Nnaemeka Okafor

See attached file(s).

To: Food and Drug Administration (FDA) / Center for Devices and Radiological Health (CDRH)
Docket Number: FDA-2026-N-7874
Re: Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion
Paper and Request for Feedback
Date: October 4, 2026

To the Members of the CDRH Digital Health Center of Excellence:
As an actively practicing emergency physician and Chief Analytics and Informatics Officer
overseeing enterprise AI deployment, my perspective on generative AI is rooted in its
operational reality at the bedside. As these tools transition from controlled test environments to
the clinical floor, regulatory frameworks must bridge mathematical validation with the complex
human behaviors surrounding their use.

1. Risk Assessment & The Blind Spot of Cognitive Deskilling

The FDA’s proposed two-axis framework—evaluating clinical function against the severity of
incorrect output—is a necessary foundation, but it structurally assumes an idealized, constant
level of human vigilance. In clinical deployment, human oversight is neither static nor
friction-free.
Generative AI produces highly fluent, syntactically coherent outputs that easily mask factual
confabulations. When evaluating risk, the Agency must account for the downstream cognitive
burden these tools create. Every clinical decision involves four sequential stages: Appraisal,
Instruction, Verification, and Commitment. A device’s true risk profile is governed by how many
of these stages are stripped away upstream. If an agentic tool or an automated summarizer
pushes a predetermined insight, the clinical instruction was decided by a vendor months ago.
This shifts a massive, unresourced verification burden onto the clinician. As the cognitive effort
required to validate an action increases—and as the AI acts as a "black box"—clinicians are
systematically driven to blindly accept the output because it is the path of least resistance. This
leads directly to automation bias and selective deskilling over time.

●​ Incorporate a Third Risk Axis: The Degree of Verification Burden. The FDA should
expand its heuristic to measure the cognitive effort required for a clinician to genuinely
validate an AI's output. The danger lies in how AI interfaces are intentionally designed to
make blind acceptance the path of least resistance. Just as it is far easier to blindly
 accept a completed order from a personal shopper than to meticulously audit every item
in the bag against the original instructions, clinical AI interfaces use polished, highly fluent
summaries to seduce the clinician into accepting the output without cross-referencing the
source data. The higher the verification burden, the higher the premarket regulatory
hurdle must be.
●​ Mandate "Provenance Accounting": GenAI clinical tools must not simply present a
finished narrative. To combat the verification burden, an AI tool must explicitly display
what facts it corroborated, what relevant data it omitted, and where it synthesized
ambiguities, much like a personal shopper providing an itemized receipt of substitutions.
●​ Align with Workforce Well-Being: In alignment with the National Academy of Medicine’s
AI Code of Conduct, the total cognitive load imposed by automated systems must be
measured. We must account for the vigilance burden AI quietly puts back onto the
cognitive plates of nurses and physicians.

2. Premarket Evidence: The Limits of Decentralized Evaluation

While voluntary Foundation Model Device Master Files and competency-based evaluations
represent a step toward transparency, transparency alone is insufficient if the end-user lacks the
capacity to interpret it.
The volume and complexity of generative AI tools have scaled exponentially beyond standard
human comprehension. We have reached a threshold where AI systems are generating other AI
systems; the sheer volume of code requires specialized AI just to contextualize and audit the
models. Expecting decentralized, individual health systems to independently evaluate technical
master files, track performance trends, and validate vendor claims is neither practical nor safe.
Furthermore, vendor-provided premarket data often reflects highly controlled environments that
do not map to the specific demographics or workflows of a local health system.
●​ Establish Centralized Certification: The industry requires an authoritative "stamp of
approval" from the FDA or FDA-accredited third-party governing entities with the
dedicated time and technical resources to review massive device master files.
●​ Translate Transparency into Actionability: Pre-market transparency rules must require
centralized bodies to distill technical evaluations into standardized, clinically actionable
risk classifications, rather than passing raw data to health systems for interpretation.
3. Postmarket Monitoring: Managing Implementation Drift

Reviewing postmarket evidence through robust device performance monitoring is critical, but it
extends far beyond tracking algorithmic degradation. A mathematically sound tool can quickly
become a liability if end-users do not understand its features or apply it outside of appropriate
use conditions.
Postmarket evidence will vary dramatically across different health systems based on localized
implementation styles, version control, and workflow integration.
●​ Promote Hub-and-Spoke Governance Models: To manage the complexity of
postmarket monitoring, health systems must adopt structured hub-and-spoke
organizational models. A central "hub" ensures standardized oversight and vendor
management, while localized "spokes" are embedded within specific clinical units. These
pods are uniquely positioned to monitor localized behavioral drift—ensuring the tool is
utilized under appropriate conditions and capturing contextual, bedside feedback that a
remote vendor dashboard would entirely miss.
●​ Monitor Human Factors Postmarket: Postmarket surveillance mandates must track
user interaction metrics alongside model accuracy. If clinicians are systematically
bypassing verification steps, this must trigger safety reviews just as readily as a drop in
predictive precision.
Thank you for your attention to these critical dimensions of clinical safety and cognitive
stewardship.
Sincerely,
Nnaemeka Okafor, MD, MS
Chief Analytics and Informatics Officer
Disclaimer: The views, thoughts, and opinions expressed in this comment belong solely to the
author in his professional capacity and do not necessarily reflect the official policy or position of
his employer or affiliated health systems.