Identification of Predictive Subnetwork for Brain Network-Based Psychiatric Diagnosis: An Information-Theoretic Perspective

Graph neural networks (GNNs) have recently been applied to develop useful diagnostic tools for psychiatric disorders. However, due to the lack of interpretability, clinicians are hard to identify quantifiable and personalizable biomarkers which provide biologically and clinically relevance. We introduce three recently proposed GNN-based psychiatric disorders diagnostic models, namely BrainIB, Graph-PRI and CI-GNN, from an information-theoretic perspective. These models are able to discriminate psychiatric patients from healthy controls and identify predictive subgraph, a.k.a. biomarkers, solely from observations. We demonstrate their improved classification accuracy and interpretability on ABIDE database. We also put forward three proposals for future research.