Improved Brain Disorder Classification in RS-FMRI via Inter-Individual Variability Characterization: A Self-Supervised Learning Approach

In the era of precision medicine, understanding and explaining inter-individual variability in brain function is in-creasingly important. Functional connectomes (FCs), typically assessed using Pearson's Correlation Coefficient (PCC) from sufficiently long resting-state fMRI (rs-fMRI) data, are known for their reliability in subject-fingerprinting, reflecting robustness to inter-individual variability. However, the reliability of PCC-based FCs deteriorates significantly as rs-fMRI signal length decreases. Furthermore, while there is a theoretical foundation linking inter-individual variability to precision medicine, practical studies confirming this link remain scarce. To address these gaps, we introduce VarCoNet, a contrastive self-supervised learning framework aimed at improving subject fingerprinting from short rs-fMRI acquisitions and enhance brain disorder classification. VarCoNet was evaluated on rs-fMRI data from two datasets including 1,095 subjects from the Human Connectome Project for subject fingerprinting, and 1,112 subjects from the ABIDE I dataset for autism spectrum disorder (ASD) identification. VarCoNet outperformed conventional PCC-based subject fingerprinting by up to 28%, and performed similarly or better than a supervised model in ASD detection. Our study demonstrates the reliability of self-learning frameworks for inter-individual variability analysis to improve brain disorders classification.