Robert Elvin Domondon

I am a critical care nurse (RN, CCRN, CSC, CMC) who has worked ICU night shift for more than twenty years. I am commenting in my personal capacity, not on behalf of my employer. I also found and advise organizations working on nurse-led AI governance; I disclose that so FDA can weigh my comments accordingly.
I want AI that takes real work off nurses’ hands at 3 a.m. And at that hour, I want to know who answers for it when it is wrong.
My central recommendation is to evaluate the whole care workflow, not only the quality of an AI answer. A system is not supervised unless clinicians have the information, time, authority, and practical means to intervene.

Docket FDA-2026-N-7874 Considerations for the
Regulation of Generative AI-Enabled Medical Devices
To the FDA Center for Devices and Radiological Health:

I am a critical care nurse (RN, CCRN, CSC, CMC) who has
worked ICU night shift for more than twenty years. I am
commenting in my personal capacity, not on behalf of my
employer. I also found and advise organizations working
on nurse-led AI governance; I disclose that so FDA can
weigh my comments accordingly.
I want AI that takes real work off nurses' hands at 3 a.m.
And at that hour, I want to know who answers for it when it
is wrong.
My central recommendation is to evaluate the whole care
workflow, not only the quality of an AI answer. A system is
not supervised unless clinicians have the information,
time, authority, and practical means to intervene.
One night, an automated deterioration score kept firing
low-level alerts, but my bedside assessment told me
something the score could not: the patient was becoming
subtly confused, cool, and harder to arouse, and the
pattern felt wrong before the numbers fully showed it.
Nursing judgment meant integrating those small changes
with the patient’s trajectory, deciding they mattered, and
escalating before the algorithm could be certain. Acting
on that judgment required time to reassess, another nurse
to help, and a workflow that gave the bedside nurse the
authority to escalate early. The lesson: the tool could flag
risk, but nursing judgment recognized what was
happening—and the workflow determined whether that
judgment could become action in time.
1. Risk must reflect bedside conditions | Questions 1, 6
FDA's two-axis framework is a useful starting point. I
recommend it also account for reversibility, time pressure,
downstream safeguards, and traceability. A draft summary
a nurse can check before use is not the same as an agent
that changes a care workflow or starts a highconsequence action. In the ICU, a missed escalation and a
flood of false ones both cause harm: the first delays
rescue, the second teaches staff to stop listening. A
human-approval label should not lower assessed risk
unless realistic testing shows that timely review and
intervention are possible.
2. Test the workflow as deployed | Questions 7, 9, 11
I support benchmarking followed by clinical confirmation
scaled to risk. Test the final device configuration: retrieval,
prompts, tools, permissions, and interface. Include
evolving deterioration, incomplete or conflicting inputs,
handoffs, interruptions, and failures of connected tools.
Measure missed and unnecessary escalation, time to
escalation, review burden, and recovery from error, not
only answer accuracy. Shadow deployment can reveal
workflow failures, but it cannot establish safety once
clinicians start relying on outputs.
3. Include nurses in evaluation | Questions 4, 14, 16, 21
The discussion paper frames clinical users largely as
generalist or specialist physicians. Yet nurses are often
the primary users of surveillance, early-warning, and
escalation tools, and the first to act on them. Where a
device touches nursing assessment, surveillance,
coordination, education, or escalation, independent
evaluation panels should include nurses with relevant
bedside experience, alongside other professionals and
patient representatives. Comparators should match the
actual task and standard of care. Evaluate how the
human-AI team performs, including whether the interface
invites overreliance or hides uncertainty. Nursing expertise
belongs in evaluation design, not only in feedback after
launch.
4. Make human control real | Question 26; Appendix A,
A.1
For agentic devices, test bounded permissions, a named
responsible human, visible action status, oversight
checkpoints, and a stop mechanism that works. Highconsequence or irreversible actions should require explicit
human authorization. Staff should be able to reject an
output, pause execution, and escalate a concern without
hunting through menus. Logs should separate what the
system proposed, what a human accepted, and what was
actually executed. Fallback procedures must work when
the model or a connected service fails.
5. Monitoring cannot replace evidence | Questions 18–
Leaning more on postmarket monitoring may be
reasonable for lower-risk uses where failures are
detectable and recovery is quick. It should not replace
premarket evidence where a failure can seriously harm a
patient before anyone notices. Approval clicks are not
proof of meaningful oversight. Any such trade should
name the remaining uncertainty, how fast problems are
detected, who owns the response, and the thresholds for
restricting or suspending use.
Reporting should be simple for errors, near misses, unsafe
recommendations, and workflow disruptions, and the
reporting clinician should hear back. Reassessment
should follow safety signals and material changes to
models, retrieval, prompts, tools, permissions, or
interfaces, including third-party model updates. Monitor
performance across clinically relevant subgroups and
watch for shifts in escalation behavior. If supervisory AI is
used, validate it independently and test for correlated
failures. Manufacturers must keep accountability;
hospitals and frontline nurses should not inherit an
undefined surveillance burden.

My bedside test
At 3 a.m., with a deteriorating patient and too few hands,
can a nurse understand the system's limits, verify its
output, know who is responsible, and stop an unsafe
action? If not, the device is not supervised, whatever its
label says. I urge FDA to build that test into its approach.
Agents may propose. Humans must judge. Nurses are
ready to steward.

Respectfully,

Robert Domondon, RN, CCRN, CSC, CMC
San Francisco, CA

Reference: FDA, Considerations for the Regulation of
Generative AI-Enabled Medical Devices: Discussion Paper
and Request for Feedback (August 2026), Sections IV–VII
and Appendix A