Abstract
Multimodal brain connectivity has emerged as a framework for characterising the structural, functional, effective, and metabolic organisation of the brain by integrating complementary neuroimaging and neurophysiological modalities. This review examines the methodological evolution of brain connectivity analysis, from classical connectivity measures and graph-theoretical approaches to contemporary artificial intelligence-based methods, with particular emphasis on Graph Neural Networks (GNNs).
First, the principal concepts of structural, functional, effective, and metabolic connectivity are introduced, together with the classical techniques used to estimate these forms of connectivity and the graph-theoretical metrics employed to characterise brain networks. Subsequently, the fundamental principles of GNN-based approaches are described, including Graph Convolutional Networks, Graph Attention Networks, Spatiotemporal Graph Neural Networks, and Graph Transformers, highlighting their capacity to model non-linear, higher-order, and dynamic relationships within brain networks. The review also examines current training strategies, including supervised, self-supervised, contrastive, transfer, federated, and few-shot learning, particularly in the context of limited and heterogeneous clinical datasets. Multimodal integration strategies are then discussed, ranging from conventional statistical and graph-based approaches to GNN architectures designed to jointly model structural, functional, electrophysiological, and metabolic information. Particular attention is given to clinical applications in epilepsy, neurodegenerative disorders, and post-stroke rehabilitation, highlighting the potential of multimodal connectivity and graph-based learning for diagnosis, prognosis, biomarker identification, and the characterisation of brain reorganisation.
Finally, current challenges related to data heterogeneity, model interpretability, clinical validation, generalisability, ethical considerations, and translation into clinical practice are discussed. Overall, the convergence of multimodal connectomics and graph-based artificial intelligence represents a promising direction towards more comprehensive and personalised models of brain organisation and neurological disease.