Srividya Narayanan

See attached file(s)

Docket:FDA-2026-N-7874
Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion
Paper and Request for Feedback

It is my opinion that he is a man who manages his own business masterfully, with knowledge of
AI/ML enabled medical devices, wearable innovations, Software as Medicine Devices (SaMD),
and medical device administration plans.

I applaud the FDA Virtual Vitality Center of Excellence for requesting comments before the initial
regulatory expectations for Generative Machine Intelligence (GenAI)-enabled healthcare devices.
The discussion document correctly acknowledges that GenAI, using a traditional AI/ML device
paradigm including a variety of freely-form outputs, hallucinations, prompt sensitivity, automation
partiality, model version uncertainty, and uncertainty about the origin and control of the training
data, can introduce a risk profile without absolute capture. FDA underlined that the present paper
is intended to be debated only and does not constitute a draft of any definitive guidance; it is
important to maintain the current unsealed and iterative technique.

The intended two-axis technique should be explicitly narrative for a longer period of time
compared to the clinical disadvantages of the incorrect final product. Furthermore, the extent to
which the model autonomy is restricted or unfettered, the ability of a trained consumer to identify
and correct errors before an impairment occurs, the reversibility of downstream movements, the
vulnerability of a tolerant population, and the clinical and information background of the device
should be considered. A system that produces a structured, clinically reviewed draft report provides
a unique risk profile from an agentic framework that initiates activities or provides a patient-facing
recommendation.
If the Food and Drug Administration will translate the term 'capability ‘into a particular, taskspecific evidentiary description, the Competence-based Methodology is likely to remain a precious
form rule. The prescriber must give details of the intended clinical enterprise, the user, the setting
of the use, the significant disadvantages, and the decision limits. Benchmarking should take into
account everyday, difficult, out-of-distribution, incomplete data, adversarial, and clinically
ambiguous cases, not only average performance testing. In addition, consistency across clinically
relevant subgroups and surroundings should be evaluated. The clinical validation of the ability of
the intended user to interpret, issue, and appropriately execute the final product on top of the device
under realistic working conditions should begin.

Human component testimony should be important for the premarket assessment of GenAI devices.
The Food and Drug Administration should expect validation of the precautionary measures for
consumers, interaction of uncertainty and restriction, prevention of harm caused by automation
and excessive reliance, escalation mechanism, and serviceability under time pressure. Technically
perfect models can still present an unacceptable threat if the presentation of the final product, the
tone of voice, or the working order causes the clinician or patient to misapply it.

The central model and the agentic arrangements. For devices that integrate the cornerstone model,
including the lineage, version, cause, and regulation for training, adjustment, and retrieval, prompt
and orchestration architecture, safety control, and validation for the final device configuration, the
FDA should establish a traceability expectation. For agentic arrangements, the manufacturer
should lay down enforceable autonomy limits, authorisations, human oversight advice, audited
account records, and safe backup countries. Structures enabling clinical, administrative, or
otherwise patient-affecting events to be triggered should be ensured to increase the conditions for
action mandate, recording, reversal, and incident investigation.

Risk-proportional monitoring is important, but it should not become an option to obtain adequate
premarket proof in the case of high-consequence applications. A potential postmarket monitoring
plan, together with established procedures for data quality control, drift and deterioration detection,
thresholds for probes, corrective measures, and methods for communicating material changes to
users, should be available for higher-risk devices. FDA should explain how the practical
performance monitoring interface, together with the reporting, correction, and deletion
responsibilities, the security procedures, and predetermine modify direct Plans (PCCPs), are
related to the reporting, correction, and deletion responsibilities of healthcare devices, as well as
the correction and deletion responsibilities and the predetermination of modified direct plans
(PCCPs). The current FDA life cycle management activity for AI-enabled devices, including the
importance of post-deployment performance monitoring, provides a useful basis for such an effort.

The specific issue of updating the basic model should be addressed by the FDA. A transformation
of an implicit model, retrieval principle, prompt organization, safety rail, or any other tool usage
capability may alter a device's clinical mannerism without obvious adaptation to its current
indication. The Bounded Modification Permit, Credence Standard, Verification Protocol,
Monitoring Epoch, Rollback Gun Button, and User Notification Method should therefore be laid
down in the PCCPs. Material changes in the behavior of the produced end product, autonomy, or
alternatively clinical reliance should be appropriately examined.

Finally, the FDA should pursue global convergence through the IMDRF and a sustained
engagement with regulators such as the Canadian Ministry of Health and MHRA, integrating
conformity with the exceptional. Machine Learning: Ethics. Harmonization of risk-based
expectations will reduce unnecessary atomization while maintaining high standards of safety,
productivity, equity, visibility, and public confidence.

Thank you for the opportunity to provide feedback.