121 Study design considerations for post-deployment monitoring of AI/ML models

Objectives/Goals: Safely and effectively translating AI classification models requires robust post-deployment monitoring. Yet there is little guidance about how to do so. Here, we outline how standard machine learning model development study designs are insufficient for post-deployment monitoring, and what specific study designs are needed. Methods/Study Population: We made original linkages between machine learning methodology and biostatistics, particularly to diagnostic testing, for study design guidance on post-deployment model performance and fairness assessment. Results/Anticipated Results: The kind of case–control sampling typically used for model development cannot give valid estimates of the Positive Predictive Value (precision) and may suffer from verification bias; in a pragmatic clinical trial, or under deployment, only the PPV can be measured, unless there is random confirmatory testing of predicted negatives. This is important for measuring sensitivity, specificity, and False Omission Rate as a fairness metric. Discussion/Significance of Impact: By linking well-understood biostatistics principles (e.g., diagnostic testing) to machine learning classifier evaluation, we show what is required for rigorous evaluation of models in clinical practice. This will establish clear guidance and rigorous standards of practice for study design and analytic aspects of post-deployment monitoring.