Greg Gutkowski

**Subject: Closing the Between Visit Gap in Generative AI Medical Device Governance**

I appreciate the FDA’s initiative to examine the regulation of generative AI enabled medical devices. I would like to address a critical gap in AI governance: the inability to consistently monitor, document, and evaluate what happens to patients between clinical visits.

**The Problem**

Healthcare information is fragmented across electronic health records, laboratories, pharmacies, specialists, wearable devices, and patient observations. Yet much of what determines a patient’s health happens between clinical encounters.

Consider a patient discharged after treatment for congestive heart failure.

Over the following two weeks, a connected scale detects weight gain, a wearable device records declining activity, and the patient reports increasing fatigue. An AI enabled medical device identifies potential deterioration and recommends clinical review.

But what happens next?

Was the recommendation communicated? Did a clinician respond? Was an intervention made? Did the patient improve or require hospitalization?

These events may be recorded in separate systems, making it difficult to determine whether the AI performed appropriately, whether human intervention occurred, and what ultimately happened to the patient.

**A Proposed Solution: HealthWallet**

I have been developing a conceptual architecture called HealthWallet, designed to connect patient information, AI recommendations, clinical actions, and subsequent outcomes.

It includes four components:

**1. Longitudinal Patient Record**

A patient centered record that consolidates authorized information from clinical systems, pharmacies, laboratories, wearables, and patient observations, while preserving data provenance and appropriate privacy controls.

**2. Continuous AI Supported Monitoring**

AI identifies significant changes, missing information, potential risks, and situations requiring clinical attention. Its role is to support monitoring and escalation, not replace licensed healthcare professionals.

**3. Explicit Clinical Accountability**

An Extended Care Navigation System (ECNS) coordinates monitoring, communication, and followup.

An Extended Care Clinic (ECC) provides professional clinical evaluation and intervention when needed.

Every significant AI alert should have a documented recipient, response, and disposition.

**4. Independent Outcome Evaluation**

A Real Outcomes Research Center (RORC) function evaluates the evidence connecting AI observations, recommendations, human decisions, interventions, and patient outcomes.

Independent reviewers could examine whether the system operated as intended, where failures occurred, and whether corrective actions were effective.

**Recommendations to the FDA**

For generative AI functions regulated as medical devices, particularly those supporting ongoing patient monitoring and clinical escalation, I recommend that the FDA consider:

1. Requiring appropriate mechanisms to document significant AI recommendations, alerts, and escalation events.
2. Evaluating whether recommendations reach authorized healthcare professionals and receive appropriate responses.
3. Linking relevant AI outputs and clinical interventions to subsequent patient outcomes where feasible.
4. Supporting independent, risk proportionate assessment of longitudinal device performance.
5. Encouraging interoperable, patient centered architectures that preserve evidence across clinical encounters while protecting privacy.

These mechanisms could help distinguish failures of AI performance from failures of communication, clinical workflow, or human intervention.

**Conclusion**

Documenting that an AI system complies with governance requirements does not establish that it contributes to safe and effective patient care.

Similarly, verifying that an AI system produced an appropriate recommendation does not establish that the recommendation was communicated, acted upon, or associated with a better outcome.

The HealthWallet model illustrates one possible architecture for closing this evidence gap. It is proposed as a conceptual framework, not as a clinically validated or deployed system.

I encourage the FDA to consider longitudinal evidence and independent outcome verification as important elements of postmarket oversight for applicable generative AI enabled medical devices.

The fundamental question should not stop at whether the AI made the right recommendation. It should also examine what happened afterward.

Greg Gutkowski