Ketryx Comments to FDA on the Regulation of Generative AI-Enabled Medical Devices

Ketryx

and

•

October 19, 2026

Risk Management

Section link

The following is a reproduction of comments Ketryx submitted to the U.S. Food and Drug Administration in response to the Center for Devices and Radiological Health's discussion paper, Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback (Docket No. FDA-2026-N-7874). The comments are reproduced as submitted.

Docket No. FDA-2026-N-7874

Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback

Introduction

Ketryx appreciates the opportunity to comment on the Center for Devices and Radiological Health’s discussion paper on generative AI (GenAI)-enabled medical devices. Ketryx develops lifecycle and quality management software used by medical device manufacturers, including developers of AI-enabled devices, to manage design controls, risk management and change control under 21 CFR Part 820 and ISO 13485.

We support CDRH’s risk-based, total product life cycle (TPLC) approach. Our comments share a single theme: the proposed framework will work best when it builds on the design control, risk management, supplier control and change control processes manufacturers already operate under the Quality Management System Regulation. Anchoring new expectations in these established processes supports least burdensome review, makes evidence verifiable over the life of the device, and helps beneficial devices reach patients sooner without reducing safety.

Our specific recommendations, organized by the discussion questions they address, follow.

Recommendations

Topic and FDA questions	Ketryx recommendation	Benefit to patients
Risk framework (Q1, Q2, Q5). Two-axis framework: device activity (how directive or autonomous the function is) against the consequence of relying on an incorrect output.	Support. We recommend that FDA align the consequences axis with severity as defined under ISO 14971 and with IEC 62304 software safety classification, so manufacturers extend the risk management process they already operate rather than run a parallel one. We also recommend adding traceability of an output to its primary source material as a risk modifier, and publishing worked examples that place representative functions on the framework. A framework built on established standards will be applied more consistently across quality, regulatory and engineering teams.	Review effort is concentrated on the functions that can cause harm, and lower-risk informational tools reach patients and clinicians sooner. Consistency with existing standards reduces the chance of errors arising from competing or duplicative risk processes.
Benchmarking as design verification (Q7, Q9, Q10). Competency-based premarket evaluation: prespecified benchmarks for safety, clinical proficiency and generalizability, followed by clinical confirmation.	Support. We recommend that FDA state that benchmark suites, scoring rubrics, datasets and results are design verification records under the Quality Management System Regulation (ISO 13485 §7.3), traced to design inputs, requirements and identified hazards. Treating benchmarking within existing design controls answers a question the paper leaves open: how design inputs and acceptance criteria for GenAI-enabled devices should be established and documented. Over time, FDA should work toward recognized consensus standards for benchmarking methods.	Benchmarks that are versioned and traceable can be re-run and audited, so safety claims remain verifiable after authorization.
Greater reliance on postmarket monitoring (Q18). Accepting greater premarket uncertainty in exchange for greater reliance on postmarket monitoring.	Support, with conditions. This trade-off is only as strong as the surveillance system behind it. Reduced premarket evidence should be accepted only where the manufacturer can demonstrate a highly reliable postmarket monitoring system and the ability to remediate quickly. That system should address not only adverse events but also general product feedback. We recommend FDA define its minimum characteristics: a prespecified monitoring plan with quantitative thresholds, defined detection timeframes, documented escalation into corrective and preventive action (CAPA), defined remediation timeframes, and traceability linking each monitoring signal to the benchmark element and hazard it relates to. This approach is less appropriate where harm may be severe, fast-acting or irreversible. GenAI-enabled devices that fail to perform as intended must be both identified and remediated quickly before harm accumulates.	Patients gain earlier access to beneficial devices, while problems that emerge in real-world use are detected and corrected quickly.
Supervisory agents in postmarket monitoring (Q19, Q20). Periodic re-benchmarking, clinician review and drift monitoring, potentially carried out by machine-based supervisory agents.	Support. We recommend FDA clarify that supervisory agents used for postmarket monitoring are quality system software, validated in proportion to risk under the Quality Management System Regulation (ISO 13485 §4.1.6), 21 CFR Part 11 where applicable, and FDA’s Computer Software Assurance guidance. Validation of the agent itself should include agreement testing against clinician-adjudicated samples, prespecified performance thresholds for the agent, periodic revalidation, and human review of every escalation. The agent is then controlled through the manufacturer’s quality system rather than reviewed as part of the device.	Continuous monitoring detects drift and unsafe outputs faster than periodic sampling alone, and clinicians spend their review time on flagged cases.
Change control and re-benchmarking (Q22, Q23). The premarket benchmark serves as the baseline for evaluating changes. FDA asks how PCCPs can work when future changes cannot be fully prespecified.	Strongly support. We propose an outcome-based approach to Predetermined Change Control Plans (PCCPs). A change that meets prespecified acceptance criteria with no regression on the safety, robustness and subgroup performance elements (S.1 to S.3, R.1 and R.2) would be managed within the manufacturer’s quality system; a regression on any of these elements would require FDA review before release. The gating elements and acceptance criteria should scale with the device’s position on the risk framework, and FDA should accept automated re-benchmarking evidence generated by the manufacturer’s controlled release process in suitable form. Many devices already log performance in clinical use; manufacturers should be able to cite this real-world performance data, under documented data integrity controls, to support PCCP justifications.	Safety improvements, such as corrected guardrails, can reach patients in weeks rather than waiting for a new submission, and a regression is caught before the change reaches patients. Real-world performance data grounds design and change decisions in actual clinical use.
Third-party foundation models (Q24, Q25). The foundation model developer can change the model underlying the device. FDA proposes voluntary Foundation Model Master Files (MAFs).	Support MAFs, though they are not sufficient on their own. We recommend that a third-party foundation model be treated as off-the-shelf software under the manufacturer’s supplier controls (ISO 13485 §7.4). This includes pinned model versions, quality agreements that require advance notice of changes, and treating any version change as a mandatory trigger for re-benchmarking before deployment.	The device a patient uses keeps the behavior that was evaluated. An upstream model update cannot change it without verification.
Additional considerations, including agentic AI (Q14, Q26). Additional considerations for premarket and postmarket evaluation of GenAI-enabled and agentic devices.	Comparator traceability. Benchmarking against a median clinician or clinician panel creates a traceability obligation whose cost and scope are not yet defined. The comparator becomes a design input, so FDA should clarify the documentation expected (panel composition, qualifications, rubric version and date) so that re-benchmarking after a change compares like with like. Timeliness and critical monitoring points. Effective oversight depends on how quickly a problem is detected and acted on. We recommend that manufacturers define critical monitoring points and maximum detection-to-action intervals, proportionate to risk. For agentic devices, critical monitoring points should include each consequential action (such as tool use, clinical orders or control of another device), with logs sufficient to reconstruct the full action sequence.	GenAI-enabled devices that fail to perform as intended can be identified quickly and corrected before harm accumulates.