Jodi Orm

See attached file for full response.

I submit this comment in response to FDA’s August 2026 discussion paper, Considerations for the Regulation of Generative AI-Enabled Medical Devices. FDA explains that the paper seeks stakeholder feedback and does not represent draft or final guidance. I urge CDRH to include nurses explicitly in risk assessment, premarket evaluation, and post market monitoring for generative AI-enabled medical devices. In many care settings, nurses will be among the clinicians who first encounter an AI output, compare it with the patient’s condition and other clinical information, communicate it across the care team, and decide whether to act, escalate, or seek clarification. Nurses must be prepared and empowered to critically appraise AI outputs through clinical judgment. Device safety cannot depend on a nurse’s vigilance alone.

These comments address Appendix B questions 1, 3, 9, 11, 14, and 19 through 21. They focus on how CDRH can evaluate whether human oversight is meaningful in practice and how nursing expertise can inform that evaluation.

Nursing Participation in FDA Evaluation of
Generative AI Medical Devices
Public comment on the August 2026 discussion paper
Docket No. FDA-2026-N-7874 | Submitted October 19, 2026

To the Center for Devices and Radiological Health:

Introduction
I submit this comment in response to FDA's August 2026 discussion paper, Considerations for the
Regulation of Generative AI-Enabled Medical Devices. FDA explains that the paper seeks stakeholder
feedback and does not represent draft or final guidance [1]. I urge CDRH to include nurses explicitly in
risk assessment, premarket evaluation, and post market monitoring for generative AI-enabled
medical devices. In many care settings, nurses will be among the clinicians who first encounter an AI
output, compare it with the patient's condition and other clinical information, communicate it across
the care team, and decide whether to act, escalate, or seek clarification. Nurses must be prepared and
empowered to critically appraise AI outputs through clinical judgment. Device safety cannot depend
on a nurse's vigilance alone.
These comments address Appendix B questions 1, 3, 9, 11, 14, and 19 through 21. They focus on how
CDRH can evaluate whether human oversight is meaningful in practice and how nursing expertise can
inform that evaluation.

Background and clinical relevance
The discussion paper identifies features that complicate evaluation of generative AI devices: openended and variable outputs, possible changes to models and other components after deployment, and
limited visibility into some third-party foundation models. It proposes a risk framework based on
device activity and the consequences of an incorrect output, and asks about safeguards, output
traceability, user expertise, competency-based testing, human-AI team performance, and clinician
roles in post market monitoring [1]. These are appropriate questions. Their answers should account
for the clinical work expected of nurses when a device is used in care.
Nurses often assess patients over time and reconcile symptoms, patient reports, changes in condition,
medication administration, response to treatment, and the plan of care. That context can reveal when
an AI output is incomplete, unsupported, or inconsistent with the patient’s current presentation.
Nursing participation can therefore help FDA and sponsors identify workflow and human-factors
risks that a benchmark of model outputs alone may miss.
The nursing literature supports this focus. A mixed-methods study of nursing informaticists identified
nurse education and participation in AI implementation planning and evaluation as important to safe
clinical use [2]. In a qualitative study of 23 critical care nurses across four hospitals, participants
described interpreting early-warning outputs through bedside assessment and contextual knowledge;
some reported that workload pressure contributed to undocumented overrides [3]. This study does
not estimate how often such events occur across settings, but it illustrates why override pathways,
workload, and user experience should be assessed as safety issues. NIST likewise emphasizes clear
human-AI roles and evaluation of the interaction between people and AI, because team performance
can differ from the performance of either component alone [4].

Recommendations for risk assessment and premarket evaluation

Assess whether human oversight is meaningful
CDRH should ask sponsors to identify the clinical roles expected to review or act on device outputs
and to describe the conditions that allow those users to assess an output. The assessment should
address whether the user has access to relevant patient context and supporting evidence, enough time
to review the output, authority and a practical mechanism to question or override it, and a clear path
to escalate uncertainty. It should also consider time pressure, reversibility, traceability, and the
consequences of both accepting and rejecting an output.
The presence of a nurse or other clinician in a workflow should not, by itself, be treated as a riskreducing safeguard. A required confirmation click, a disclaimer, or a nominal human-in-the-loop label
does not show that the user can recognize an error or respond safely. Risk assessment should reflect
the intended user’s relevant clinical knowledge and the actual work environment, including whether
the output is action-directing, difficult to verify, time-sensitive, or hard to reverse. If nurses are
expected to use a device, the evidence should represent the roles, training, and clinical settings in
which they will use it.

Evaluate the nurse and device as a clinical team
For devices intended for nurse use or deployed in workflows where nurses interpret or act on
outputs, CDRH should encourage sponsors to include practicing nurses and nursing informaticists in
benchmark design, clinical confirmation, and human-factors evaluation. The participants should
match the intended use and setting. Depending on that use, the evaluation may include registered
nurses, advanced practice nurses, and other relevant clinicians. Nurses should help identify realistic
cases and judge whether outputs are understandable, sufficiently supported, and usable within the
care workflow.
Evaluation should compare device-alone performance with the performance of the intended humanAI team. It should test whether the nurse can detect and respond to materially incorrect, incomplete,
or overconfident outputs under realistic conditions, including multi-turn interactions, interruptions,
time pressure, missing or conflicting information, and clinically relevant subgroups. Measures should
include:
• Clinically significant errors and omissions, including under-escalation and over-escalation.
• Whether users recognize, verify, correct, override, or escalate an unsafe output, and how long
those actions take.
• Whether the interface communicates uncertainty, limitations, and relevant supporting
information clearly enough for review.
• Workload, alert burden, and downstream effects on decisions and care, evaluated in the intended
workflow and across relevant patient groups.
When safety depends on a clinician identifying an error, that capability should be tested as part of the
device’s intended human-AI use. FDA should not infer that a device is safe in practice from standalone
model accuracy or from evidence that users can theoretically override it. Training can support critical
appraisal, but it cannot replace adequate device performance, usable displays, access to evidence, and
workflow conditions that permit review.
Recommendations for post market monitoring and accountability
CDRH should encourage post market plans that include periodic review by independent clinicians,
with nursing reviewers when nurses are intended users or routinely encounter the device in practice.
Monitoring should examine real-world output quality and changes in the human-AI team, including
discrepancies between an output and the patient’s condition, overrides, near misses, escalation
patterns, workload, and signs of performance degradation. Review cadence and triggers should
account for changes to the underlying model, prompts, retrieval or tool-use components, interface,
intended workflow, and the population using the device.
Frontline nurses need a low-burden way to report concerning outputs, near misses, and workflow
barriers. Monitoring should distinguish a device error from a user’s interpretation of that error and
from an organizational condition that prevented review. This information can help identify whether
the problem is recurring and whether a device change affects the user’s ability to detect and manage
risk. CDRH should consider nursing participation in the monitoring roles raised in question 21 while
preserving the manufacturer’s accountability for device monitoring and response [1].
FDA and healthcare organizations should also support the clinical judgment that safe use requires.
Sponsors should clearly identify intended users, device limits, and the evidence supporting outputs,
and should provide role-appropriate training materials. Healthcare organizations should provide
time, training, and reporting processes that allow nurses to question outputs without punitive
barriers. Nurses’ responsibility to exercise professional judgment must not become a substitute for
manufacturer responsibility or for institutional support.

Conclusion
CDRH should make nursing participation explicit in its continuing work on generative AI-enabled
devices. Nurses should contribute to risk assessment, design of clinical benchmarks, confirmation of
human-AI team performance, and post market review. Their ability to critically evaluate and challenge
AI outputs should be treated as a safety capability that must be supported and tested. This approach
would help FDA assess how a device performs in the clinical work where it will be used and would
make the monitoring process more responsive to risks observed at the bedside.
Respectfully submitted,

References
1. U.S. Food and Drug Administration, Center for Devices and Radiological Health. (2026, August). Considerations for
the regulation of generative AI-enabled medical devices: Discussion paper and request for feedback.
https://www.fda.gov/medical-devices/digital-health-center-excellence/considerations-regulation-generative-aienabled-medical-devices-discussion-paper-and-request
2. Turchioe, M. R., Pepingco, C., Lytle, K., & Austin, R. (2024). Nurses’ roles in artificial intelligence implementation:
Results from a mixed-methods study. Studies in Health Technology and Informatics, 315, 223–227.
https://doi.org/10.3233/SHTI240140
3. Samy, R., Ramadan, O. M. E., & Elsayed, G. E. A. (2026). Trusting the algorithm or trusting the nurse? Critical care
nurses’ experiences of automation bias and professional autonomy in AI-assisted early warning. Nursing in
Critical Care, 31(5), e70620. https://doi.org/10.1111/nicc.70620
4. Tabassi, E. (2023). Artificial intelligence risk management framework (AI RMF 1.0) (NIST AI 100-1). National
Institute of Standards and Technology. https://doi.org/10.6028/NIST.AI.100-1