Program On Regulation, Therapeutics, And Law (PORTAL) and MIT Critical Data

Please see attached file.

Division of Pharmacoepidemiology and Pharmacoeconomics
Brigham and Women’s Hospital • Harvard Medical School
41 Avenue Louis Pasteur, Suite 0B17, Boston, MA 02115
portalresearch.org • portal@bwh.harvard.edu

Laboratory for Computational Physiology
Massachusetts Institute of Technology
77 Massachusetts Avenue, E25-505, Cambridge, MA 02139
criticaldata.mit.edu

October 19, 2026

Dr. Michelle Tarver, M.D., Ph.D.
Director, Center for Devices and Radiological Health
Food and Drug Administration (FDA)
5630 Fishers Lane, Rm 1061
Rockville, MD 20852

Subject: Comments on FDA Discussion Paper, Considerations for the Regulation of Generative AIEnabled Medical Devices, Docket Number FDA-2026-N-7874

Dear Dr. Tarver:

As members of PORTAL and MIT Critical Data, we appreciate the opportunity to comment on CDRH’s
discussion paper. [1]

We agree with the premise that the evaluation of AI-enabled medical devices should be conducted in the
context in which the device is configured and used.

We generally agree with the sequence of steps proposed in the discussion paper, beginning with a risk
assessment based on the two axes of device activity and the consequences of relying on an incorrect
output, and progressing to device benchmarking, clinical confirmation, and ongoing postmarket
monitoring. [1] For each step, we recommend that the FDA confirm that the conditions outlined below
are met before relying on the results. Our comments are based on research conducted at PORTAL on
medical device regulation and research conducted at MIT Critical Data on the evaluation of clinical AI.
[2, 3, 4, 5]

Re: questions 1, 2, and 26

Before deciding the type of evidence or amount of safety testing a GenAI-enabled device needs, the
FDA should clearly map the bounds of what the device can and cannot do and what could happen if it
makes a mistake. A GenAI-enabled device that explains medications to patients poses different risks
from one that is given the ability to renew prescriptions without a physician’s review. [6, 7] The twoaxis framework should be designed to place a product only after the manufacturer has stated what it may
do, for which patients, with what information, and when a clinician must take over. A product’s place on
PORTAL • Program On Regulation, Therapeutics, And Law MIT Critical Data

the consequences axis should reflect the most serious harm an incorrect output could plausibly cause,
weighed against how often such an output is likely, even when the device’s usual tasks are harmless. Its
location along the activity axis should also reflect the developer’s promotional claims and advertised use
cases rather than its terms of service alone, since terms of service are often written to limit liability and
may describe a narrower product than the one being marketed. This is consistent with the discussion
paper, which indicates that directiveness should be judged on substance (e.g., a disclaimer to consult a
clinician does not make an action-directing output less directive). For agentic devices that carry out
several steps or act through a committee of AI agents, redundancy among agents should not be credited
as a safeguard without evidence from the specific configuration under review. In preliminary
experiments by MIT Critical Data across six public datasets, an agent that ignored a misleading shortcut
when working alone sometimes adopted it once two confident peer agents asserted the same wrong
answer, and a peer offering a detailed justification was not more persuasive than two peers offering
none. [8] These findings come from a single study of one model family in simulated settings, but they
illustrate why agreement among agents should be verified rather than assumed to be independent.

Re: questions 7, 8, 10, 11, 14, and 15

The FDA should require clinical evidence that reflects how the device will be used. Performance on
benchmarks or synthetic test cases can help establish a device's capabilities and limitations, but should
not alone be sufficient to demonstrate how the device will perform in clinical practice. [4, 9, 10] The
FDA should require manufacturers to demonstrate that their evaluations reflect the patients, clinical
decisions, and care settings in which the device is intended to be used. [11, 12, 13] For devices that can
initiate treatment or in circumstances in which errors could cause serious harm, this should ordinarily
include a prospective clinical study with independent safety oversight, with the study design
proportionate to the risk and to what is feasible in the intended setting. [14, 15] The comparator in such
studies should be the care the intended patients would otherwise receive, and ‘no care’ is an appropriate
comparator only when that is an acceptable alternative for the study population. A favorable average
comparison does not by itself justify accepting a risk of severe harm to a subset of patients, and the
reference standard against which the device is judged should itself be examined for the biases it may
carry. [16, 17]

Re: questions 3, 4, 6, and 9

The FDA should require that each device be tested among the clinicians and patients who will actually
use it. For patient-facing GenAI-enabled devices, real-world patients and caregivers across patient
populations should evaluate whether the device’s outputs are understandable at their level of health
literacy and feasible to act on, and qualified clinicians with the requisite specialty training should
evaluate whether those outputs are medically appropriate. [18, 19] For example, when the device advises
a patient about whether and how urgently to seek care from a human clinician, testing should follow
each patient afterward and record whether the patient understood how urgent the condition was, whether
the patient sought human care, whether the patient received care, and how long each step took. This
PORTAL • Program On Regulation, Therapeutics, And Law MIT Critical Data

would extend the discussion paper’s benchmark elements for escalation, communication, and
comprehension from simulated encounters to real patients followed after the encounter. Cases in which
the device failed to recognize an emergency and cases in which it sent a patient to care unnecessarily
should be counted separately, and a patient who is lost to follow-up should be recorded as an unknown
outcome. [20] Whenever possible, results should be stratified across the intended-use population, so that
favorable overall statistics do not inadvertently hide unacceptable risks for some patients. [16, 17, 21,
22, 23] When the evidence base for a given group of patients is thin in the testing of a given device, the
FDA should require further study, interim precautions, or labeling that states the limits of the evidence,
rather than accept an assumption of equal performance.

Re: FDA questions 10 and 16

The discussion paper already expects adjudicators to be independent of the sponsor and model
developer. We recommend that FDA go further to build independent evaluation into the review of
evidence used to authorize GenAI-enabled devices. Manufacturers should continue to be responsible for
conducting their own studies, but for higher-risk devices the FDA may need to seek independent sources
of evidence to determine whether those devices are safe. [12] The FDA should identify independent
experts who have no conflicting interests in the outcome of the evaluation and empower them to oversee
the selection of safety-critical cases, evaluate the clinical results, and audit the underlying data. [24, 25,
26, 27, 28] These experts should have access to all relevant study results, including unsuccessful tests,
and be able to report safety concerns directly to the FDA without needing the manufacturer’s approval.
Trade secret and other intellectual property claims should not prevent qualified third-party reviewers
from examining the data and model attributes needed to assess safety; the FDA’s existing confidentiality
protections should be extended to such reviewers so that proprietary information is protected without
being placed beyond scrutiny.

Re: FDA questions 18, 19, 20, and 21

The FDA should not allow sponsors to replace required safety testing before market authorization with
promises to study a GenAI-enabled device only after the device is introduced into patient care. [5, 29]
Postmarket monitoring of GenAI-enabled devices is necessary and should be designed to complement
the evidence needed to establish device safety and effectiveness prior to initial authorization for clinical
use. [30] Manufacturers should be required to develop postmarket monitoring plans for GenAI-enabled
devices before authorization, with regular evaluations of device safety and effectiveness at
predetermined intervals after deployment. [3, 31] Each postmarket monitoring plan should clearly
establish how device-related safety concerns will be identified, who is responsible for investigating and
addressing identified safety concerns, and what actions must be taken if unacceptable risks arise or
required studies are not completed on time.

A supervisory AI agent may assist with postmarket monitoring, but its reliability should be tested for
each intended task as a device function in its own right. MIT Critical Data compared three automated
PORTAL • Program On Regulation, Therapeutics, And Law MIT Critical Data

monitors on the same clinical multi-agent runs and found that how well a monitor worked depended as
much on the task as on the monitor. A judge that read the full deliberation performed well on text
questions but no better than a simple agreement check on chest radiographs, and the one monitor that
performed acceptably across both still raised enough false alarms to belong in a review queue rather than
in front of a clinical decision. [8] These results are from one study and should be confirmed in other
settings, but they suggest that a monitor’s performance cannot be assumed to transfer from one task to
another.

The FDA should therefore require that a supervisory agent’s rate of missed serious errors and its falsealarm rate be reported for each task, and that human review include a sample of cases the agent did not
flag.

Re: FDA questions 22, 23, 24, and 25

Evidence showing that a GenAI-enabled device was safe and effective when first authorized does not
necessarily establish that the device will remain safe and effective after modifications, updates, or being
used in a different patient population. [2, 32] The FDA should require manufacturers to assess how
changes to a GenAI-enabled device and its clinical context could affect the device’s safety and clinical
performance. Replacing the underlying AI model, expanding the scope of functions of the device,
reducing physician oversight in the use of the device, or changing the patient information the device
uses may introduce new risks, even if the user interface appears unchanged. [33, 34, 35] Manufacturers
should conduct additional testing of the modified device based on the potential clinical consequences of
each proposed change and applicable FDA requirements. Manufacturers should also compare the
updated device with both the most recently tested version and the version originally reviewed by the
FDA, so that the combined effects of a series of updates are not overlooked. When manufacturers rely
on AI models developed by other companies, manufacturers should be responsible for identifying and
evaluating updates to the underlying models before incorporating the updates into the clinical device,
and reporting modifications to the FDA when required. The FDA’s proposed voluntary Foundation
Model Device Master File could help manufacturers and regulators access information about the
underlying models, but should not replace evidence demonstrating that the complete GenAI-enabled
device remains safe and effective. [36] Regardless of whether a Master File exists, sponsors should be
expected to secure change notification commitments from model developers as a condition of relying on
those models. A predetermined change control plan should state the kinds of change permitted, their
limits, the testing each requires, and the point at which the FDA must review the device again. Passing
the original test should not by itself authorize a new treatment, a new patient population, or less
supervision, and changes outside the plan should remain subject to the FDA’s standard premarket
submission requirements. [2, 33, 34]
PORTAL • Program On Regulation, Therapeutics, And Law MIT Critical Data

Re: FDA’s scope and legal authority

The FDA’s proposed framework addresses how to evaluate the risks of GenAI-enabled medical devices,
but does not resolve which clinical AI functions fall under federal regulation. Clarifying this is
increasingly important as AI systems move beyond providing general health information to
recommending treatments and prescribing medications. Without clear guidance, developers of GenAI
tools may interpret the ambiguity in the FDA’s statutory mandate and regulatory authority in ways that
avoid costly but important safeguards, and developers acting in good faith are left without a reliable
basis for planning. States may temporarily relax professional-practice requirements through regulatory
mitigation agreements, but state authorization should not be treated as evidence that an AI system is safe
and effective for autonomous clinical use or as a substitute for any federal review required under
medical-device law. [6, 7] State regulatory relief also cannot exempt manufacturers from applicable
federal requirements or shield manufacturers from federal enforcement. [12]

The FDA should explain when a clinical-AI function falls within its device jurisdiction before applying
the paper’s risk framework. PORTAL examined related questions in its work on mobile health
regulation. [37] The statutory software exclusions and the FDA’s clinical decision support guidance
remain central to that analysis. [38, 39] However, the non-device clinical decision support criteria apply
only to software intended for health care professionals, with patient-facing decision support, health
chatbots, and other tools still needing regulatory clarification. We recommend that the FDA publish its
assessment of examples covering patient-specific treatment advice, prescription renewal, and initiation
of clinical orders. The examples should distinguish a function outside device jurisdiction from a device
within jurisdiction for which the FDA does not require a particular submission or has chosen not to
enforce certain requirements.

Our work on Utah’s clinical-AI sandbox, which has supported useful experimentation, also highlights
the importance of independent oversight when federal, state, and professional responsibilities intersect.
[12] The FDA should coordinate with state authorities and explain any applicable federal review
requirements. A state pilot’s permission to progress toward autonomous care should not be mistaken for
evidence that autonomous use has been evaluated.

How a product is marketed also bears on whether it is a device. Under 21 CFR § 801.4, intended use can
be evidenced by labeling, advertising, design, and the circumstances of distribution. [40] A
manufacturer promoting autonomous treatment should either substantiate that claim or correct it;
contradictory terms of service should not be treated as resolving the question. The FDA’s multiplefunction policy also provides a way to assess how supporting software affects a device without assuming
that every component separately becomes a device. [41]

The FDA should use its existing powers when they apply. For example, medical device reporting and
section 522 surveillance are different mechanisms, and section 522 applies to specified Class II and III
devices meeting statutory criteria rather than to all health AI. [31, 42] A general system of expiring
PORTAL • Program On Regulation, Therapeutics, And Law MIT Critical Data

authorizations, compulsory evaluation fees, or direct duties on suppliers that do not market devices, such
as the measures we and others have proposed elsewhere, would need separate authority analysis. So do
proposals for broader institutional reporting, protected reporting channels, public monitoring results, and
universal periodic review of accumulated changes. Some may require legislation.

CDRH’s next step should be to publish practical examples of the evidence expected, develop
independent testing arrangements, and specify how safety findings will change continued use. The
agency should distinguish steps available now from those requiring a new program, rulemaking, or
legislation. Our groups would welcome the opportunity to contribute to that work, with patient partners
and frontline clinicians involved from the outset.

Conclusion

GenAI-enabled medical devices may help expand patient access to care and improve outcomes, and the
same capabilities that make them useful also create risk. The FDA’s approach to regulating these tools
should account for the different clinical roles they can play, whether providing information, directing
action, or taking action with or without clinician supervision, and the level of scrutiny they receive
should be commensurate with the risks associated with their claimed capabilities. The core of the
FDA’s strategy should be clinical evidence, generated before authorization in the settings where the
device will be used, that establishes safety and effectiveness. Longitudinal, real-world evidence of safety
and effectiveness, validated with the patient and clinician communities affected by each tool, should
then confirm that pre-authorization performance. There should be initial benchmarks that establish a
baseline level of safety before any device is used on patients, as well as ongoing requirements to ensure
the device’s performance does not degrade as models, clinical use cases, or patient populations change.
The FDA should use all of the regulatory levers at its disposal, and clearly articulate to GenAI
developers what AI clinical activities fall within its jurisdiction. GenAI-enabled devices may become
part of how the health care system addresses access and sustainability, but that will depend on whether
they are shown to work for the patients they are meant to serve. By requiring clear, consistent testing
grounded in specific clinical contexts across the life cycle of these tools, the FDA can help ensure that
the benefits of these technologies are realized without exposing patients to avoidable harm.

Respectfully submitted,

Nikhil Jaiswal, BSc
Rahul Gorijavolu, MSE
Shailin Thomas, MD, JD
Thomas J. Hwang, MD
Leo Anthony Celi, MD, MS, MPH
Aaron S. Kesselheim, MD, JD, MPH
PORTAL • Program On Regulation, Therapeutics, And Law MIT Critical Data

References
1. U.S. Food and Drug Administration. Considerations for the Regulation of Generative AI-Enabled Medical Devices:
Discussion Paper and Request for Feedback. August 2026. Docket FDA-2026-N-7874.
https://www.fda.gov/media/194242/download
2. Hwang TJ, Kesselheim AS, Vokinger KN. Lifecycle regulation of artificial intelligence- and machine learning-based
software devices in medicine. JAMA. 2019;322(23):2285–2286. doi:10.1001/jama.2019.16842
3. Kramer DB, Tan YT, Sato C, Kesselheim AS. Postmarket surveillance of medical devices: a comparison of strategies in
the US, EU, Japan, and China. PLOS Med. 2013;10(9):e1001519. doi:10.1371/journal.pmed.1001519
4. Bielick CG, Awwad A, Ellen J, et al. Moving beyond the benchmarks: five foundational principles for meaningful AI
evaluation in healthcare. PLOS Digit Health. 2026;5(5):e0001115. doi:10.1371/journal.pdig.0001115
5. Abulibdeh R, Cajas Ordóñez SA, Celi LA, et al. 1,357 AI medical devices cleared, 3 actually tested on patient outcomes.
PLOS Digit Health. 2026;5(8):e0001597. doi:10.1371/journal.pdig.0001597
6. Utah Department of Commerce, Office of Artificial Intelligence Policy. Doctronic AI Regulatory Mitigation Agreement.
https://commerce.utah.gov/ai/regulatory-relief-4/authorized-pilots/doctronic/
7. Jaiswal N, Gorijavolu R, Hernandez-Boussard T, Celi LA. When an algorithm renews a prescription. JAMA Intern Med.
Published online August 31, 2026. doi:10.1001/jamainternmed.2026.3480
8. Cajas Ordóñez SA, Munnangi A, Marzullo A, et al. Agents catching agents: shortcut cascades and benchmark gaming in
clinical multi-agent systems. arXiv. 2026. doi:10.48550/arXiv.2608.03744
9. Mbakwe AB, Lourentzou I, Celi LA, Mechanic OJ, Dagan A. ChatGPT passing USMLE shines a spotlight on the flaws of
medical education. PLOS Digit Health. 2023;2(2):e0000205. doi:10.1371/journal.pdig.0000205
10. Bean AM, Payne RE, Parsons G, et al. Reliability of LLMs as medical assistants for the general public: a randomized
preregistered study. Nat Med. 2026;32(2):609–615. doi:10.1038/s41591-025-04074-y
11. Celi LA, Jaiswal N, Gorijavolu R, Hernandez-Boussard T. Evidence, transparency, and scope: concerns about Doctronic.
BMJ. 2026;393:s833. doi:10.1136/bmj.s833
12. Gorijavolu R, Jaiswal N, Kesselheim AS, Sakran JV, Hernandez-Boussard T, Celi LA. What Utah’s clinical AI sandbox
reveals about independent oversight. Nat Med. 2026;32(7):2330–2332. doi:10.1038/s41591-026-04418-2
13. Basu S, Patel SY, Sheth P, et al. AI safety evaluation in an underrepresented population: real-world performance of
clinical decision support and frontier language models on Medicaid patient messaging triage. BMC Med Inform Decis Mak.
2026. doi:10.1186/s12911-026-03763-z
14. Cruz Rivera S, Liu X, Chan AW, Denniston AK, Calvert MJ; SPIRIT-AI and CONSORT-AI Working Group. Guidelines
for clinical trial protocols for interventions involving artificial intelligence: the SPIRIT-AI extension. Nat Med.
2020;26(9):1351–1363. doi:10.1038/s41591-020-1037-7
15. Liu X, Cruz Rivera S, Moher D, Calvert MJ, Denniston AK; SPIRIT-AI and CONSORT-AI Working Group. Reporting
guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension. Nat Med.
2020;26(9):1364–1374. doi:10.1038/s41591-020-1034-x
16. Wong AI, Charpignon M, Kim H, et al. Analysis of discrepancies between pulse oximetry and arterial oxygen saturation
measurements by race and ethnicity and association with organ dysfunction and mortality. JAMA Netw Open.
2021;4(11):e2131674. doi:10.1001/jamanetworkopen.2021.31674
17. Zack T, Lehman E, Suzgun M, et al. Assessing the potential of GPT-4 to perpetuate racial and gender biases in health
care: a model evaluation study. Lancet Digit Health. 2024;6(1):e12–e22. doi:10.1016/S2589-7500(23)00225-X
18. U.S. Food and Drug Administration. Applying Human Factors and Usability Engineering to Medical Devices. Final
guidance. August 2026. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/applying-humanfactors-and-usability-engineering-medical-devices
19. U.S. Food and Drug Administration. Incorporating Voluntary Patient Preference Information over the Total Product Life
Cycle. Final guidance. March 2026. https://www.fda.gov/regulatory-information/search-fda-guidancedocuments/incorporating-voluntary-patient-preference-information-over-total-product-life-cycle
PORTAL • Program On Regulation, Therapeutics, And Law MIT Critical Data

20. Ramaswamy A, Tyagi A, Hugo H, et al. ChatGPT Health performance in a structured test of triage recommendations. Nat
Med. 2026;32(5):1671–1675. doi:10.1038/s41591-026-04297-7
21. Vokinger KN, Feuerriegel S, Kesselheim AS. Mitigating bias in machine learning for medicine. Commun Med.
2021;1:25. doi:10.1038/s43856-021-00028-w
22. Celi LA, Cellini J, Charpignon ML, et al. Sources of bias in artificial intelligence that perpetuate healthcare disparities: a
global review. PLOS Digit Health. 2022;1(3):e0000022. doi:10.1371/journal.pdig.0000022
23. Windecker D, Baj G, Shiri I, et al. Generalizability of FDA-approved AI-enabled medical devices for clinical use. JAMA
Netw Open. 2025;8(4):e258052. doi:10.1001/jamanetworkopen.2025.8052
24. Gallifant J, Afshar M, Ameen S, et al. The TRIPOD-LLM reporting guideline for studies using large language models.
Nat Med. 2025;31:60–69. doi:10.1038/s41591-024-03425-5
25. Raji ID, Xu P, Honigsberg C, Ho D. Outsider oversight: designing a third party audit ecosystem for AI governance. In:
Proceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society. 2022:557–571. doi:10.1145/3514094.3534181
26. U.S. Food and Drug Administration. Accreditation Scheme for Conformity Assessment (ASCA).
https://www.fda.gov/medical-devices/division-standards-and-conformity-assessment/accreditation-scheme-conformityassessment-asca
27. U.S. Food and Drug Administration. Medical Device Development Tools (MDDT). https://www.fda.gov/medicaldevices/medical-device-development-tools-mddt
28. Hwang TJ, Bucha S, Kesselheim AS. Comments on reauthorization of the Medical Device User Fee Amendments, FY
2028–2032. Program On Regulation, Therapeutics, And Law; August 2026. https://portalresearch.org/comments-onreauthorization-of-the-medical-device-user-fee-amendments-fy-2028-2032/
29. Deshmukh AD, Kesselheim AS, Rome BN. Timing of confirmatory trials for drugs granted accelerated approval based
on surrogate measures from 2012 to 2021. JAMA Health Forum. 2023;4(3):e230217.
doi:10.1001/jamahealthforum.2023.0217
30. Subasri V, Krishnan A, Kore A, et al. Detecting and remediating harmful data shifts for the responsible deployment of
clinical AI models. JAMA Netw Open. 2025;8(6):e2513685. doi:10.1001/jamanetworkopen.2025.13685
31. U.S. Food and Drug Administration. Medical Device Reporting (MDR): How to Report Medical Device Problems. See
also 21 CFR part 803. https://www.fda.gov/medical-devices/medical-device-safety/medical-device-reporting-mdr-howreport-medical-device-problems
32. Rome BN, Kramer DB, Kesselheim AS. FDA approval of cardiac implantable electronic devices via original and
supplement premarket approval pathways, 1979–2012. JAMA. 2014;311(4):385–391. doi:10.1001/jama.2013.284986
33. U.S. Food and Drug Administration. Marketing Submission Recommendations for a Predetermined Change Control Plan
for Artificial Intelligence-Enabled Device Software Functions. Final guidance. August 2025. https://www.fda.gov/regulatoryinformation/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-planartificial-intelligence
34. Federal Food, Drug, and Cosmetic Act § 515C, 21 USC § 360e-4. Predetermined change control plans for devices.
https://www.law.cornell.edu/uscode/text/21/360e-4
35. U.S. Food and Drug Administration. Deciding When to Submit a 510(k) for a Software Change to an Existing Device.
Final guidance. October 2017. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/deciding-whensubmit-510k-software-change-existing-device
36. Gallifant J, Fiske A, Levites Strekalova YA, et al. Peer review of GPT-4 technical report and systems card. PLOS Digit
Health. 2024;3(1):e0000417. doi:10.1371/journal.pdig.0000417
37. Cortez NG, Cohen IG, Kesselheim AS. FDA regulation of mobile health technologies. N Engl J Med. 2014;371(4):372–
379. doi:10.1056/NEJMhle1403384
38. Federal Food, Drug, and Cosmetic Act § 520(o), 21 USC § 360j(o). Software functions.
https://www.law.cornell.edu/uscode/text/21/360j
39. U.S. Food and Drug Administration. Clinical Decision Support Software. Final guidance. January 2026.
https://www.fda.gov/regulatory-information/search-fda-guidance-documents/clinical-decision-support-software
PORTAL • Program On Regulation, Therapeutics, And Law MIT Critical Data

40. 21 CFR § 801.4. Meaning of intended uses. https://www.ecfr.gov/current/title-21/chapter-I/subchapter-H/part801/subpart-A/section-801.4
41. U.S. Food and Drug Administration. Multiple Function Device Products: Policy and Considerations. Final guidance. July
2020. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/multiple-function-device-products-policyand-considerations
42. Federal Food, Drug, and Cosmetic Act § 522, 21 USC § 360l. Postmarket surveillance.
https://www.law.cornell.edu/uscode/text/21/360l