Equity-Aware Self-Supervised Multi-Domain Connectome Representation Learning for Brain Functional Connectivity

Reliable functional-connectivity learning requires representations that capture complementary neural interactions, generalize across sites with limited labels, and reduce demographic bias. Yet most rs-fMRI methods rely on correlation-based connectivity and address fairness only at prediction time. We introduce \textbf{RESOLVE} (\textbf{R}obust and \textbf{E}quitable \textbf{S}elf-supervised learning for functi\textbf{O}na\textbf{L} connecti\textbf{V}ity mod\textbf{E}ling), an equity-aware graph framework that jointly learns time- and frequency-domain connectomes, preserves brain-network topology through reconstruction, and promotes demographic invariance via Optimal Transport. Evaluated on REST-meta-MDD for cross-site MDD diagnosis, RESOLVE achieves the best AUC, accuracy, and fairness--accuracy trade-off among supervised, self-supervised, and graph-based baselines.