Comment from Elvan Ceyhan

Comments on “Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products” In what follows, I provide the comment on a particular section and then provide suggestions afterwards. I. INTRODUCTION The introduction provides a general overview but could benefit from a clearer statement of the specific AI applications targeted. Consider elaborating on whether the guidance applies to AI models used in exploratory research or only to those directly influencing regulatory decisions. Specify whether AI applications in early-stage research are within the guidance’s scope or if the focus is exclusively on decision-making for drug approval and post-market evaluation. II. SCOPE The scope section provides exclusions (e.g., drug discovery) but does not specify whether AI used in patient recruitment or trial design is covered. Clarify if AI applications in patient recruitment, trial simulations, or protocol optimization fall within the regulatory framework outlined in this guidance. IV. CONSIDERATIONS FOR AI USE IN THE DRUG PRODUCT LIFE CYCLE IV-A. A Risk-Based Credibility Assessment Framework: •The 7-step framework is well-structured but lacks explicit mention of how sponsors should handle multi-stage AI models (e.g., ensemble learning methods). Include guidance on evaluating complex AI architectures, such as ensemble models or transfer learning, particularly in regulatory submissions. •The guidance introduces context of use (COU) but does not provide concrete examples of regulatory precedents where COU was effectively defined. Provide a case study or historical example where COU definition played a pivotal role in AI model acceptance or rejection, if available. •The model risk matrix provides a useful framework, but the guidance does not explain how sponsors should document model influence assessments. Recommend standardized documentation templates for assessing AI model influence and decision consequence. •While the guidance stresses data quality, it does not mention synthetic data as a potential solution for bias mitigation. Consider including recommendations for the use of synthetic data to supplement training datasets and reduce bias in model development. •Performance metrics are discussed, but it is unclear whether FDA expects continuous monitoring of AI performance post-deployment. Clarify whether AI models used in regulatory decisions require ongoing performance validation or if a single validation submission is sufficient. IV-B. Special Consideration: Life Cycle Maintenance of the Credibility of AI Model Outputs in Certain Contexts of Use The guidance touches on model drift (in the form of data drift) but lacks a structured approach for AI model revalidation when performance degrades over time. Introduce a systematic revalidation schedule or monitoring threshold to determine when AI models should be retrained or recalibrated. IV-C. Early Engagement The guidance encourages early engagement but does not specify response timelines or expected feedback formats. Provide estimated response timelines and examples of the level of detail expected in initial AI model discussions with FDA. Disclaimer: The views expressed in this comment are my own and do not represent the views of Auburn University or any other affiliated institution.