Alliance for Nursing Informatics (Karen Dunn-Lopez, Nancy Beale)

See attached file(s)

October 19, 2026

U.S. Department of Health and Human Services
Food and Drug Administration
Dockets Management Staff
5630 Fishers Lane
Room 1061 (HFA-305)
Rockville, MD 20852

Re: Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion
Paper and Request for Feedback

Submitted electronically at: Regulations.gov, Docket ID FDA-2026-N-7874

Dear Commissioner Diamantas:

As nursing stakeholders, the Alliance for Nursing Informatics (ANI) is pleased to offer comments
on the Considerations for the Regulation of Generative AI-Enabled Medical Devices:
Discussion Paper and Request for Feedback.

The Alliance for Nursing Informatics (ANI), cosponsored by the American Medical Informatics
Association (AMIA) and the Healthcare Information and Management Systems Society
(HIMSS), advances nursing informatics leadership, practice, education, policy, and research
through a unified voice of nursing informatics organizations. We transform health and
healthcare through nursing informatics and innovation. ANI is a collaboration of organizations
representing more than 25,000 nurse informaticists and bringing together 29 distinct nursing
informatics groups globally. ANI spans academia, practice, industry, and nursing specialty
boundaries and collaborates with more than 4 million nurses in practice today.

ANI appreciates the opportunity to comment on this crucial discussion paper. ANI supports
developing a risk-proportionate, total-product-lifecycle approach to generative AI (GenAI)-
enabled medical devices. The paper’s attention to competency-based evaluation, clinical
confirmation, and postmarket performance is valuable.

However, the proposed two-axis framework may oversimplify risk, particularly when applied to
diverse clinical activities, users, and environments. Drawing on nursing informatics expertise in
clinical workflows, documentation, patient safety, and technology implementation, ANI offers
the following recommendations. These comments address regulated device functions and
identify areas where the boundary between regulated and nonregulated applications warrants
clarification.
1. Refine the risk framework beyond a presumed linear progression (Questions 1–6)
The four activity categories, from non-directive information to fully autonomous action, are
useful descriptors but should not imply that risk necessarily rises in that order. An actiondirecting system offering unsafe chest-pain triage advice could present a far greater patient
safety risk than an autonomous scheduling system performing routine administrative tasks.
Similarly, the consequences of a clinician-supervised action depend on the clinical decision
involved and the effectiveness of actual oversight.
FDA should explicitly state that autonomy is one modifier of risk, not a substitute for
evaluating intended use and severity of possible harm. Further, risk characterization should
incorporate the particular action, clinical condition, patient population, intended user,
deployment setting, reversibility, time sensitivity, and available safeguards. We recommend
defining what constitutes “action-taking” and “fully autonomous” in realistic clinical
workflows, including where a human reviews an action, has authority to intervene, and can
reasonably identify an error before harm occurs. FDA should also explain how the proposed
heuristic would interact with existing medical-device classification and software-device
frameworks rather than functioning as a parallel categorization system.
Explainability and interpretability at the point of care should be explicit risk modifiers.
Clinicians need clinically meaningful information about the data, evidence, and reasoning
relevant to a recommendation, even when technical model internals cannot be fully disclosed.
Proprietary design should not preclude an appropriate basis for evaluating and contesting
consequential outputs. Risk also depends on whether outputs are understandable to patients
with different health literacy levels and whether apparently low-risk information becomes
directive over a multi-turn exchange.

2. Clarify scope and assess ambient AI in real nursing environments (Questions 1, 9, 11, 17,
26)
FDA should provide clearer examples of which GenAI-enabled functions fall within medicaldevice oversight, including ambient documentation systems and agentic workflow tools.
Although many documentation applications may be outside the device definition, inaccurate
generated content can affect downstream care, and capabilities can evolve from transcription
or summarization toward clinical decision support. The distinction between regulated and
nonregulated functions should be transparent, including for products with multiple functions
or changing intended uses.
In particular, ambient documentation systems present use-context risks that the
activity axis does not capture. Nurses frequently document in noisy, interrupted, nonprivate
environments, including shared rooms and open bays. Evaluation should assess background
speech contamination, attribution of statements to the wrong patient or speaker, transcription
errors involving negation, missing or conflicting information, and accuracy when encounters
are fragmented across multiple interactions. Patient privacy and inadvertent capture of
bystander conversations also require attention. For outputs entering the health record,
sponsors should evaluate source fidelity, clinically consequential omissions, fabricated
statements, editability, auditability, and the clinician’s ability to verify content before use.
Clinical confirmation should test actual nurse–AI interactions and realistic work conditions, not
output accuracy alone. Standardized nursing simulations and representative workflows, such
as handoffs, medication administration, patient education, and escalation, can complement
retrospective and shadow evaluations. Studies should examine whether users detect and
correct errors amid interruptions, varying workload, and competing priorities.
3. Establish reproducible, clinically meaningful evaluation expectations (Questions 7–16)
ANI supports competency-based benchmarking paired with risk-appropriate clinical
confirmation. Because generative outputs can vary across repeated runs and model updates,
evaluation must assess the reliability of safety-critical behavior, including responses to
semantically equivalent inputs, wording changes, background noise, user health literacy, and
longer conversations. Sponsors should prespecify clinically meaningful failure thresholds,
report variability and subgroup performance, and demonstrate that performance remains
acceptable under expected deployment conditions. A high average score should not obscure
infrequent but serious errors.
Evaluation comparators and adjudication panels should reflect the device’s actual
function and intended users. Questions framed exclusively around generalist and specialist
physicians should include nurses, advanced practice registered nurses, and other qualified
clinicians when relevant. Nursing expertise is necessary for evaluating nursing-facing
applications, and the performance of the clinician–AI team should be distinguished from standalone model performance. Independent testing and expert adjudication can strengthen
confidence, particularly when benchmark cases reflect the messy realities of care rather than
idealized single-turn scenarios.

4. Make postmarket surveillance actionable and accountable (Questions 18–24)
ANI agrees that a premarket assessment cannot fully anticipate open-ended inputs, evolving
model behavior, and changing deployment environments. However, postmarket monitoring
should not be treated as an unqualified substitute for premarket evidence, particularly for
high-consequence functions. FDA should specify risk-based expectations for continuous or
periodic evaluation, clinically meaningful triggers for reassessment, independent review of
representative outputs, and clear processes for remediation, suspension, or rollback when
unacceptable performance is detected.
Monitoring should include structured opportunities for nurses and other frontline
clinicians to report AI-related near misses, workflow disruptions, unsafe recommendations,
documentation errors, and unanticipated patterns of use. These reports can identify problems
that routine accuracy metrics and developer-led surveillance may overlook. FDA should
explore ways for clinical organizations and professional societies to aggregate and
communicate these signals through pathways that complement existing safety reporting, while
manufacturers remain responsible for evaluating and correcting device problems. Changes to
underlying foundation models should trigger appropriate notification and regression testing,
whether manufacturers or third-party developers initiate them.

5. Include nursing informatics expertise in FDA deliberation and oversight
Nurses comprise a major share of the healthcare workforce and are integral to how AIgenerated information is interpreted, acted on, documented, and communicated to patients.
FDA advisory deliberations, clinical evaluation design, independent adjudication, and
postmarket oversight should reflect their perspectives. ANI encourages FDA to seek qualified
nursing and nursing informatics representation on the Digital Health Advisory Committee and
related working groups, and to engage nursing organizations when developing subsequent
guidance and evaluation methods.

ANI appreciates FDA’s openness to stakeholder input and welcomes opportunities to develop
further clinically grounded, feasible approaches to GenAI safety and effectiveness. We
encourage a framework that recognizes the multidimensional nature of risk, the realities of
nursing workflows, and the essential contribution of frontline clinicians to evaluation and
surveillance.

Sincerely,

Karen Dunn-Lopez, PhD, MPH, RN, FAAN Nancy Beale, PhD, RN, NI-BC, FAMIA
ANI Co-chair ANI Co-chair

The Alliance for Nursing Informatics (ANI), cosponsored by AMIA & HIMSS, advances nursing informatics
leadership, practice, education, policy, and research through a unified voice of nursing informatics
organizations. We transform health and healthcare through nursing informatics and innovation. ANI is a
collaboration of organizations representing more than 25,000 nurse informaticists and bringing together
29 distinct nursing informatics groups globally. ANI crosses academia, practice, industry, and nursing
specialty boundaries and collaborates with more than 4 million nurses in practice today. Contact ANI.