Appendix G: FDA AI-Enabled Device Evidence Tracker
FDA marketing authorization tells a physician that a device met a regulatory standard for a stated intended use; it does not say whether the device has been tested in a trial, in a prospective study, or only on archived cases. This tracker records both. It keeps a dated, versioned copy of FDA’s list of AI-enabled medical devices, summarizes it by year, pathway, and review panel, and assigns an evidence grade to 33 widely deployed device families after reading the primary peer-reviewed studies behind each one.
At a Glance
- Version 1.0.0, published October 30, 2026. FDA data downloaded October 30, 2026; grades assigned October 30, 2026.
- Graded device families: 33. Grade A: 8. Grade B: 11. Grade C: 13. Grade D: 1.
- A grade describes the strongest study design found for a device. It does not say whether the device works in a given hospital, and a grade A device can have mixed or null results.
- Downloads: graded devices (CSV), full FDA list snapshot (CSV), tracker record (JSON).
What FDA’s List Does and Does Not Mean
- It does mean FDA authorized the device for marketing in the United States for a stated intended use, through 510(k) clearance, De Novo authorization, or premarket approval, and that FDA identified the device as AI-enabled from its public records.
- It does not mean the device improves patient outcomes, outperforms clinicians, or has been tested in a randomized trial. Most entries are 510(k) clearances, which rest on substantial equivalence to an earlier device rather than on new clinical outcome data.
- It does not mean the device has been validated in a given hospital’s population, equipment, or workflow, or that a later software version performs like the version FDA reviewed.
- It does not mean the list is complete. FDA states the list is not comprehensive and is updated periodically.
Authorization is not evidence of clinical benefit. Entries outside the graded set below are marked “not graded” in the data files, not given a default grade.
How to Read the Evidence Grades
Each grade answers one question: what is the strongest peer-reviewed study design that has tested this device, or the algorithm it was built from, in people? Grades are assigned from the study’s own methods section or abstract, not from the vendor’s description of it.
| Grade | Definition | Typical study |
|---|---|---|
| A | A randomized trial, or a prospective study with a concurrent comparison group, that measured patient outcomes or care processes with the device in clinical use, published in a peer-reviewed journal. | Cluster randomized trial of an alerting tool; randomized screening trial |
| B | A prospective peer-reviewed accuracy or performance study in the intended population and setting, including FDA pivotal studies published in journals. | Consecutive patients scanned with the device and checked against a reference standard |
| C | Retrospective peer-reviewed validation only. | Reader study on an enriched set of archived images; retrospective cohort |
| D | No peer-reviewed study of the device found at the grading date. FDA’s decision summary is the only public performance source. | 510(k) summary with sponsor-run testing |
Terms used in the tables:
- 510(k) clearance means FDA found the device substantially equivalent to a legally marketed predicate device. De Novo authorization creates a new low-to-moderate-risk device type when no predicate exists. PMA (premarket approval) is the pathway for the highest-risk class of devices.
- A reader study asks clinicians to interpret a fixed set of archived cases with and without AI. It measures diagnostic performance under test conditions, not what happens to patients.
- An enriched case set contains more disease than routine practice, which inflates positive predictive value and can change how readers behave.
- Outcome direction states what the strongest studies found: favorable, mixed, null, or accuracy only when no outcome was measured.
Graded Devices
The FDA record column links the submission the grade is anchored to, usually the first authorization of that product on FDA’s list. Later versions and modules of the same product often have their own submission numbers; the graded devices file lists every matching entry on the current FDA list.
Radiology
| Device and company | FDA record | Grade | Strongest peer-reviewed evidence | What the study showed | Limits |
|---|---|---|---|---|---|
| Viz LVO (ContaCT), Viz.ai. Flags suspected large vessel occlusion on CT angiography and alerts the stroke team by phone. | DEN170073, De Novo, February 2018 | A | Martinez-Gutierrez et al., 2023, JAMA Neurology. Stepped-wedge cluster randomized trial at 4 comprehensive stroke centers; 243 patients treated with thrombectomy. | Door-to-groin time was 11.2 minutes shorter after activation of automated LVO detection with secure messaging (95% CI 4.2 to 18.2); functional independence did not differ in exploratory analysis (odds ratio 1.3, 95% CI 0.42 to 4.0). Also Sachdev et al., 2026, AJNR: Retrospective single-center comparison: Viz.ai detected 108 of 147 LVOs (73%) on CT angiography versus 144 (98%) for RapidAI. | Outcome: Favorable (care process). Version: Trial used the Viz.AI LVO software; version not reported in the abstract. Process outcome; the trial was not powered for functional outcome. Cluster activation order was random except that the academic center went last. |
| Rapid LVO, iSchemaView (RapidAI). Flags suspected large vessel occlusion on CT angiography for triage. | K200941, 510(k), July 2020 | C | Sachdev et al., 2026, AJNR. Retrospective review of 1,589 consecutive stroke alerts at one comprehensive stroke center, both products run in parallel. | RapidAI detected 144 of 147 LVOs (98%) compared with 108 (73%) for Viz.ai, and correctly classified 94% versus 91% of LVO-negative cases. | Outcome: Accuracy only. Version: Version not reported in the abstract. Single center. No randomized or prospective outcome study of this product was found. |
| Aidoc BriefCase (intracranial hemorrhage triage), Aidoc. Triage and notification software for CT findings; this row grades the intracranial hemorrhage module only. | K180647, 510(k), August 2018 | C | Chavoshi et al., 2025, npj Digital Medicine. Retrospective evaluation of 101,944 non-contrast head CTs in a 17-facility academic health system; reference labels extracted from reports by a language model. | Sensitivity 82.2% and specificity 97.6% for detecting intracranial hemorrhage; sensitivity was lower for subacute (45.5%) and chronic (54.8%) bleeds. | Outcome: Accuracy only. Version: Commercial BriefCase ICH Triage as deployed 2023 to 2025. Listed BriefCase triage entries cover several findings; evidence for the hemorrhage module does not transfer to other modules. |
| Aidoc CARE multi-triage CT, Aidoc. Foundation-model triage of multiple acute findings on CT. | K252970, 510(k), January 2026 | D | No peer-reviewed study of the cleared product found. | FDA 510(k) summary is the only public performance source located. | Outcome: Not assessed. Grade D reflects absent published evidence at the grading date, not evidence of poor performance. |
| Transpara, ScreenPoint Medical. Assigns a malignancy risk score to screening mammograms and marks suspicious regions. | K181704, 510(k), November 2018 | A | Lång et al., 2023, Lancet Oncology (MASAI). Randomised, controlled, non-inferiority screening-accuracy trial at 4 Swedish screening sites; 80,033 women; prespecified clinical safety analysis. | Cancer detection was 6.1 versus 5.1 per 1,000 screened with AI-supported versus standard double reading (ratio 1.2, 95% CI 1.0 to 1.5), with 44.3% lower screen-reading workload. | Outcome: Favorable for workload; the higher detection rate was not statistically significant (p=0.052). Version: Transpara 1.7.0, which is on the FDA list as K210404. The trial’s primary endpoint, interval cancer rate, was still in follow-up at this publication. Swedish double-reading workflow. |
| Lunit INSIGHT MMG, Lunit. Detects and scores suspicious findings on screening mammograms. | K211678, 510(k), November 2021 | B | Dembrower et al., 2023, Lancet Digital Health (ScreenTrustCAD). Prospective, population-based, paired-reader non-inferiority study at one Stockholm hospital; 55,581 women. | One radiologist plus AI detected 261 cancers versus 250 for two radiologists (relative proportion 1.04, 95% CI 1.00 to 1.09). | Outcome: Favorable (detection). Version: INSIGHT MMG version 1.1.6. Single site, Philips equipment, Swedish double reading with consensus. Funding included Lunit. |
| PowerLook Tomo Detection / ProFound AI for DBT, iCAD. Detects suspicious soft-tissue lesions and calcifications on digital breast tomosynthesis. | P160009, PMA, March 2017 | C | Conant et al., 2019, Radiology: Artificial Intelligence. Retrospective multireader study; 24 radiologists read 260 DBT exams (65 cancers) with and without AI. | Reader AUC for detecting malignant lesions rose from 0.795 without AI to 0.852 with AI; reading time fell 52.7%. | Outcome: Favorable (reader study). Version: PowerLook Tomo Detection 2.0, on the FDA list as K182373. Enriched case set; several authors were iCAD employees. |
| Koios DS, Koios Medical. Classifies breast ultrasound lesions into risk categories. | K190442, 510(k), July 2019 | C | Mango et al., 2020, AJR. Retrospective multicenter reader study; 900 lesions read by 15 physicians with and without decision support. | Mean reader AUC for classifying lesions rose from 0.83 with ultrasound alone to 0.87 with decision support. | Outcome: Favorable (reader study). Version: Koios DS for Breast. Enriched case mix (47.8% malignant). |
| Annalise Enterprise CXR, Annalise-AI (Harrison.ai). Triage of specified findings on chest radiographs, such as pneumothorax. | K213941, 510(k), February 2022 | C | Seah et al., 2021, Lancet Digital Health. Retrospective multireader multicase study; 20 radiologists, 2,568 enriched cases, 127 findings. | Macroaveraged reader AUC across 127 findings rose from 0.713 unassisted to 0.808 with the model. | Outcome: Favorable (reader study). Version: Multi-finding research model underlying Annalise CXR; FDA entries cover triage of specific findings, not all findings studied. Company-funded and largely company-authored. |
| Lunit INSIGHT CXR Triage, Lunit. Sorts chest radiographs into normal, non-urgent, and urgent for worklist triage. | K211733, 510(k), November 2021 | B | Sridharan et al., 2024, European Journal of Radiology. Prospective clinical evaluation in Singapore (authors’ design label); 43 radiologists blinded to AI output. | For the urgent category, sensitivity 82% and specificity 99% against radiologist classification. | Outcome: Accuracy only. Version: LUNIT INSIGHT CXR Triage. Graded from the abstract; site count not stated there. |
| BoneView, Gleamer. Identifies and highlights fractures on radiographs as a concurrent reading aid. | K212365, 510(k), March 2022 | C | Guermazi et al., 2022, Radiology. Retrospective multireader multicase study; 480 exams, 24 readers from six specialties, fracture prevalence set at 50%. | Per-patient sensitivity for fractures rose from 64.8% without AI to 75.2% with AI; specificity was noninferior. | Outcome: Favorable (reader study). Version: Gleamer BoneView. Funded by Gleamer, which built the study sample and design. |
| Optellum Virtual Nodule Clinic, Optellum. Estimates malignancy risk for indeterminate pulmonary nodules on CT. | K202300, 510(k), March 2021 | C | Kim et al., 2022, Radiology. Retrospective multireader multicase study; 300 chest CTs, 12 readers (radiologists and pulmonologists). | Readers’ average AUC for malignancy rose from 0.82 to 0.89 with the tool; interobserver agreement improved. Also Massion et al., 2020, AJRCCM: Development and retrospective external validation of the underlying LCP-CNN model. | Outcome: Favorable (reader study). Version: Virtual Nodule Clinic version 2.0.0. Enriched set, 50% malignant. |
Cardiology
| Device and company | FDA record | Grade | Strongest peer-reviewed evidence | What the study showed | Limits |
|---|---|---|---|---|---|
| HeartFlow FFRct Analysis, HeartFlow. Computes fractional flow reserve from coronary CT angiography. | K152733, 510(k), January 2016 | A | Curzen et al., 2021, European Heart Journal (FORECAST). Randomized trial; 1,400 patients with stable chest pain at 11 UK centres. | CT angiography with selective FFRct did not change 9-month cardiac costs or major adverse events compared with standard care, but reduced invasive angiography (19% versus 25%). Also Douglas et al., 2015, European Heart Journal (PLATFORM): Prospective multicenter comparison of consecutive cohorts at 11 sites: among patients with planned invasive angiography, no obstructive disease was found in 12% with FFRct-guided care versus 73% with usual care. | Outcome: Mixed (fewer invasive tests; no cost or event difference). Version: Earlier HeartFlow versions; the first entry on the FDA list is K152733. Earlier HeartFlow authorizations that are not on the FDA AI list are not shown. |
| FFRangio, CathWorks. Derives fractional flow reserve from routine coronary angiograms. | K182149, 510(k), December 2018 | B | Fearon et al., 2019, Circulation (FAST-FFR). Prospective multicenter international accuracy study; 10 centers, 301 patients, 319 vessels. | Per-vessel sensitivity 94% and specificity 91% for predicting pressure-wire FFR of 0.80 or less. | Outcome: Accuracy only. Version: FFRangio as used in FAST-FFR. |
| Caption Guidance, Caption Health (Bay Labs at authorization). Guides novice operators in real time to acquire diagnostic echocardiographic views. | DEN190040, De Novo, February 2020 | B | Narang et al., 2021, JAMA Cardiology. Prospective multicenter diagnostic study at 2 hospitals; 8 nurses without echo experience scanned 240 patients. | Nurse scans with AI guidance were judged diagnostic for left ventricular size, function, and pericardial effusion in 237 of 240 patients (98.8%). | Outcome: Favorable (image acquisition). Version: Deep-learning guidance software as tested in 2019. Measures image quality, not patient outcomes. |
| Us2.ai, Us2.ai (Eko.ai Pte. Ltd.). Automated echocardiographic measurement and reporting. | K210791, 510(k), July 2021 | C | Tromp et al., 2022, Lancet Digital Health. Multicohort validation against manual measurements in existing datasets, with a 142-study prospective re-measurement subset. | Automated measurements agreed with expert sonographer measurements across cohorts from Asia, Canada, Taiwan, and the US. | Outcome: Accuracy only. Version: Research workflow underlying the product. Graded conservatively: the prospective component is a small re-measurement subset. |
| EchoGo Heart Failure, Ultromics. Detects heart failure with preserved ejection fraction from one apical four-chamber echo clip. | K240013, 510(k), September 2024 | C | Akerman et al., 2023, JACC: Advances. Model development with an independent multisite test set of 646 cases and 638 controls. | In independent testing, sensitivity 87.8% for identifying HFpEF; 7.3% of studies were nondiagnostic. | Outcome: Accuracy only. Version: Research model preceding EchoGo Heart Failure 2.0. Case-control design; supported by Ultromics. |
| ECG-AI Low Ejection Fraction, Anumana. Flags adults likely to have low left ventricular ejection fraction from a 12-lead ECG. | K232699, 510(k), September 2023 | A | Yao et al., 2021, Nature Medicine (EAGLE). Pragmatic cluster randomized trial; 120 primary care teams, 45 clinics, 22,641 adults. | New diagnoses of low EF within 90 days were 2.1% with AI results versus 1.6% with usual care (odds ratio 1.32). | Outcome: Favorable (diagnosis yield). Version: Trial tested the Mayo Clinic AI-ECG algorithm later licensed to Anumana; trial threshold EF 50% or less, cleared indication EF 40% or less. Outcome is new diagnosis, not clinical events. |
| Eko Low Ejection Fraction Tool (ELEFT), Eko Health. Flags low ejection fraction from single-lead ECG recorded by an ECG-enabled stethoscope. | K233409, 510(k), March 2024 | B | Bachtiger et al., 2022, Lancet Digital Health. Prospective, observational, multicentre study in London; 1,050 patients referred for echocardiography. | AUROC 0.85 for identifying LVEF of 40% or lower at the pulmonary auscultation position. | Outcome: Accuracy only. Version: AI-ECG retrained for single-lead input on the Eko DUO stethoscope; equivalence to the cleared tool is not stated. |
| Eko Murmur Analysis Software, Eko Health. Detects heart murmurs from digital stethoscope recordings. | K213794, 510(k), June 2022 | B | Chorba et al., 2021, Journal of the American Heart Association. Algorithm trained on prior recordings and tested in 962 patients enrolled in a prospective clinical study (NCT03458806). | Sensitivity 76.3% and specificity 91.4% for detecting murmurs against echocardiography and expert annotation. | Outcome: Accuracy only. Version: Eko murmur detection algorithm as described in 2021; an earlier Eko Analysis Software entry (K192004) is also on the FDA list. |
| Irregular Rhythm Notification Feature, Apple. Notifies smartwatch users of pulse irregularity suggestive of atrial fibrillation. | K212516, 510(k), October 2021 | B | Perez et al., 2019, NEJM (Apple Heart Study). Prospective, siteless, single-arm pragmatic study; 419,297 participants. | 0.52% received notifications; positive predictive value 0.84 for atrial fibrillation on a simultaneously worn ECG patch. | Outcome: Accuracy only. Version: Algorithm version studied predates the first listed entry. Funded by Apple. No comparison group. |
| Tempus ECG-AF, Tempus AI. Estimates risk of developing atrial fibrillation from a 12-lead ECG. | K233549, 510(k), June 2024 | C | Raghunath et al., 2021, Circulation. Retrospective development and validation on 1.6 million ECGs from 430,000 patients at one health system. | AUROC 0.85 for predicting new-onset atrial fibrillation within 1 year. | Outcome: Accuracy only. Version: Geisinger research model; Tempus co-authors and funding. |
| Viz HCM, Viz.ai. Flags suspected hypertrophic cardiomyopathy on a 12-lead ECG. | DEN230003, De Novo, August 2023 | C | Hillis et al., 2025, Circulation: Heart Failure. Retrospective case selection at 3 hospitals; 293 HCM-positive and 2,912 HCM-negative cases. | Sensitivity 68.4% and specificity 99.1% for detecting HCM; positive predictive value 13.7% at an assumed 1-in-500 prevalence. | Outcome: Accuracy only. Version: Viz HCM device. |
Ophthalmology
| Device and company | FDA record | Grade | Strongest peer-reviewed evidence | What the study showed | Limits |
|---|---|---|---|---|---|
| IDx-DR (LumineticsCore), Digital Diagnostics. Autonomous detection of more-than-mild diabetic retinopathy in primary care. | DEN180001, De Novo, April 2018 | A | Wolf et al., 2024, Nature Communications (ACCESS). Parallel randomized controlled trial; 164 youth aged 8 to 21 with diabetes at one pediatric center. | Diabetic eye exam completion within 6 months was 100% with point-of-care autonomous AI versus 22% with referral. Also Abràmoff et al., 2018, npj Digital Medicine: Prospective pivotal trial, 900 adults at primary care sites: sensitivity 87.2% and specificity 90.7% for detecting more-than-mild DR against reading-center grading. | Outcome: Favorable (exam completion). Version: IDx-DR as authorized in 2018; the trial population was younger than the adult indication. ACCESS tested youth, outside the adult indication of the 2018 authorization. |
| EyeArt, Eyenuk. Autonomous detection of referable and vision-threatening diabetic retinopathy. | K200667, 510(k), August 2020 | B | Ipp et al., 2021, JAMA Network Open. Prospective multicenter cross-sectional diagnostic study; 893 patients at 15 primary care and eye care sites. | Sensitivity 95.5% and specificity 85.0% for detecting more-than-mild DR without dilation against reading-center grading. | Outcome: Accuracy only. Version: EyeArt version 2.1.0. |
| AEYE-DS, AEYE Health. Autonomous detection of more-than-mild diabetic retinopathy, including with a handheld camera. | K221183, 510(k), November 2022 | B | Dvey-Aharon et al., 2026, Frontiers in Digital Health. Three prospective pivotal validation studies; more than 1,200 patients; desktop and handheld cameras. | Sensitivity 92% to 93% and specificity 89% to 94% for detecting more-than-mild DR against reading-center grading. | Outcome: Accuracy only. Version: AEYE-DS pivotal versions. Company-authored report of the FDA pivotal studies. |
Other Specialties
| Device and company | FDA record | Grade | Strongest peer-reviewed evidence | What the study showed | Limits |
|---|---|---|---|---|---|
| Acumen Hypotension Prediction Index, Edwards Lifesciences. Predicts intraoperative hypotension from the arterial waveform. | DEN160044, De Novo, March 2018 | A | Wijnberge et al., 2020, JAMA (HYPE). Single-center randomized trial; 68 patients undergoing elective noncardiac surgery. | Median time-weighted average of hypotension was 0.10 mm Hg with the early warning system versus 0.44 mm Hg with standard care. Also Maheshwari et al., 2020, Anesthesiology: Randomized pilot trial, 214 patients: index guidance did not reduce hypotension; half of alerts were not followed by treatment. | Outcome: Mixed (one positive and one null randomized trial). Version: Hypotension Prediction Index software. Benefit depended on a treatment protocol and on clinicians acting on alerts. |
| GI Genius, Cosmo Artificial Intelligence (Medtronic). Real-time computer-aided detection of colorectal polyps during colonoscopy. | DEN200055, De Novo, April 2021 | A | Repici et al., 2020, Gastroenterology. Multicenter randomized trial; 685 patients at 3 Italian centers. | Adenoma detection rate was 54.8% with computer-aided detection versus 40.4% without (relative risk 1.30). | Outcome: Favorable (adenoma detection). Version: GI Genius. Adenoma detection is a surrogate; interval cancer and mortality were not measured. |
| DermaSensor, DermaSensor. Handheld elastic scattering spectroscopy with an algorithm that recommends investigate further or monitor. | DEN230008, De Novo, January 2024 | B | Merry et al., 2025, Journal of Primary Care & Community Health (DERM-SUCCESS). Blinded, prospective, multicenter pivotal study at 22 primary care clinics in the US and Australia; 1,579 lesions. | Sensitivity 95.5% and specificity 20.7% for detecting skin cancer against histopathology; positive predictive value 16.6%. | Outcome: Accuracy only. Version: DermaSensor pivotal device. Low specificity: most lesions flagged for further investigation were benign. |
| Paige Prostate, Paige.AI. Flags suspicious areas on digitized prostate biopsy slides. | DEN200080, De Novo, September 2021 | C | da Silva et al., 2021, Journal of Pathology. Independent application to 600 biopsy part-specimens from 100 consecutive patients. | Part-specimen sensitivity 0.99 and specificity 0.93 for identifying slides needing further review. Also Raciti et al., 2020, Modern Pathology: Retrospective reader study with Paige Prostate Alpha: pathologist sensitivity rose from 74% to 90%. | Outcome: Accuracy only. Version: Paige Prostate (clinical-grade); the reader study used the Alpha version. |
| Sepsis ImmunoScore, Prenosis. Estimates sepsis risk within 24 hours from laboratory and clinical data once a blood culture is ordered. | DEN230036, De Novo, April 2024 | B | Bhargava et al., 2024, NEJM AI. Prospective enrollment at 5 US institutions; derivation, internal validation, and external validation (698 patients) cohorts. | AUC 0.81 for identifying sepsis within 24 hours in the external validation cohort. | Outcome: Accuracy only. Version: Sepsis ImmunoScore. Requires that a clinician already suspects infection. |
| Bayesian Health Sepsis Flagging Device (TREWS), Bayesian Health. Continuous EHR-based sepsis risk flag for adults in the emergency department and inpatient units. | K250680, 510(k), April 2026 | A | Adams et al., 2022, Nature Medicine. Prospective multi-site cohort study at 5 hospitals; 590,736 patients monitored, 6,877 sepsis patients analyzed. | When providers confirmed the alert within 3 hours, adjusted in-hospital mortality was 3.3 percentage points lower than when they did not. | Outcome: Favorable (observational). Version: TREWS, renamed Bayesian Health Sepsis Flagging Device. Observational: the comparison is by provider response to the alert, not by randomization. |
| Ceribell Status Epilepticus Monitor, Ceribell. Estimates seizure burden on point-of-care EEG and alerts for electrographic status epilepticus. | K223504, 510(k), May 2023 | C | Sheikh et al., 2026, Epilepsia. Retrospective test of algorithm version 8 on a previously used rapid-response EEG dataset. | Sensitivity 100% for detecting electrographic status epilepticus at 10% to 20% burden thresholds, with specificity 0.90. | Outcome: Accuracy only. Version: Automated Seizure Burden Estimator version 8. |
What the Graded Set Shows
Eight of the 33 graded device families have randomized or prospective comparative evidence of a change in care, and only two of those report clinical events or costs as main outcomes: the HeartFlow FFRct randomized trial found no difference in costs or major adverse events, and the TREWS sepsis study, which reported lower mortality, was observational. The rest of the comparative evidence measures intermediate endpoints: time to thrombectomy, cancer or adenoma detection, intraoperative hypotension, new diagnoses of low ejection fraction, and eye exam completion.
Grade C is the most common grade. Most radiology products graded here rest on reader studies of enriched archived cases, and those results do not transfer automatically to routine prevalence, local equipment, or unassisted workflow.
Version drift is common. In 9 of the 33 graded families, the strongest study tested an earlier version of the software, a research model the product was built from, or a population outside the cleared indication; the Limits column names each case.
The Full FDA List in Numbers
FDA describes its list as a resource for identifying AI-enabled devices authorized for marketing in the United States, built primarily by searching authorization summaries and device classifications for AI-related terms, and states that it is not comprehensive (FDA AI-Enabled Medical Devices).
The FDA list downloaded on October 30, 2026 (list page content current as of September 4, 2026) contains 1,614 AI-enabled devices with final decisions from September 1995 through June 29, 2026 (FDA AI-Enabled Medical Devices list, CSV export).
| Pathway | Devices | Share of list |
|---|---|---|
| 510(k) | 1,553 | 96.2% |
| De Novo | 40 | 2.5% |
| PMA | 21 | 1.3% |
| Total | 1,614 | 100% |
| Year | 510(k) | De Novo | PMA | Total |
|---|---|---|---|---|
| 1995–2015 | 31 | 2 | 9 | 42 |
| 2016 | 17 | 0 | 1 | 18 |
| 2017 | 23 | 2 | 2 | 27 |
| 2018 | 58 | 7 | 0 | 65 |
| 2019 | 80 | 0 | 0 | 80 |
| 2020 | 113 | 1 | 0 | 114 |
| 2021 | 122 | 6 | 1 | 129 |
| 2022 | 160 | 1 | 1 | 162 |
| 2023 | 219 | 6 | 1 | 226 |
| 2024 | 225 | 9 | 1 | 235 |
| 2025 | 329 | 2 | 4 | 335 |
| 2026 (through June 29) | 176 | 4 | 1 | 181 |
| Lead review panel | Devices | Share of list |
|---|---|---|
| Radiology | 1,230 | 76.2% |
| Cardiovascular | 154 | 9.5% |
| Neurology | 73 | 4.5% |
| Anesthesiology | 30 | 1.9% |
| Gastroenterology-Urology | 27 | 1.7% |
| Hematology | 22 | 1.4% |
| Pathology | 10 | 0.6% |
| Ophthalmic | 10 | 0.6% |
| General and Plastic Surgery | 9 | 0.6% |
| Clinical Chemistry | 9 | 0.6% |
| Orthopedic | 8 | 0.5% |
| Dental | 7 | 0.4% |
| Microbiology | 7 | 0.4% |
| General Hospital | 6 | 0.4% |
| Obstetrics and Gynecology | 5 | 0.3% |
| Clinical Toxicology | 5 | 0.3% |
| Ear Nose & Throat | 1 | 0.1% |
| Immunology | 1 | 0.1% |
Counts reflect the lead review panel FDA assigned, not the specialty that uses a device. A cardiac MRI analysis tool reviewed by the radiology panel counts as radiology.
Method
FDA data. The tracker downloads FDA’s CSV export of the AI-Enabled Medical Devices list, keeps the downloaded file (line endings normalized) with its SHA-256 checksum, and normalizes the records: ISO dates, the regulatory pathway inferred from the submission number prefix (K for 510(k), DEN for De Novo, P for PMA), and a link to each record in FDA’s databases in the form the FDA list page uses.
Inclusion in the graded set. Device families were chosen for clinical reach, not for strength of evidence: products that are widely deployed in U.S. practice, that recur in specialty chapters of this handbook, or that were the first of their kind on FDA’s list. Coverage spans radiology (stroke, breast, chest, musculoskeletal, and body imaging), cardiology, ophthalmology, gastroenterology, dermatology, pathology, anesthesiology, sepsis, and critical care EEG. Ambient documentation tools are not graded because they are not FDA-authorized medical devices.
Grading. For each family, PubMed and the publisher record were searched for peer-reviewed studies that named the product or its documented predecessor algorithm. The study with the strongest design was read in full where openly available, otherwise from its abstract; the Evidence basis field in the data file records which. Bibliographic identity (first author, journal, year, DOI) was checked against PubMed or Crossref. When a design label was ambiguous, the lower grade was assigned. Preprints, conference abstracts, vendor white papers, and FDA summaries do not raise a grade.
What a row grades. A row grades the algorithm and use studied. When the study tested an earlier version, a research model later licensed to the company, or a population outside the cleared indication, the grade is kept and the difference is stated, because FDA records do not establish equivalence between versions.
Refresh. A maintenance script (scripts/fda_ai_device_tracker.py) rebuilds the counts and data files from a new FDA export and writes a new dated snapshot. The graded rows change only when a person reviews new evidence. Each refresh records a new version and a changelog line below.
Limitations
- FDA’s list is not comprehensive, is updated periodically, and lags authorizations whose decision summaries were not yet published. Counts here describe the list, not every AI-enabled device on the market.
- The graded set is curated and covers a small fraction of the list. It should not be used to estimate the evidence base of all AI-enabled devices.
- A grade reflects study design, not quality within a design. Two grade A trials can differ greatly in size, risk of bias, and relevance.
- Absence of a peer-reviewed study (grade D) can reflect publication lag rather than absence of testing.
- Many graded studies were funded, designed, or co-authored by the manufacturer. The Limits column notes this where the study reports it.
- Searches were limited to English-language peer-reviewed literature indexed in PubMed and publisher records at the grading date.
Download the Data
| File | Contents |
|---|---|
| evidence-graded-devices.csv | One row per graded device family: grade, anchor submission, every matching entry on the current FDA list, study DOI, design, finding, version link, and evidence basis |
| fda-ai-enabled-devices.csv | The full FDA list, normalized, with a tracker identifier and the family evidence grade on entries that belong to a graded family; every other entry reads “not graded” |
| fda-ai-device-tracker.json | Version, retrieval date, FDA file checksum, grade definitions, summary counts, changelog, and graded rows |
| evidence-grades.csv | The reviewed evidence input from which the graded file is built |
| snapshots/2026-10-30.csv and .json | Dated copy of the normalized FDA list as retrieved on October 30, 2026, with the FDA file checksum and counts. Each refresh adds a new dated pair and leaves earlier ones unchanged. |
The data are released under the same CC BY 4.0 license as the handbook. FDA’s underlying list is a U.S. government work.
Cite This Tracker
Version 1.0.0, updated October 30, 2026. FDA data retrieved October 30, 2026.
Tegomoh B. FDA AI-Enabled Device Evidence Tracker, version 1.0.0 (FDA data retrieved 2026-10-30). In: The Physician AI Handbook. Updated October 30, 2026. https://physicianaihandbook.com/appendices/fda-ai-device-tracker.html@misc{tegomoh_fda_ai_tracker_1_0_0,
author = {Tegomoh, Bryan},
title = {FDA AI-Enabled Device Evidence Tracker},
version = {1.0.0},
year = {2026},
howpublished = {The Physician AI Handbook},
url = {https://physicianaihandbook.com/appendices/fda-ai-device-tracker.html},
note = {FDA data retrieved 2026-10-30; accessed YYYY-MM-DD}
}Cite the primary study, not this tracker, when a statement depends on a specific trial result.
Tracker Changelog
| Version | Date | Change |
|---|---|---|
| 1.0.0 | 2026-10-30 | First release. FDA list downloaded October 30, 2026 (list page content current as of September 4, 2026); 33 device families graded. |