Harnessing machine learning for functional connectivity-based feature discovery in post-traumatic epilepsy

•Leakage-free nested CV-RFE identifies a reproducible connectivity signature that characterizes the PTE-diseased state in our cohort.•Data-driven feature selection is essential for classification performance, with only four stable functional connections (from an original feature space of 4950) able to distinguish PTE from TBI with an AUC of 0.95.•We identify a specific connectivity phenotype in our PTE cohort which includes the Visual network as a hub of PTE-related changes, left-hemisphere overrepresentation and hyper-coupling, and right-hemisphere un-coupling.•The resting-state network disruptions identified by our ML paradigm went unrecognized by direct comparison between groups and traditional statistics, supporting the use of advanced data-driven methodologies for exploring functional connectivity in PTE.