Brain-State-Dependent Biomarker Discovery from Multi-Condition EEG: A Structured Signal Framework for Schizophrenia with Interpretable Graph Neural Analysis

Multi-condition EEG is a structured physiological signal whose utility as a clinical digital biomarker is fundamentally brain-state-dependent—a property ignored by single-condition paradigms that assume state-invariant biomarkers. We propose a dual-stream GCN-Transformer on epoch-wise dynamic ROI-level PLV graphs across six cognitive paradigms on 60 subjects (40 schizophrenia, 20 healthy controls), using structured graph interventions to quantify state-dependent regional importance. Music perception yields the strongest digital biomarker signal (84.73%, ±1.23%)—outperforming resting-state by 4.36% with 2× lower variance—while we observe that occipital regions are dominant in five of six conditions and frontal regions emerge exclusively under post-task and working-memory paradigms, with the same region being informative in one state and suppressive in another. LOSO validation achieves 92.00±2.45% (Warsaw) and 89.00±2.62% (SUSZ), outperforming all baselines, with a two condition music-arithmetic protocol as a promising biomarker acquisition strategy