Abstract
Predicting cognition from neuroimaging data in healthy individuals offersinsights into the neural mechanisms underlying cognitive abilities, withpotential applications in precision medicine and early detection ofneurological and psychiatric conditions. This study systematically benchmarkedclassical machine learning (Kernel Ridge Regression (KRR)) and advanced deeplearning (DL) models (Graph Neural Networks (GNN) and Transformer-GNN (TGNN))for cognitive prediction using Resting-state (RS), Working Memory, and Languagetask fMRI data from the Human Connectome Project Young Adult dataset. Our results, based on R2 scores, Pearson correlation coefficient, and meanabsolute error, revealed that task-based fMRI, eliciting neural responsesdirectly tied to cognition, outperformed RS fMRI in predicting cognitivebehavior. Among the methods compared, a GNN combining structural connectivity(SC) and functional connectivity (FC) consistently achieved the highestperformance across all fMRI modalities; however, its advantage over KRR usingFC alone was not statistically significant. The TGNN, designed to modeltemporal dynamics with SC as a prior, performed competitively with FC-basedapproaches for task-fMRI but struggled with RS data, where its performancealigned with the lower-performing GNN that directly used fMRI time-series dataas node features. These findings emphasize the importance of selectingappropriate model architectures and feature representations to fully leveragethe spatial and temporal richness of neuroimaging data. This study highlights the potential of multimodal graph-aware DL models tocombine SC and FC for cognitive prediction, as well as the promise ofTransformer-based approaches for capturing temporal dynamics. By providing acomprehensive comparison of models, this work serves as a guide for advancingbrain-behavior modeling using fMRI, SC and DL.