Erwin Sole

I am a bachelor’s-prepared registered nurse with five years of clinical experience across multiple specialties, currently practicing at the bedside. I also maintain a published registry of clinical retrieval-augmented generation failure modes.

My full comment is attached. It addresses Questions 1, 18, 19, 20, and 21, and concerns risk framing and postmarket monitoring. A summary follows.

The discipline the paper asks for already exists in nursing. Much of clinical practice is longitudinal: clinicians track and trend findings over time specifically in order to act on a developing problem before it becomes an event. The obstacle to clinician participation in postmarket monitoring is not willingness or aptitude. It is that the data pipeline can flatten the trend, while the mandatory reporting framework is organized primarily around discrete adverse events and qualifying malfunctions.

QUESTION 1. I support including downstream safeguards and time pressure among the risk dimensions, and I specify how each operates at the bedside. The safeguard against a wrong directive output is a clinician with time to notice it, and that capacity varies by unit, shift, and census. A sensitivity producing an actionable alert volume under adequate staffing can produce unactionable volume under inadequate staffing, which turns the safeguard into a rubber stamp. Time pressure belongs on the risk side: during a rapid response or a code, the human reviewer the framework relies on is delivering care, not reviewing.

QUESTION 18. Before accepting more premarket uncertainty in exchange for postmarket monitoring, I ask that two things be stated: the specific clinical problem justifying the trade, and the expected detection lag. Premarket evidence prevents harm before it occurs; monitoring limits harm after it starts. The patients affected inside that interval are what the trade purchases, and the interval should be estimated rather than left implicit.

QUESTION 19. Structured-data monitoring becomes less reliable as clinical acuity rises, precisely when the consequences of missed degradation may be greatest, because high-acuity events generate the least contemporaneous structured documentation. Clinical meaning frequently lives in narrative and does not reliably survive extraction: a documented denial of chest pain in a patient admitted for unstable angina can extract to the same value as the same words from a healthy outpatient, losing the fact that an anticipated emergency was being actively watched for. I support sample-based clinician review, with conditions: reviewers need access to the narrative record, clinical competence in the setting under review, and a working understanding of how the system can fail. That second competency cannot be assumed. Reassessment triggers should include changes at the deploying institution and to the external standard of care, not only manufacturer-initiated changes, with notification running in both directions and a named individual on each side.

QUESTION 20. A machine supervisory agent reading clinical documentation inherits the failure modes of the system it monitors, including retrieval over a corpus that never received the document and semantic loss at the extraction boundary. In medicine the absence of a finding is itself information, interpreted differently depending on whether anyone looked for it. A system returning the same confident output whether something was ruled out, never assessed, or never reached the index produces results that read as reassuring regardless. Evaluation of such an agent must test for what it fails to see, not only the accuracy of what it reports.

QUESTION 21. The party obligated to monitor and the party holding the signal are often not the same. Much of the evidence of degradation sits in health system records the manufacturer may not access, which is why clinicians need a defined role. A clinician is accountable for decisions they make, and the system recommending it protects no license. Detecting population-level degradation is a different duty, concerning patients they never cared for, over a period they do not observe, using evidence they do not aggregate. Assigning it informally relocates accountability without conferring authority, time, or data access. The existing mandatory reporting triggers are organized around discrete events that have already occurred, and a device whose output is gradually declining fits them poorly. I ask that any clinician role include a defined review responsibility rather than an expectation, protected time, stated reviewer competency requirements, a confidential channel for reporting suspected degradation before a reportable adverse event, and an explicit statement that the manufacturer’s obligation is not discharged by clinician participation.

Submitted in my individual capacity as a practicing registered nurse. The views expressed are my own.

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

Submitted by Sole Erwin, BSN, RN, in an individual capacity.

I am a bachelor’s-prepared registered nurse with five years of clinical experience across multiple
specialties, currently practicing at the bedside. I also maintain a public dataset of CMS-0057-F payer
disclosure infrastructure and a published registry of clinical retrieval-augmented generation failure
modes.

My comments address five questions, in the order they appear in the paper. They concern risk framing
and postmarket monitoring, where I believe the framework may assume a quality of clinical signal that
is often unavailable in practice.

I want to note at the outset that the discipline the paper is asking for already exists in nursing. Much of
clinical practice is longitudinal: clinicians track and trend findings over time specifically in order to act
on a developing problem before it becomes an event. The obstacle to clinician participation in
postmarket monitoring is not willingness or aptitude. It is that the data pipeline can flatten the trend,
while the mandatory reporting framework is organized primarily around discrete adverse events and
qualifying malfunctions.

Question 1. Dimensions of the risk framework
The two axes are reasonable, and I support the paper’s inclusion of downstream safeguards and time
pressure among the dimensions under consideration. I would like to specify how each operates at the
bedside.

Downstream safeguards depend on staffing capacity. The safeguard against a wrong directive
output is a clinician with enough time to notice it. That capacity is not a property of the device. It varies
by unit, by shift, and by census. The same tool at the same level of directiveness can carry materially
different risk on a fully staffed unit than on one running two nurses short.

This has a direct consequence for threshold setting. A sensitivity that produces an actionable number of
alerts under adequate staffing can produce unactionable volume under inadequate staffing, and
sustained alert volume turns the human safeguard into a rubber stamp. A risk assessment that treats
clinician attention as a constant will tend to overestimate the safeguard in precisely the conditions
where it is weakest.

Time pressure belongs on the risk side rather than the safeguard side. The framework credits a
human reviewer with catching a wrong output, but during a rapid response or a code that reviewer is
not available to review. They are not inattentive. They are delivering care. A device used in time-critical
settings warrants a higher risk rating for that reason, because the safeguard the framework relies on is
least available at the moment it matters most.

Question 18. Greater reliance on postmarket monitoring
Before accepting more premarket uncertainty in exchange for postmarket monitoring, I would ask that
two things be stated explicitly.

First, the specific clinical problem that justifies the trade. Speed of technological development is not,
by itself, a reason to accept more uncertainty in a patient-facing system. If an unmet need is urgent
enough to justify less premarket evidence, it should be possible to name that need.

Second, the expected detection lag, and what happens inside it. Premarket evidence is intended to
prevent harm before it occurs. Monitoring can only limit harm after it has started. The real cost of the
trade is the patients affected between the point where performance begins to fall off and the point
where someone notices. That interval is the thing being purchased, and I would ask that it be estimated
rather than left implicit.

Question 19. Monitoring approaches, cadence, and triggers
Structured-data monitoring becomes less reliable as clinical acuity rises,
precisely when the consequences of missed degradation may be greatest.

Monitoring systems watch discrete fields, because discrete fields can be counted and compared over
time. Level of consciousness selected from a list, pain scored zero to ten, flowsheet rows, coded
diagnoses, vital signs.

High-acuity events tend to generate the least contemporaneous structured documentation, because
during them the clinician is delivering care rather than charting. Documentation is frequently
reconstructed afterward, compressed, sometimes hours later, and sometimes by a person who was not
the primary caregiver. The events in which a failing device matters most are often the events that
produce the weakest and most delayed signal.

Clinical meaning frequently lives in narrative, and narrative does not reliably
survive extraction.

Angina is chest pain caused by reduced blood flow to the heart. A patient admitted for unstable angina
is admitted because that pain is occurring unpredictably and may signal an impending heart attack. In
such a patient, chest pain is anticipated rather than merely possible, and new or worsening pain would
prompt immediate escalation under the applicable clinical protocol.

Consider that patient having a period of tachycardia during treatment without reporting any pain. The
clinician assesses and documents that the patient denies chest pain.

A natural language processing system extracting that entry will typically record a negated finding for
chest pain. The extracted value can be indistinguishable from the one produced by a healthy outpatient
answering a routine review-of-systems question.

The same four words carry different clinical weight depending on who they were written about. For a
patient with no cardiac history, they record an unremarkable answer to a screening question. For this
patient, they record that a specifically anticipated emergency did not occur at the moment of that
assessment, during a period when the clinician had reason to watch for it. The extracted value can be
identical in both cases.

A monitoring program reading extracted values may be unable to distinguish between a finding that is
reassuring and a finding that is reassuring only because someone was watching for the opposite.

Sample-based clinician review is the right instrument, and it requires
specification.
I support sample-based clinician review, with conditions, because an unspecified version of it is unlikely
to work.

A review conducted against structured data reproduces the blind spot it is meant to correct. The
reviewer needs access to the narrative record.

The reviewer also needs two distinct competencies, and clinical licensure supplies only one of them.
The first is clinical competence in the care setting under review, including current familiarity with its
workflow. The second is a working understanding of how the system under review can fail: what it does
with incomplete input, what it does when a document never reaches it, and what it does when meaning
is lost in transformation. A clinician who is unfamiliar with these failure modes is unlikely to recognize
one.

There is reason to think this second competency cannot be assumed. A survey of 2,041 healthcare
leaders across 90 countries, conducted by the Digital Medicine Society with Google for Health, found
that 68 percent did not feel very confident using or evaluating AI tools, and reported that critically
evaluating AI systems was the lowest-confidence area across nearly every role. A review requirement
that assumes this competency may produce sign-off rather than scrutiny.

Triggers should include local and external events, not only manufacturerinitiated changes.
Predetermined change control plans address changes the manufacturer makes. But the meaning of a
device’s inputs can change without the device changing at all. A revised local protocol. A modified EHR
build. A retired flowsheet row. A new service line. A changed staffing model. An updated clinical
practice guideline from a specialty society, which occurs continually and which changes both what
clinicians do and how they document it.

I was not able to identify a standardized notification pathway that consistently connects changes at the
deploying institution with manufacturer reassessment, or manufacturer-initiated changes with
notification to affected clinical users.

The second direction deserves particular attention. When a manufacturer updates an underlying
foundation model, the clinicians using the device may be the only party positioned to notice that its
behavior has changed, and they are frequently not informed that anything has.

I would suggest that reassessment triggers include changes at the deploying institution and changes to
the external standard of care, with a defined period of heightened monitoring following each, and that
notification run in both directions. I would further suggest naming a specific individual on each side,
one at the deploying institution and one at the manufacturer, with a standing channel between them, so
that there is a person to notify rather than an organization.

Question 20. Machine-based supervisory agents
A machine supervisory agent reading clinical documentation inherits the failure modes of the system it
is monitoring. I maintain a public registry of these failure modes at https://github.com/solehealthtech/clinical-rag-failure-mode-registry

Two are directly relevant.

The first is retrieval over a corpus that never received the document. If a note, a result, or a form never
entered the system’s index, the system returns an answer built from everything else, and that answer
can look complete. Nothing in the output indicates a gap. The clinician reading it may be unable to
distinguish between a record in which nothing relevant exists and a record containing something the
system never saw.

That distinction is the entire clinical question. In medicine, the absence of a finding is itself
information, and it is interpreted differently depending on whether anyone looked for it. A documented
absence means something was assessed and ruled out. A silent absence may mean nothing was
assessed, or that the assessment exists somewhere the system cannot reach. A system that returns the
same confident output in all three situations produces results that read as reassuring regardless of
which one is true.

The second is semantic loss at the extraction boundary, described above.

The defining property of both is that the absence is difficult to detect from inside the pipeline. A
supervisory agent cannot report that something is missing when the missing thing never arrived.
Evaluation of such an agent would therefore need to test specifically for what it fails to see, rather than
only for the accuracy of what it reports.

Question 21. Roles for clinicians, institutions, and
professional societies
The party obligated to monitor and the party holding the signal are often not the same party. Much of
the clinical evidence needed to identify degradation in a deployed device resides in health system
records that the manufacturer may not have access to.
That is why clinicians need a defined role. Not as a courtesy, and not as a gesture toward clinical input,
but because they may hold the only available record of the relevant evidence.

I would ask that the role be defined precisely, for a reason that matters to the people who would carry
it.

A licensed clinician is accountable for the clinical decisions they make. The statement that the system
recommended it does not protect a license, and it should not.

Detecting population-level degradation in a deployed device is a different duty. It concerns patients the
individual clinician may never have cared for, over a period they do not observe, using evidence they do
not aggregate. Assigning that duty informally relocates accountability without conferring the authority,
the time, or the data access that would make it achievable.

On the reporting path

The existing mandatory mechanism is not well suited to this problem.

Under medical device reporting requirements, manufacturers, importers, and device user facilities
have mandatory reporting obligations, with different requirements depending on the reporter and the
event. Manufacturers must report deaths and serious injuries their device may have caused or
contributed to, and malfunctions likely to cause or contribute to a death or serious injury if they recur.
Device user facilities must report suspected device-related deaths and serious injuries, but are not
required to report malfunctions under the user-facility reporting provision. An individual clinician’s
route is voluntary submission of Form FDA 3500 through MedWatch.

These mandatory triggers are organized around discrete events that have already occurred. A
generative device whose output quality is gradually declining, with no identifiable serious injury yet
attributable to it, does not fit them well. By the time it does, the detection lag discussed in Question 18
has already run.

This is the point at which the framework and clinical practice diverge most sharply. Clinicians are
trained to act on a trend before it becomes an event. The mandatory reporting structure is built to
receive events.

I would also note a practical problem with relying on voluntary individual reporting. Clinicians who
raise concerns about a tool their institution has purchased are not always positioned to do so freely.
Compliance functions in healthcare generally provide a confidential channel for exactly this reason. I
was not able to identify an equivalent confidential route for reporting that a clinical AI tool appears to
be performing worse than it did.

If clinicians are to participate in postmarket monitoring, I would ask that the mechanism include a
defined review role rather than an expectation, protected time, stated competency requirements for
reviewers, a confidential reporting channel for suspected degradation that has not yet produced a
reportable adverse event, and an explicit statement that the manufacturer’s monitoring obligation is
not discharged by clinician participation.

Submitted in my individual capacity as a practicing registered nurse. The views expressed are my own.