TS-GAT+FC+Adv: Connectivity-Guided Graph Attention for EEG Schizophrenia Detection under Subject-Independent Evaluation

Schizophrenia is a chronic psychiatric disorder diagnosed almost entirely through clinical interview. Early detection matters, since longer untreated psychosis is associated with poorer outcome, yet no objective electrophysiological marker is in routine use. Deep learning on resting-state electroencephalography has been reported to separate patients from controls with very high accuracy, but most such reports do not state how the training and test data were divided. Where epochs from one recording fall on both sides of that division, a network can succeed by recognising the individual rather than the illness. Before such models can be considered clinically, the field needs evidence of how they behave on unseen people. We provide that evidence on a public cohort of twenty-eight participants, half diagnosed with paranoid schizophrenia, recorded over nineteen scalp channels. Four architectures were trained under one common budget and evaluated twice. Under record-level cross-validation, all four exceed ninety-nine percent and are indistinguishable. Under participant-grouped cross-validation of the same trained models, accuracy falls to the mid-to-upper seventies. Our graph attention network, which guides inter-electrode attention by scalp geometry and channel coupling and is trained adversarially against participant identity, attains the best grouped accuracy, the smallest fall between protocols and the most stable estimates. Paired testing nevertheless separates none of the four models, so the protocol effect is the firm finding.