Loading
Skip to content
Miracle Minds AI
  • Try “HIPAA”, “model risk”, “FinOps” or “agentic”.

AI-enabled medical devices

An AI-enabled device pays off when it clears review on evidence that holds up in use, and when the model can then be updated without a new submission each time. Both depend on work done before the first filing: independent validation data and a change control plan the regulator has reviewed.

Part of AI for Life Sciences & Healthcare. 8 sources, checked .

AI-enabled medical devices authorized by FDA for marketing in the United States as of September 2026
1,600+
AI-enabled medical devices authorized by FDA for marketing in the United States as of September 2026
U.S. Food and Drug Administration, 2026(opens in a new tab)
of 182 recall events linked to FDA-cleared AI-enabled devices occurred within 12 months of clearance (950 devices, recall data matched November 2024)
43.4%
of 182 recall events linked to FDA-cleared AI-enabled devices occurred within 12 months of clearance (950 devices, recall data matched November 2024)
JAMA Health Forum, 2025(opens in a new tab)
of 794 AI-enabled medical devices authorized by FDA from 2023 through 2025 had an authorized predetermined change control plan
5.4%
of 794 AI-enabled medical devices authorized by FDA from 2023 through 2025 had an authorized predetermined change control plan
JAMA Health Forum, 2026(opens in a new tab)

The problem

FDA had authorized over 1,600 AI-enabled medical devices for marketing in the United States as of September 2026.1 Clearance does not settle whether a device performs in use. A study of 950 FDA-cleared AI-enabled devices found that 60 (6.3%) were associated with 182 recall events, and 79 of those recalls (43.4%) came within the first 12 months of clearance, about double the rate reported for all 510(k) devices.7 Diagnostic or measurement errors were the largest cause, with 109 recalls. Devices without reported clinical validation had more recalls per device, a mean of 3.4, than those with retrospective (1.9) or prospective (2.0) validation.7

The second test is whether the model can change after clearance. FDA's final guidance on predetermined change control plans, issued on August 18, 2025, lets a manufacturer have planned modifications reviewed once, as part of the marketing submission. Modifications that are specified and implemented in accordance with an authorized plan can then be made without a new marketing submission.2 Few manufacturers have used the route so far. Of 794 AI-enabled devices authorized from 2023 through 2025, 43 (5.4%) had an FDA-authorized plan.8 Without one, a significant change to the model generally waits for a new submission.

The quality system rules changed in 2026. The Quality Management System Regulation took effect on February 2, 2026. It amends 21 CFR Part 820 by incorporating ISO 13485:2016 by reference and, in FDA's words, "now specifically requires risk management."3 FDA reissued its premarket cybersecurity guidance on February 3, 2026, written against the new regulation.4 Under section 524B of the FD&C Act, a submission for a device that includes software, can connect to the internet and could be vulnerable to cybersecurity threats must include a plan to address postmarket vulnerabilities and a software bill of materials.4 FDA still lists its guidance on lifecycle management and marketing submissions for AI-enabled device software functions as a draft. For generative AI, it lists a discussion paper and advisory committee meetings in November 2024 and November 2025.1

Selling in the EU adds the AI Act on top of the MDR or IVDR. Joint guidance from the Artificial Intelligence Board and the Medical Device Coordination Group says an AI system that is a medical device, or a safety component of one, is high-risk under the AI Act when the device needs notified body assessment. That guidance gave 2 August 2027 as the date the high-risk obligations apply.5 The Digital Omnibus on AI has since moved it to 2 August 2028.6 The guidance also says that changes pre-determined by the manufacturer and assessed at the initial conformity assessment do not count as a substantial modification, the same idea as a US change control plan.5

What we build

  1. 01

    Design controls and traceability for the model lifecycle

    We connect requirements, risk analysis, datasets, training runs and verification results inside your quality system, so the design history is produced as the work happens and not rebuilt before a submission.

  2. 02

    Data management and sequestered test sets

    Versioned datasets with provenance, a documented reference standard process, test data locked away from developers, and coverage reports that show how well each subgroup and site in the intended use population is represented.

  3. 03

    Performance evaluation harness

    Pre-specified metrics, acceptance criteria and statistical tests that run on every candidate version, comparing it with the original and the last released version, overall and by subgroup. It covers imaging, diagnostic and generative AI functions.

  4. 04

    Predetermined change control plans and release gates

    We draft the description of modifications, modification protocol and impact assessment with your regulatory team, then encode the protocol in the release pipeline so an update that misses its criteria cannot ship.

  5. 05

    Post-market performance and vulnerability monitoring

    Monitoring of inputs, outputs and confirmed outcomes by site and software version, drift triggers that open a quality record, a software bill of materials for each release and a tracked patch cycle.

Where it goes wrong

The failure modes we design against from the first week, and the control for each.

  • Filing on thin clinical evidence

    Recalls of cleared AI-enabled devices were concentrated early after clearance and predominantly involved products lacking clinical validation.7 Control: validate on data from sites that played no part in development, fix acceptance criteria before results are seen, and report performance by subgroup in the submission.

  • Test data that leaks into development

    FDA's guidance describes training, tuning and test data as sequestered to prevent overfitting and overstated test performance, and asks manufacturers to describe the controls that keep test data separate and independent from development.2 Control: locked test sets with access controls, lineage for every record, and a written reference standard protocol.

  • Model changes made outside the authorized plan

    Significant modifications that are not specified in the authorized plan, or not implemented in accordance with it, likely require a new marketing submission, and FDA says deviations would generally cause the device to be adulterated and misbranded.2 Control: limit the plan to a small number of specific, verifiable modifications and enforce the modification protocol in the release pipeline.

  • Performance drift nobody measures

    Diagnostic or measurement errors accounted for 109 of 182 recall events in one study of cleared AI-enabled devices.7 FDA anticipates that manufacturers will monitor safety and effectiveness over time as modifications are implemented.2 Control: monitor inputs and outputs by site and version, set drift triggers for re-evaluation, and route signals into complaint handling and corrective action.

Rules that apply

  • FDA guidance on Predetermined Change Control Plans for AI-enabled device software functions (August 2025)

    Primary text on fda.gov(opens in a new tab)

    A plan should contain a Description of Modifications, a Modification Protocol and an Impact Assessment, and must be reviewed and established as part of a marketing authorization. Changes made in accordance with it need no new marketing submission.

  • Quality Management System Regulation, 21 CFR Part 820 (effective February 2, 2026)

    Primary text on fda.gov(opens in a new tab)

    Applies to finished device manufacturers. Incorporates ISO 13485:2016 by reference, including design and development requirements, and specifically requires risk management.

  • FD&C Act section 524B and FDA premarket cybersecurity guidance (February 2026)

    Primary text on fda.gov(opens in a new tab)

    Submissions for cyber devices must include a plan to monitor, identify and address postmarket vulnerabilities, processes that give reasonable assurance the device is cybersecure, including updates and patches, and a software bill of materials.

  • EU AI Act with the MDR and IVDR: joint guidance MDCG 2025-6 (June 2025)

    Primary text on health.ec.europa.eu(opens in a new tab)

    AI that is a medical device, or a safety component of one, and needs notified body assessment under the MDR or IVDR is a high-risk AI system. It must comply with both regimes, with the AI Act requirements assessed in the MDR or IVDR conformity assessment.

How to measure it

  • Performance on sequestered, multi-site test data against pre-specified acceptance criteria, overall and by subgroup
  • Real-world performance compared with the cleared claims, by site and software version
  • Share of model updates released under the authorized change control plan without a new submission
  • Time from a drift signal or complaint to a validated, released update
  • Time to patch known vulnerabilities, and releases shipped with a complete software bill of materials

Questions buyers ask

Do we need a new 510(k) every time we retrain an AI medical device model?

Not if the change was planned and reviewed in advance. FDA's August 2025 guidance lets you include a predetermined change control plan in the marketing submission. It describes the planned modifications, how each will be developed and validated, and an impact assessment. Changes made in accordance with an authorized plan need no new submission. Significant changes outside the plan likely do, so define modifications narrowly and enforce the protocol in your release process.

Does the EU AI Act apply to AI medical device software?

Yes, if the software is a medical device, or a safety component of one, and the device needs notified body assessment under the MDR or IVDR. It is then a high-risk AI system. After the 2026 Digital Omnibus on AI, those obligations apply from 2 August 2028. Joint EU guidance says the AI Act requirements are assessed within the MDR or IVDR conformity assessment, and its quality system elements can sit inside your existing one.

Why do AI-enabled medical devices get recalled?

Mostly for errors in what the device measures or reports. In a 2025 study of 950 FDA-cleared AI-enabled devices, 6.3% were associated with a recall, and diagnostic or measurement errors accounted for 109 of 182 recall events. About 43% of recalls came within a year of clearance. Devices with no reported clinical validation had more recalls each. Independent validation before filing and performance monitoring after release are the practical response.

Sources

Every figure on this page comes from the sources below. We checked each link on .

  1. 1.Artificial Intelligence-Enabled Medical Devices(opens in a new tab)U.S. Food and Drug Administration, September 22, 2026 · fda.gov
  2. 2.Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions(opens in a new tab)U.S. Food and Drug Administration, August 18, 2025 · fda.gov
  3. 3.Quality Management System Regulation (QMSR)(opens in a new tab)U.S. Food and Drug Administration, February 2, 2026 · fda.gov
  4. 4.Cybersecurity in Medical Devices: Quality Management System Considerations and Content of Premarket Submissions(opens in a new tab)U.S. Food and Drug Administration, February 3, 2026 · fda.gov
  5. 5.MDCG 2025-6 / AIB 2025-1: Interplay between the Medical Devices Regulation (MDR) & In vitro Diagnostic Medical Devices Regulation (IVDR) and the Artificial Intelligence Act (AIA)(opens in a new tab)Medical Device Coordination Group and Artificial Intelligence Board (published by the European Commission), June 19, 2025 · health.ec.europa.eu
  6. 6.Regulation (EU) 2026/1744 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 (Digital Omnibus on AI)(opens in a new tab)European Parliament and Council of the European Union, July 8, 2026 · eur-lex.europa.eu
  7. 7.Early Recalls and Clinical Validation Gaps in Artificial Intelligence-Enabled Medical Devices(opens in a new tab)JAMA Health Forum, August 22, 2025 · jamanetwork.com
  8. 8.Use and Public Reporting of Predetermined Change Control Plans for AI-Enabled Medical Devices(opens in a new tab)JAMA Health Forum, May 22, 2026 · jamanetwork.com