PI-FC: Pre-training Individual-specific Functional Connectome through State-invariant Contrastive Learning

Abstract

Functional MRI enables non-invasive mapping of brain connectivity, yet its clinical translation remains hindered by uncontrolled state-dependent variability that obscures individual-specific signatures during routine scanning. Here we introduce PI-FC -- a deep learning framework leveraging state-invariant contrastive learning to extract stable individual brain signatures across diverse arousal levels, cognitive states, and temporal scales spanning tens of seconds to hours. PI-FC achieves equivalent phenotypic prediction accuracy using substantially reduced scanning time, and eliminates state-dependent effects varying task demands and brain states. Trained on 36,119 subjects across 8 independent datasets, our model demonstrates superior cross-site generalization and outperforms traditional functional connectome (FC) in predicting neuropsychiatric conditions including schizophrenia, autism, depression, and anxiety. Furthermore, PI-FC enables zero-shot inference of brain age, biological sex, and cognitive ability without site-specific retraining. Overall, PI-FC represents a robust, clinically scalable framework that overcomes fundamental barriers to real-world deployment of precision functional neuroimaging.