Ketryx (Erez Kaminski; Paul Jones)

See attached file(s)

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 Ketryx recommendation Benefit to patients
questions

Risk framework (Q1, Support. We recommend that FDA align the Review effort is
Q2, Q5). Two-axis consequences axis with severity as defined under concentrated on the
framework: device ISO 14971 and with IEC 62304 software safety functions that can cause
activity (how classification, so manufacturers extend the risk harm, and lower-risk
directive or management process they already operate rather informational tools
autonomous the than run a parallel one. We also recommend reach patients and
function is) against adding traceability of an output to its primary clinicians sooner.
the consequence of source material as a risk modifier, and publishing Consistency with
relying on an worked examples that place representative existing standards
incorrect output. functions on the framework. A framework built on reduces the chance of
established standards will be applied more errors arising from
consistently across quality, regulatory and competing or
engineering teams. duplicative risk
processes.
Benchmarking as Support. We recommend that FDA state that Benchmarks that are
design verification benchmark suites, scoring rubrics, datasets and versioned and traceable
(Q7, Q9, Q10). results are design verification records under the can be re-run and
Competency-based Quality Management System Regulation (ISO audited, so safety claims
premarket 13485 §7.3), traced to design inputs, requirements remain verifiable after
evaluation: and identified hazards. Treating benchmarking authorization.
prespecified within existing design controls answers a question
benchmarks for the paper leaves open: how design inputs and
safety, clinical acceptance criteria for GenAI-enabled devices
proficiency and should be established and documented. Over time,
generalizability, FDA should work toward recognized consensus
followed by clinical standards for benchmarking methods.
confirmation.

Greater reliance on Support, with conditions. This trade-off is only as Patients gain earlier
postmarket strong as the surveillance system behind it. access to beneficial
monitoring (Q18). Reduced premarket evidence should be accepted devices, while problems
Accepting greater only where the manufacturer can demonstrate a that emerge in realpremarket highly reliable postmarket monitoring system and world use are detected
uncertainty in the ability to remediate quickly. That system should and corrected quickly.
exchange for greater address not only adverse events but also general
reliance on product feedback. We recommend FDA define its
postmarket minimum characteristics: a prespecified monitoring
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. GenAIenabled devices that fail to perform as intended
must be both identified and remediated quickly
before harm accumulates.
Supervisory agents Support. We recommend FDA clarify that Continuous monitoring
in postmarket supervisory agents used for postmarket monitoring detects drift and unsafe
monitoring (Q19, are quality system software, validated in outputs faster than
Q20). Periodic re- proportion to risk under the Quality Management periodic sampling alone,
benchmarking, System Regulation (ISO 13485 §4.1.6), 21 CFR Part and clinicians spend
clinician review and 11 where applicable, and FDA’s Computer Software their review time on
drift monitoring, Assurance guidance. Validation of the agent itself flagged cases.
potentially carried should include agreement testing against clinicianout by machine- adjudicated samples, prespecified performance
based supervisory thresholds for the agent, periodic revalidation, and
agents. human review of every escalation. The agent is
then controlled through the manufacturer’s quality
system rather than reviewed as part of the device.

Change control and Strongly support. We propose an outcome-based Safety improvements,
re-benchmarking approach to Predetermined Change Control Plans such as corrected
(Q22, Q23). The (PCCPs). A change that meets prespecified guardrails, can reach
premarket acceptance criteria with no regression on the patients in weeks rather
benchmark serves as safety, robustness and subgroup performance than waiting for a new
the baseline for elements (S.1 to S.3, R.1 and R.2) would be submission, and a
evaluating changes. managed within the manufacturer’s quality system; regression is caught
FDA asks how PCCPs a regression on any of these elements would before the change
can work when require FDA review before release. The gating reaches patients. Realfuture changes elements and acceptance criteria should scale with world performance data
cannot be fully the device’s position on the risk framework, and grounds design and
prespecified. FDA should accept automated re-benchmarking change decisions in
evidence generated by the manufacturer’s actual clinical use.
controlled release process in suitable form. Many
devices already log performance in clinical use;
manufacturers should be able to cite this realworld performance data, under documented data
integrity controls, to support PCCP justifications.

Third-party Support MAFs, though they are not sufficient on The device a patient
foundation models their own. We recommend that a third-party uses keeps the behavior
(Q24, Q25). The foundation model be treated as off-the-shelf that was evaluated. An
foundation model software under the manufacturer’s supplier upstream model update
developer can controls (ISO 13485 §7.4). This includes pinned cannot change it
change the model model versions, quality agreements that require without verification.
underlying the advance notice of changes, and treating any
device. FDA version change as a mandatory trigger for reproposes voluntary benchmarking before deployment.
Foundation Model
Master Files (MAFs).
Additional Comparator traceability. Benchmarking against a GenAI-enabled devices
considerations, median clinician or clinician panel creates a that fail to perform as
including agentic AI traceability obligation whose cost and scope are intended can be
(Q14, Q26). not yet defined. The comparator becomes a design identified quickly and
Additional input, so FDA should clarify the documentation corrected before harm
considerations for expected (panel composition, qualifications, rubric accumulates.
premarket and version and date) so that re-benchmarking after a
postmarket change compares like with like.
evaluation of GenAIenabled and agentic Timeliness and critical monitoring points. Effective
devices. 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.

Ketryx appreciates CDRH’s engagement on this topic and would welcome the opportunity to discuss
these recommendations further.

Respectfully submitted,

Erez Kaminski, CEO, Ketryx

Paul Jones, VP of RA/QA, Ketryx

Contact email: paul@ketryx.com