Dynamic Adaptive Spatio-Temporal Graph Convolution for fMRI Modelling

The characterisation of the brain as a functional network in which connections between regions are represented by correlation values across time series has been very popular last years. Although this representation advanced our understanding function, it represents simplified model connectivity that complex dynamic spatio-temporal nature. Oversimplification data may hinder merits applying non-linear feature extraction algorithms. To end, we propose adaptive graph convolution (DAST-GCN) to overcome shortcomings pre-defined static correlation-based structures. proposed approach allows end-to-end inference via layer-wise structure learning module while mapping phenotype supervised framework. This leverages computational power model, and targets represent connectivity, could enable identification potential biomarkers for target question. We evaluate pipeline on UKBiobank dataset age gender classification tasks from resting-state scans show outperforms currently adapted linear methods neuroimaging. Further, assess generalizability inferred transferring pre-trained an independent same task. Our results demonstrate task-robustness against different scanning parameters demographics.