Disrupted Cerebello-Thalamo-hippocampal network topology in internet gaming disorder: Evidence from explainable graph convolutional networks.

BACKGROUND: Internet gaming disorder (IGD) is associated with disrupted brain network topology, but explainable, network-level neuroimaging biomarkers remain scarce, since conventional vector-based classifiers discard the topological relationships among brain regions that a graph convolutional network (GCN) can directly exploit. METHODS: We trained an explainable GCN on individualized whole-brain functional connectivity graphs (AAL-116 atlas) from 784 young adults (453 with IGD, 331 healthy controls; aged 17-32 years) under 10-fold stratified cross-validation, benchmarked it against a linear support vector machine (SVM), identified the most discriminative regions using class activation mapping (CAM) and gradient-weighted CAM (Grad-CAM), and examined associations between regional nodal efficiency and symptom severity indexed by DSM-5 criteria and the Internet Addiction Test (IAT) within the IGD group. RESULTS: The GCN outperformed the SVM (balanced accuracy 62.60% ± 4.27% vs 57.65% ± 5.38%) with a more balanced sensitivity-specificity profile; CAM, Grad-CAM, and SVM each implicated the cerebellum among the most discriminative regions, with CAM and Grad-CAM sharing seven of their top-10 regions; and after FDR correction, nodal efficiency was significantly and negatively correlated with DSM-5 severity in the left cerebellar lobule X, vermal lobule IX, and left thalamus, with a marginally significant correlation in the right hippocampus; no nodal efficiency-IAT correlations survived correction. CONCLUSIONS: These convergent classification, attribution, and nodal-efficiency findings converge on the cerebellum-alongside the thalamus and hippocampus-as explainable-AI-identified regional substrates whose nodal efficiency tracks DSM-5-defined symptom severity in IGD.