B. Braun

I appreciate FDA’s effort to develop a risk-based framework for generative AI-enabled medical devices and agree that many of the considerations described in the discussion paper are appropriate for AI systems that generate clinical interpretations, recommendations, diagnoses, triage decisions, or other outputs on which users must directly rely. However, I encourage FDA to distinguish more explicitly between these systems and AI that serves primarily as a technical enabling technology within a medical device.

Many future medical devices are expected to incorporate AI not to provide clinical reasoning or interpretation, but to improve the performance of a technical function. Examples may include image enhancement, signal processing, sensor fusion, measurement optimization, robotic control, autofocus functions, workflow optimization, or other performance-enhancing technologies. In these cases, AI may significantly contribute to device performance, yet the user is not relying on an AI-generated clinical interpretation or recommendation. Instead, the user evaluates the resulting device performance directly.

This distinction is important because the regulatory questions differ substantially. For clinical or cognitive AI functions, users must often trust information generated by the AI, creating concerns regarding explainability, reasoning, uncertainty, bias, and appropriateness of clinical recommendations. By contrast, for technical AI functions, the primary question is often whether the device performs safely, effectively, accurately, and consistently within its intended use. The internal method used to achieve that performance, whether conventional software, a statistical model, or AI, may be of secondary importance provided that performance can be objectively verified and validated.

As currently written, many elements of the proposed competency-based framework appear oriented toward systems exhibiting clinical knowledge, reasoning, communication, decision support, or conversational behavior. These concepts are highly relevant for clinical decision-support systems and generative AI applications but may have limited applicability for AI components whose role is purely technical. For example, benchmarking against clinician performance or evaluating clinical reasoning competencies may not provide meaningful evidence for an AI-enabled image enhancement or robotic control function.

I encourage FDA to consider explicitly distinguishing between:

(1) Interpretation-centered AI, where users rely on AI-generated clinical information, recommendations, or reasoning; and

(2) Performance-centered AI, where AI functions primarily as a technology that enables or improves a technical device capability.

For the latter category, objective performance verification and validation, supported by established engineering and risk-management principles, may often provide a more appropriate basis for demonstrating safety and effectiveness than evaluation approaches focused on explainability or clinician-like competencies.

Without such a distinction, there is a risk that regulatory expectations developed for clinical reasoning systems could unintentionally be applied to technical AI functions for which those concepts are not meaningful or risk-proportionate. Explicit recognition of these different categories could improve regulatory predictability, maintain a risk-based approach, and avoid unnecessary barriers to innovation while continuing to ensure the safety and effectiveness of AI-enabled medical devices.