PulseMed LLC (Sonal Patel)

Pulse Med’s Recommendation:
Thank you for the opportunity to comment. Pulse Med is a physician owned and operated AI company bringing the voices of physician to AI while keeping their IP. Our unifying principle is accountability for what an AI actually does with patients, not what it promises in its instructions. Pulse Med argues that convincing-but-wrong AI output can harm patients, so the FDA should judge these tools on how they perform across real clinical use, match scrutiny to the level of potential harm, and keep validating them after approval as models change and drift rather than certifying a static product at a single point in time.

Generative AI-Enabled Medical Devices: A Risk-Proportionate, Clinically Grounded Framework for
Safety Across the Total Product Life Cycle

Submitted by Pulse Med and the undersigned physicians
Docket: FDA-2026-N-7874
Re:Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and
Request for Feedback
Date: October 19, 2026

To the Center for Devices and Radiological Health:

Pulse Med and the undersigned physicians appreciate the opportunity to comment on FDA’s discussion
paper, Considerations for the Regulation of Generative AI-Enabled Medical Devices. We support FDA’s
effort to develop scientifically sound, risk-proportionate approaches that can enable beneficial innovation
while protecting patients and supporting clinicians.

Generative AI can assist with clinical information, communication, analysis, and workflow. But a model’s
fluency is not evidence of clinical reliability. An answer that sounds sure is not always right. A system
may produce plausible but incorrect information, behave differently across equivalent encounters, or
change when its model, prompts, retrieval sources, tools, or deployment environment change. In clinical
practice, these are not merely technical issues: these mistakes can affect a patient’s care and safety.

We therefore urge FDA to build on the discussion paper with a risk-based, total-product-life-cycle
framework that evaluates the complete device in its intended clinical context, accounts for realistic
interactions and foreseeable changes, and makes clear who is responsible for keeping it safe. We offer
these recommendations:

Main Recommendations:

1. Judge the real-world risk, not the disclaimer.
2. Test full conversations, not single answers.
3. Report missed care and excess care separately.
4. Match evidence to risk.
5. Monitor after launch.
6. Track third-party model changes.
7. Require human review before high-stakes actions.
8. Protect independent testing without over-burdening.
Detailed Analysis of Key Recommendations

1. Judge the real-world risk, not the disclaimer.

We support FDA’s proposed use of device activity and the consequences of an incorrect output as
organizing dimensions for risk assessment. We recommend that FDA also explicitly consider:

1. Clinical context and time sensitivity: whether the function is used in emergency, inpatient,
outpatient, home, or other settings, and how quickly a harmful error could affect care.
2. Reversibility and downstream safeguards: whether an action can be readily paused, reviewed,
corrected, or reversed before patient harm occurs.
3. User and patient context: whether the output is intended for a patient, a generalist clinician, a
specialist, or a care team, and whether the intended user can reasonably recognize an error.
4. Degree of direction and personalization: whether the device provides general information,
interprets an individual’s circumstances, recommends a course of action, or issues a specific
instruction.
5. Interaction-level behavior: whether a series of turns can shift the device from informational
support to diagnosis, triage, treatment direction, or another higher-risk function.
6. Autonomy and ability to act: whether the device can initiate or change orders, communicate with
patients, control another system, or otherwise act without a meaningful human checkpoint.

Risk reviews should consider who uses the tool, where and when it is used, how serious a mistake could
be, whether a person can catch the mistake, and whether the action can be stopped or fixed. A note saying
“talk to your doctor” should not lower the risk rating if the AI is still giving strong advice about an
important health decision.

FDA should look at how the AI is really used, not only at its name, label, or warning notice. A tool that
gives general health information may pose less risk than one that tells a patient what treatment to choose
or whether to seek emergency care.

2. Test full conversations, not single answers.

Patients often share information over several messages. A safety problem may appear only as a
conversation continues. For example, the AI may fail to ask an important question, forget a warning sign,
or not change its advice when new information comes up.

For conversational devices, FDA should encourage or expect evaluation across complete, clinically
realistic interaction trajectories, rather than relying principally on isolated prompts and responses. A
single-turn test can miss important failures: the system may not ask a necessary follow-up question, may
lose a safety-critical fact, may become more directive over time, or may fail to escalate after new
information changes the level of concern.
Evaluation should include representative multi-turn scenarios that test, as appropriate:

1. Evolving or ambiguous presentations, including initially reassuring information followed by red
flags.
2. Missing, contradictory, corrected, or low-quality information and changes in patient condition.
3. User pressure, emotional distress, repeated questioning, prompt injection, and attempts to move
beyond the device’s intended scope.
4. The timing, clarity, and appropriateness of escalation, referral, refusal, and human handoff.
5. Whether important facts and limitations are preserved and communicated through the interaction
and any resulting handoff.

Where a device is intended to support a workflow rather than a single response, testing should reflect that
workflow, including relevant interfaces, retrieval systems, tools, and human-AI handoffs.

3. Report missed care and excess care separately.

For tools that guide people about where or when to get care, two different problems matter. The AI might
fail to tell someone to get needed care. Or it might tell too many people to seek urgent care when it is not
needed. For triage and care-escalation functions, both failure to recommend necessary care and
unnecessary escalation can cause harm. We recommend that FDA ask sponsors to prespecify and report
under-escalation and over-escalation separately, with clinically justified thresholds and analyses suited to
the intended use.

Evaluation should examine not only whether escalation occurred, but also whether it was timely,
appropriately specific, understandable to the intended user, and supported by a useful handoff. The
balance between missed escalation and unnecessary escalation will differ by clinical context; a single
aggregate accuracy score may hide an important safety issue. Sponsors should explain how they set
acceptance criteria and how they evaluated effects on patient safety, access, resource use, and trust.

4. Match evidence to risk.

The amount of testing should match the possible harm. The product tested should be the complete tool
that patients or healthcare workers will use, not only the AI model behind it.

For higher-risk tools, computer tests alone are not enough. Companies need to test the product with real
clinical cases, healthcare workers, or patients, as appropriate. FDA should encourage reporting of
important failures and differences in how the tool works for different groups, not just average results.

Benchmarking should use prespecified methods and clinically justified acceptance criteria. As applicable,
it should address safety-critical recognition and escalation; scope adherence and refusal behavior;
calibration and uncertainty communication; clinical knowledge and task performance; information
gathering and reasoning; quantitative interpretation; communication and user comprehension; reliability
and robustness; and performance across clinically relevant populations and settings. Agentic functions
should additionally be evaluated for planning, sequencing, tool use, handling tool failures, and
compliance with human-oversight checkpoints.
Benchmarking should not be treated as sufficient evidence by itself where real-world use, user interaction,
or workflow integration could materially affect performance. Tests should use clear rules and realistic
examples, check important safety skills, such as spotting danger signs, staying within the tool’s purpose,
explaining uncertainty, and working well for different groups of people. For higher-risk functions
particularly those that direct or take consequential actions, operate with limited human review, or could
cause serious harm more rigorous clinical confirmation may be warranted before broad deployment.

If the tool can take several steps or use other computer tools, testing should check those actions too.

FDA should encourage reporting of clinically meaningful failure modes and subgroup performance, not
only average scores. Where synthetic data are used, sponsors should justify its relevance to real-world
populations and settings and explain how they have assessed whether synthetic data could mask
performance gaps. Public benchmark results should be supplemented, where appropriate, by independent
or sequestered test data to reduce contamination and test-set optimization.

5. Monitor after launch.

Premarket testing cannot anticipate every input, workflow, population, or change a generative system may
encounter. For devices whose risks warrant it, FDA should expect a postmarket monitoring plan that
combines scheduled reassessment with event-triggered review. AI products may change, and the people or
settings where they are used may change too. For devices where the risks call for it, companies should
have a plan to keep checking safety after launch. Companies should explain how they will find problems,
decide when a change is unsafe, and take action to protect patients.

Potential triggers should include changes to the underlying model or its version; prompts or system
instructions; retrieval content or retrieval methods; guardrails; orchestration logic; external tools;
interfaces; deployment environment; intended use; and clinically meaningful patterns of near-misses,
complaints, adverse events, or performance degradation. The plan should specify how changes are
detected, how their safety impact is evaluated, what thresholds prompt escalation or corrective action, and
when FDA notification or review may be needed under applicable requirements.

Monitoring can include periodic re-benchmarking against prespecified safety-critical capabilities,
sample-based review of real-world interactions by qualified clinicians, and analysis for drift or changes in
relevant populations and inputs. Review should include complete interaction records where needed to
assess how the system reached or communicated its output. Machine-based monitoring tools may assist,
but should not substitute for a validated monitoring plan or obscure the manufacturer’s responsibility for
the device.

6. Track third-party model changes.

FDA should encourage ways to track which model and version were used, get notice of important updates
when possible, test changes before wider use, and limit or pause a feature if a change may make it unsafe.
Moreover, a device manufacturer remains responsible for the safety and effectiveness of its device, but it
may not control updates to a third-party foundation model or service. FDA should encourage mechanisms
that enable manufacturers to detect, assess, and respond to such changes, including:
 1. Version identification and change documentation sufficient to determine what model and relevant
configuration were used.
2. Advance notice of material updates where feasible, and contractual commitments addressing
notice, support, and access to safety-relevant information.
3. Technical controls, such as version pinning or staged deployment, where feasible and appropriate
to the risk.
4. Prespecified change-control plans that define the types and bounds of changes that can be
managed without a new submission, the tests required for each change category, and triggers for
additional review.
5. A safe response when a material change cannot be characterized in time, including pausing,
reverting, restricting, or otherwise mitigating the affected function when clinically appropriate.

FDA should clarify how these controls can fit within Predetermined Change Control Plans and other
existing pathways. A change should not be presumed safe solely because it was initiated by a third-party
model provider or described as a routine platform update.

7. Require human review before high-stakes actions.

Agentic systems that can plan, call tools, or execute multiple steps can create risks beyond those of
systems that only provide information such as place orders or send instructions. For actions that could
seriously affect a patient or be hard to undo, a qualified person should review and approve the action first,
unless strong evidence supports a safe alternative. FDA should expect a meaningful, appropriately
informed human authorization step unless the sponsor provides compelling evidence and safeguards
supporting a different approach.

The human reviewer should receive enough context to understand the proposed action and its material
basis, including relevant uncertainty, missing information, and tool outputs. Oversight should be designed
so that approval is not merely a perfunctory click. Where possible, there should be a way to stop or undo
an action. Evaluation should test whether the system respects authorization boundaries, stops when
required information is unavailable, detects erroneous or conflicting tool results, and fails safely when a
tool or service is unavailable.

For agentic functions, sponsors should maintain records sufficient to reconstruct the clinically relevant
sequence of inputs, model or system version, plans, tool calls and results, proposed actions, human
approvals, and final actions. These records should be protected against unauthorized alteration and
available for appropriate safety investigation, while respecting privacy and security obligations.
Companies should also keep records of the information the AI used, the steps it took, any outside tools it
used, the human review, and the final action.

8. Protect independent testing without over-burdening.

Independent clinical adjudication and independent test assets can strengthen confidence that a device has
not simply been optimized for its own test set. FDA should consider ways to support qualified,
independent reviewers and sequestered evaluation datasets, with transparent qualifications and
conflict-of-interest safeguards.
At the same time, not every device needs the same kind or amount of testing. Requirements should remain
proportionate to risk and should not make prospective trials or a single evaluation method mandatory for
every device. FDA can support innovation and timely access by allowing multiple scientifically justified
forms of evidence, while requiring stronger evidence and safeguards where the potential consequences of
error, degree of autonomy, or difficulty of independent review are greater. Clear examples and transparent
criteria would help manufacturers, clinicians, and patients understand how these principles apply.

Conclusion

Generative AI-enabled medical devices may offer meaningful clinical benefits, but their safety cannot be
established through model capability claims or isolated benchmark results alone. A confident-sounding
answer does not prove that a device is safe. A sound framework should evaluate the complete device in
realistic use, examine interactions over time, distinguish different kinds of clinical error, establish
proportionate clinical confirmation, and maintain effective monitoring and change control after
deployment.

We ask the FDA to create clear rules that support helpful innovation while protecting patients and
healthcare workers while ensuring that patients and clinicians are not left to bear the burden of
unmeasured performance changes or opaque third-party dependencies. Companies should be able to
explain how their products were tested, how safety will be watched over time, and how problems or
updates will be handled. We appreciate FDA’s consideration of these comments.

Respectfully submitted on behalf of Pulse Med Team:

Sonal Patel, MD,
CEO
Pulse Med
www.pulsemed.ai