A Prediction-Coupled Self-Explaining Functional Connectivity Learning Framework for rs-fMRI-Based Brain Disorder Classification

Abstract Existing resting-state functional magnetic resonance imaging (rs-fMRI) functional connectivity (FC) models can achieve useful predictive performance, but their explanations are often generated after prediction and are not systematically evaluated for faithfulness and reproducibility. We propose SE-FC, a prediction-coupled self-explaining functional connectivity framework for interpretable FC learning. The explanation pathway learns signed coefficients over FC edges, whose interaction with standardized FC values produces signed edge evidence that directly contributes to the final prediction. Deletion, insertion, and average output probability change (AOPC) analyses showed that SE-FC-derived evidence was more strongly associated with model output than random edge rankings. Cross-run analysis further quantified explanation reproducibility by identifying recurrent consensus patterns while characterizing substantial variability in exact edge selection across independent training runs. Predictive capability was evaluated as supporting evidence under strict leave-one-site-out validation on ABIDE (435 participants, 10 sites) and SRPBS (925 participants, six sites), with three independent repeated training runs for each dataset. SE-FC achieved AUC values of 0.6133 ± 0.0044 on ABIDE and 0.6127 ± 0.0051 on SRPBS, indicating competitive cross-site predictive capability. Overall, SE-FC provides prediction-coupled connectivity evidence with quantitatively assessed model-level faithfulness and cross-run reproducibility while retaining useful predictive performance.