Comment from George Evgrafov
AnonymousSupportIndividual
Summary: The commenter suggests that the FDA should strengthen the draft guidance by requiring sponsors to perform specific bias detection and mitigation as part of their AI model credibility assessments. They argue that these requirements, including subgroup performance analysis and documentation of fairness checks, are necessary to ensure patient safety and regulatory integrity.
Bias and Fairness in AI Outputs
The guidance briefly touches on data quality but does not adequately address the issue of bias in model outputs, which can impact regulatory decisions and patient safety.
Recommendation: I suggest the FDA encourage sponsors to perform bias detection and mitigation as part of their credibility assessments. This may include subgroup performance analysis across demographics, documentation of model fairness checks, and plans to mitigate discovered bias. These expectations should be part of the validation and model monitoring strategy.