Ty Vachon

See attached file(s)

Re: Docket FDA-2026-N-7874

Considerations for the Regulation of Generative AI-Enabled Medical Devices

I am a board-certified diagnostic radiologist with experience in clinical informatics, AI-enabled
radiology workflows, clinical quality assurance, and medical technology development. I am
actively involved in evaluating workflows in which generative AI produces preliminary radiology
interpretations that are subsequently reviewed, modified, and finalized by licensed radiologists.

I appreciate FDA's thoughtful approach to developing a risk-proportionate regulatory framework.
I offer the following recommendations based on practical considerations in AI-assisted
diagnostic imaging.

1. Distinguish report concordance from diagnostic accuracy (Questions 14
and 17)

One of the most important challenges in evaluating generative AI for radiology is establishing a
meaningful reference standard.

An AI-generated preliminary report may demonstrate high agreement with a final radiologist
report. However, agreement does not necessarily establish clinical accuracy, particularly when
the radiologist has already reviewed the AI output.

Likewise, an AI report accepted without modification should not automatically be considered a
clinically correct interpretation.

I recommend that FDA encourage evaluation frameworks that distinguish:

●​ Agreement between AI-generated and physician-finalized reports.
●​ Accuracy against an independently established clinical reference standard.
●​ The ability of physician reviewers to identify and correct clinically significant AI errors.
●​ Residual errors in finalized reports despite AI assistance.

Independent expert adjudication of appropriately sampled studies should remain an important
component of validation and ongoing quality assessment.

2. Evaluate the complete human-AI workflow (Questions 4 and 14)

Human oversight should be assessed as an actual risk-mitigating capability rather than simply
the presence of a physician signature.

In radiology, an AI-assisted workflow may include automated image analysis, preliminary report
generation, clinical review, and final physician attestation.
The safety of that workflow depends on the reviewer's relevant expertise, access to source
images and prior examinations, ability to recognize incorrect AI outputs, and available review
time.

FDA should encourage evaluations that measure the performance of the combined human-AI
system under realistic operating conditions, including clinical workload and reading volume.

A critical question is not only whether the AI makes errors, but whether the intended reviewer
reliably detects and corrects those errors before they affect patient care.

3. Establish imaging-specific requirements for multimodal models
(Question 17)

Multimodal vision-language models introduce risks that may not be captured by conventional
text-based evaluations.

Imaging-specific assessment should include:

●​ Detection of clinically significant findings and omissions.
●​ Hallucinated findings or measurements not supported by the images.
●​ Correct utilization of prior examinations and clinical history.
●​ Laterality, anatomical localization, and image-to-report consistency.
●​ Recognition of incomplete studies, missing sequences, and technical limitations.

Testing should reflect the full deployed system, including image ingestion, metadata handling,
and integration into clinical reporting workflows.

4. Support risk-proportionate clinical confirmation (Question 11)

I support FDA's consideration of alternatives to prospective clinical trials when justified by the
device's intended use and risk profile.

For AI systems producing preliminary interpretations subject to meaningful physician review,
appropriately designed retrospective validation, independent adjudication, and prospective
shadow-mode evaluation may provide substantial evidence of clinical performance without
exposing patients to autonomous decision-making.

However, the presence of physician review should not automatically exempt a system from
appropriate premarket evaluation or other applicable regulatory requirements.

Evidence requirements should reflect the actual potential for patient harm and the demonstrated
effectiveness of safeguards.

5. Make postmarket clinical QA a central component of oversight
(Questions 19 and 21)
Generative AI performance cannot be fully characterized through a single premarket evaluation.

I encourage FDA to recognize structured, longitudinal clinical quality assurance as an important
component of a risk-proportionate total product lifecycle approach.

A practical radiology framework could include:

●​ Routine random sampling supplemented by risk-targeted review.
●​ Tracking of clinically significant AI errors, including errors corrected before finalization.
●​ Independent review of selected finalized reports to identify errors missed by the
supervising clinician.
●​ Stratification by modality, examination type, clinical setting, and relevant patient
populations.
●​ Version-linked records supporting investigation of performance changes.

Importantly, postmarket evaluation should distinguish AI errors successfully intercepted by
clinicians from errors that persist into finalized reports.

Manufacturers should retain responsibility for device performance monitoring and corrective
action, while clinical organizations contribute workflow-specific quality data and oversight.

Conclusion

I encourage FDA to develop a regulatory approach that recognizes meaningful differences
between autonomous AI diagnosis and AI-assisted interpretation under demonstrated clinical
supervision.

The objective should not be to regulate generative AI based solely on its underlying model
architecture, nor to treat human review as an automatic safety guarantee.

Instead, regulation should focus on intended clinical function, the actual deployed workflow,
independently demonstrated performance, and the ability to detect and mitigate clinically
meaningful errors over time.

Such an approach can promote both patient safety and responsible innovation in diagnostic
imaging.

Respectfully submitted,

Ty Vachon, MD

Board-Certified Diagnostic Radiologist

Submitted in an individual professional capacity.