Fig. 1. Process diagram depicting the AFISP workflow.

AFISP is an end-to-end framework for identifying subgroups on which an ML model may perform poorly. Given a dataset, pre-trained model, and set of user-specified features, the Stability Analysis phase (blue) allows a user to identify the worst-performing subset of the dataset on which the model has significantly deteriorated performance. Then, this data subset is processed to determine concrete subgroup phenotypes (interpretable subgroup descriptions; orange) present within the subset. Finally, an AFISP user applies model performance diagnostics (purple) to each identified subgroup to evaluate if the observed subgroup performance disparity is correctable through changes to the modeling pipeline.