Validation of interpretability findings in ASD classification models

Ascertain whether interpretability outputs from deep learning classifiers trained on resting-state fMRI functional connectivity data for Autism Spectrum Disorder classification reflect genuine ASD-related neurobiological characteristics rather than dataset-specific patterns, by establishing validation procedures that can verify the correspondence between attributed features and ASD mechanisms.

Background

The paper highlights a gap in ASD research: interpretability methods are underused for functional connectivity data, and when used, their findings are rarely validated. Without validation, it is uncertain whether the highlighted features capture ASD-specific neurobiology or merely reflect idiosyncrasies of the dataset.

The authors explicitly state that this lack of validation leaves uncertainty about what models are actually learning, motivating the open problem of developing and applying robust validation frameworks for interpretability in ASD neuroimaging.