Brandon Kruse

I am a board-certified psychiatric-mental health nurse practitioner and family nurse practitioner operating a telehealth psychiatry practice across six states. I use AI tools daily in clinical workflows. I write to urge FDA to account for mental-health realities as it develops its framework for generative AI-enabled devices.

**1. The risk framework’s consequence axis underweights psychiatric harm.**
The discussion paper’s two-axis framework — degree of device activity versus consequences of relying on an incorrect output — is sound, but "consequences" reads as if harm means a device malfunction. In psychiatry, the harm pathway is a changed clinical decision. A confabulated but fluent medication summary, risk assessment, or treatment suggestion does not look broken; it looks authoritative. There is no lab value, image, or vital sign downstream to catch it — only a clinician’s judgment, often exercised under time pressure, over video, with a patient in crisis. The framework should name changed-treatment-decision as a consequence category, not just device failure modes.

**2. Premarket evaluation must include mental-health-specific adversarial testing.**
Sponsors should be required to test how a device performs with patients expressing suicidality, psychosis, and mania — not just general medical Q&A. Two failure modes deserve explicit attention: (a) sycophantic outputs that affirm delusional or grandiose thinking, which feel "helpful" while being dangerous; and (b) confident de-escalation advice that is subtly wrong in ways only a trained psychiatric clinician would catch. Standard accuracy benchmarks will not surface these. Evaluation should include structured adversarial cases drawn from real psychiatric presentations.

**3. Postmarket monitoring must happen in real telehealth workflows.**
Performance degradation across the product life cycle will not be visible in lab conditions. Require sponsors to monitor device outputs in deployed settings — including telehealth, where the clinician cannot read the room the way an in-person visit allows — across diverse patient populations. Critically, uncertainty must be visible to the clinician using the tool, not buried. A device that reports confidence it does not have is more dangerous than one that admits limits.

**4. Treat the clinician handoff as part of the device’s risk surface.**
For any function above nondirective information, the person answering the patient’s "what does this mean?" is usually a nurse or nurse practitioner. The framework should treat that handoff — what the clinician sees, what caveats travel with the output, how disagreement is surfaced — as part of what gets evaluated, not as someone else’s problem.

**5. Solicit the practitioner voice before guidance.**
The clinicians who will live with these tools are underrepresented in this docket. I urge the Center to actively solicit comment from practicing nurses, nurse practitioners, and frontline prescribers — not only manufacturers and health-system executives — before drafting guidance. We are the ones who will explain these outputs to patients, notice when they are wrong, and decide what to do next.

Respectfully submitted,
Brandon Kruse, PMHNP-BC, FNP-BC