Daniel Dvoskin

Docket FDA-2026-N-7874 — Comment on Considerations for the Regulation of Generative AI-Enabled Medical Devices

I am a pediatrician (DO) practicing in Texas, trained at Baylor College of Medicine/Texas Children’s Hospital. I own an independent pediatric practice, and I also build clinical software. I am writing to urge FDA to adopt a strict premarket and postmarket framework for GenAI-enabled devices, with explicit protections for children.

1. The burden of proving safety must rest on the manufacturer, for each population claimed.
GenAI devices should not be authorized on benchmark performance alone or through substantial equivalence to predicates that are not generative AI. The manufacturer should have to show safety and effectiveness in each intended population before marketing. A device without that evidence should not be marketed or labeled for that population.

2. Children are currently unprotected by default.
Of AI-enabled devices FDA authorized through mid-2024, only 4.4% carried specific pediatric labeling, and only five were exclusively pediatric. Earlier analyses found that most AI device authorizations either made no reference to pediatric use or were adult-specific, which leaves clinicians to apply adult-trained models to children off-label. Children’s physiology, disease presentation, and normal ranges differ by age. A model validated on adults is not validated on a 2-year-old. I request that FDA:
- require age-stratified performance data using FDA’s pediatric subpopulation ranges;
- require labeling that explicitly states the age ranges in which the device has not been evaluated; and
- prohibit pediatric claims unless they are supported by pediatric data.

3. Clinical confirmation, not just benchmarks.
The discussion paper compares GenAI evaluation to physician training. That comparison supports my position. Physicians are supervised and credentialed separately for pediatric care, and a GenAI device should face the same population-specific confirmation.

4. The deployed version must be the reviewed version.
Any change to the underlying model should trigger re-evaluation, including updates to a third-party foundation model. The only exception should be changes covered by a predetermined change control plan with pediatric performance thresholds set in advance.

5. Mandatory postmarket surveillance.
Adverse event reporting for GenAI devices should be mandatory, not voluntary. Each report should include patient age and model version. Manufacturers should monitor real-world performance against pre-specified thresholds that automatically trigger suspension or relabeling.

6. Support clinician oversight.
Texas law (SB 1188) already requires practitioners to review AI-generated output before clinical decisions. Federal labeling should make that review possible by requiring outputs to show their uncertainty and the limits of their intended use.

7. Make strict standards predictable.
Pediatric AI devices already face longer reviews and higher evidentiary expectations without formal requirements. An informal higher bar discourages manufacturers from developing pediatric devices at all. I ask FDA to publish explicit pediatric evidence standards and to support pediatric data infrastructure through its pediatric device consortia. That way a strict standard becomes a clear path for manufacturers rather than an unwritten barrier.

Children should not be the population on which GenAI medical devices are tested after authorization.

Conflict of interest: I have no financial relationships with AI or medical device manufacturers.

Daniel Dvoskin, DO
Jolly Giant Pediatrics, Taylor, Texas