic, biomarker, imaging, and language-model evidence. A biomarker-guided trial is not automatically an AI trial.
Anesthesiology / Critical Care
| Tool | Vendor | Directory role | Principal use to verify | Evidence boundary |
|---|---|---|---|---|
| Masimo sedation-monitoring functions | Masimo | Non-AI or algorithmic comparator, depending on function | Physiologic monitoring and sedation-related indices | Verify the exact device and algorithm; advanced monitoring is not necessarily machine learning |
| Edwards Hypotension Prediction Index | Edwards Lifesciences | Device and evidence lead | Hemodynamic warning before hypotensive events | Accuracy and intervention studies answer different questions; verify exact platform, indication, and endpoint |
| ICU early-warning systems | Multiple vendors | Product class | Sepsis, deterioration, respiratory failure, or other risks | No class-wide performance claim is defensible; local workflow, threshold, population, and response protocol are part of the intervention |
See AI in Critical Care. A prediction system does not improve outcomes merely by producing an earlier score; benefit depends on actionable information, response capacity, and comparative evaluation.
Pulmonology
| Tool | Vendor | Directory role | Principal use to verify | Evidence boundary |
|---|---|---|---|---|
| qXR | Qure.ai | Device and implementation lead | Chest-radiograph decision support, including tuberculosis-related workflows | Verify exact product, jurisdiction, population, and intended use |
| CAD4TB | Delft Imaging | International screening-workflow lead | Computer-aided chest-radiograph interpretation for tuberculosis screening | WHO recommendations concern defined screening uses and performance requirements, not a universal endorsement of every version or deployment |
| VIDA quantitative lung products | VIDA Diagnostics | Device lead | COPD, emphysema, airway, or interstitial-lung-disease quantification, depending on module | Verify each product and submission separately |
Mental Health / Psychiatry
| Tool | Vendor | Directory role | Principal use to verify | Evidence boundary |
|---|---|---|---|---|
| Woebot products | Woebot Health | Product-history and research lead | Conversational mental-health support | Product status, target condition, and availability have changed; verify the exact current offering and do not generalize from a bounded trial |
| Ginger / Headspace behavioral-health services | Headspace | Service and workflow lead | Coaching, triage, and human-supported behavioral care | A blended service is not equivalent to a standalone chatbot or autonomous clinical model |
| Mindstrong | Mindstrong Health | Historical research and product lead | Smartphone-derived behavioral signals | The company ceased operations; retain as a historical example, not a current procurement option |
Mental-health AI requires direct evaluation of crisis response, inappropriate reassurance, escalation, privacy, dependency, and population-specific harm. The Psychiatry chapter reviews these risks without asserting that conversational systems cause psychosis.
Obstetrics / Maternal-Fetal Medicine
| Tool | Vendor | Directory role | Principal use to verify | Evidence boundary |
|---|---|---|---|---|
| Sonio products | Sonio | Device and non-U.S. product lead | Fetal-ultrasound guidance and anomaly-related support, depending on module | Verify the exact product and jurisdiction; Sonio Suspect and Sonio Detect must not be conflated |
| Delivery Date AI | Ultrasound AI | Primary-record verified | Radiological software for delivery-date prediction | FDA granted De Novo DEN250007 on February 11, 2026; no March decision date, study denominator, or broader intended-use claim is asserted here |
| Prenatal and fetal-anomaly research models | Multiple investigators and vendors | Research category | Fetal-anomaly detection, measurement, or image-quality support | Study design, population, gestational age, equipment, and endpoint must be stated for each model |
Emergency Medicine
Many emergency AI tools cross specialties (e.g., head CT hemorrhage detection, CXR pneumothorax). See Radiology section for trauma/emergency imaging AI.
| Tool | Vendor | Directory role | Principal use to verify | Evidence boundary |
|---|---|---|---|---|
| Viz.ai modules | Viz.ai | Device and workflow-evidence leads | Stroke, pulmonary-embolism, aneurysm, or other triage functions, depending on exact module | The randomized stroke evidence applies to the evaluated Viz.ai stroke workflow, not every module or company product |
| Aidoc modules | Aidoc | Device and workflow leads | Imaging triage for intracranial hemorrhage, pulmonary embolism, fractures, and other findings, depending on exact submission | Verify the exact trade name, submission, anatomy, modality, and output; do not use one submission to support a platform-wide claim |
Urology
| Tool | Vendor | Directory role | Principal use to verify | Evidence boundary |
|---|---|---|---|---|
| Arterys prostate-imaging products | Arterys lineage | Product-history lead | Prostate MRI analysis | Verify current trade name, owner, exact record, and whether the function is detection, segmentation, or scoring |
| Watson for Genomics | IBM, discontinued | Historical product lead | Molecular interpretation support | Retain as implementation history; published oncology evidence mainly assessed concordance rather than patient outcomes |
Bladder-cancer image models and prostate tools are reviewed in Surgical Subspecialties. Company-reported Breakthrough Device Designation is not marketing authorization.
Orthopedics
OsteoDetect remains the clearest primary-record example in this directory. Its FDA De Novo order is limited to assisting identification and highlighting of distal-radius fractures on adult wrist radiographs. Broader fracture catalogs, operative planning systems, and recovery applications require their own records and evidence. See AI in Orthopedics.
Surgery
| Tool | Vendor | Primary FDA record | Intended-use boundary | Evidence boundary |
|---|---|---|---|---|
| Claire OCT System | Perimeter Medical Imaging AI | PMA P250008, approved March 3, 2026 | Adjunctive three-dimensional imaging with computer-aided detection used concurrently with physician interpretation and other standard methods to evaluate margins of excised lumpectomy tissue in patients with biopsy-confirmed breast cancer | FDA’s record authorizes the labeled use and an associated predetermined change control plan; this directory does not claim stand-alone performance, universal margin benefit, or superiority beyond the reviewed indication |
The word “adjunctive” is clinically important: Claire supplements physician interpretation and standard margin-evaluation methods; it does not replace them.
Limitations and Caveats
1. Regulatory pathways answer different questions: A 510(k) clearance generally rests on substantial equivalence to a legally marketed predicate. De Novo is a risk-based route for a novel class I or II device type. PMA evaluates whether valid scientific evidence provides reasonable assurance of safety and effectiveness for a class III device’s intended use. None of these labels alone establishes comparative patient-outcome benefit in every deployment setting.
2. Evidence design must match the claim: Retrospective external validation can address transportability. Prospective workflow studies can assess use in practice. Reader studies can assess assisted interpretation. Randomized or other credible comparative designs are needed for causal benefit claims. A randomized trial is not automatically the best or only design for every diagnostic-performance question.
3. Product identity changes: Trade names, owners, modules, software versions, indications, and commercial availability can change. FDA explicitly states that its public AI-enabled medical-device list is periodically updated and not comprehensive. The exact submission record and current labeling are the controlling public sources for a U.S. authorization claim.
4. Reimbursement is separate: Authorization does not create payment. Coverage, coding, contracting, and reimbursement vary by payer, setting, date, service, and billing arrangement. Current payer and coding sources must be checked before an economic model is built.
5. Integration is part of safety: An authorized device can still fail operationally through poor worklist integration, delayed alerts, alert duplication, incompatible equipment, absent downtime procedures, or unclear ownership of follow-up.
6. Performance varies by population: Models trained on one population may underperform in yours. Validate locally before widespread use.
7. Evidence provenance matters: Vendor studies can provide useful technical information, but vendor-reported findings should be labeled as such. Independent peer-reviewed evaluations, primary regulatory records, transparent protocols, and local monitoring answer different parts of the adoption question.
AI Tool Evaluation Framework
Before purchasing/implementing:
- Resolve the exact identity: Record the trade name, manufacturer, software version, submission number, decision date, product code, indication, contraindications, limitations, compatible equipment, and required user.
- Define the intended local role: Detection, triage, diagnosis, quantification, prognosis, acquisition guidance, documentation, or administrative support. State whether the output is autonomous, adjunctive, or informational.
- Map evidence to claims: Separate standalone performance, assisted-user performance, workflow effects, patient outcomes, equity, usability, and economics. A study should support only the endpoint and product configuration it actually evaluated.
- Assess transportability: Compare prevalence, spectrum of disease, demographic mix, equipment, acquisition protocol, clinical setting, reader expertise, missingness, and operational constraints.
- Model failure and response: Identify false-negative and false-positive consequences, out-of-distribution inputs, unusable outputs, alert routing, downtime procedures, override documentation, and escalation ownership.
- Evaluate integration: Confirm interface specifications, latency, identity matching, cybersecurity review, access controls, auditability, update procedures, support, and recovery from partial failure.
- Examine total cost: Include licensing, integration, validation, training, monitoring, incident response, upgrades, clinician time, and decommissioning. Use current local prices rather than generic return-on-investment assumptions.
- Run a risk-proportionate local evaluation: Choose sample size, comparator, endpoints, subgroup analyses, and prospective monitoring from the use case and consequences of error. A silent evaluation may be appropriate before outputs affect care.
- Predefine adoption and stop rules: Base thresholds on clinical consequences, current performance, uncertainty, workflow capacity, and stakeholder agreement. Do not copy arbitrary values from another institution.
- Monitor the deployed system: Track data quality, subgroup performance, alert burden, overrides, workflow failures, version changes, incidents, and patient-relevant outcomes where feasible.
The Vendor Evaluation Framework, AI Evaluation and Monitoring, Safety, and Workflow Integration chapters provide the full implementation route.
Does FDA authorization prove that a medical AI tool improves patient outcomes?
No. Authorization applies to a defined device, version, intended use, and regulatory standard. Patient-outcome benefit, local transportability, workflow safety, and comparative value require additional evidence.
Which FDA pathway applies to medical AI?
Depending on the device and its risk classification, an AI-enabled medical device may reach the U.S. market through 510(k) clearance, De Novo authorization, or PMA approval. Some software functions are not devices or may fall outside premarket review, but that determination must be made for the exact function.
Can evidence for one vendor’s product support another vendor’s tool?
No. Evidence is connected to a specific product, version, population, workflow, comparator, and endpoint. For example, the Viz.ai stroke-workflow trial should not be cited as randomized evidence for RapidAI or Brainomix.
What should a physician verify before using a listed tool?
The exact product and version, current regulatory record, intended use, limitations, evidence design, local population, integration, failure modes, institutional approval, monitoring plan, privacy and security controls, and current commercial availability should all be verified.
Conclusion
This directory provides a preservation-first starting point for exploring specialty-specific AI tools. It retains product names that readers search for while distinguishing verified primary records from research, history, international deployment, non-device support, and non-AI comparison.
The current FDA resource is the AI-Enabled Medical Devices list. Because the agency states that the list is not comprehensive and is updated periodically, each adoption decision should continue into the exact submission database, labeling, and relevant postmarket information.
Remember: AI is a clinical tool, not a substitute for accountable judgment, contextual understanding, or the patient relationship.