Explainable EEG Foundation Models for Schizophrenia Detection Across Independent Datasets

Abstract Schizophrenia is a complex psychiatric disorder associated with altered brain electrophysiology. EEG foundation models provide promising approaches for learning transferable neural representations; however, their effectiveness and interpretability for schizophrenia detection remain insufficiently investigated. Three pretrained EEG foundation models, EEGMamba, EEGPT, and BIOT, were systematically evaluated for schizophrenia classification, with Simple Mamba included as an additional comparison. Two independent resting-state EEG datasets were analyzed: ASEEG, comprising 51 individuals with schizophrenia (SZ) and 50 healthy controls (HC), and Warsaw, comprising 14 SZ and 14 HC participants. Subject-wise cross-validation was employed. Integrated Gradients (IG) and channel occlusion were used to investigate spatial importance and explanation consistency. EEGMamba achieved the strongest performance on both datasets. On ASEEG, it obtained 79.18 ± 8.00% accuracy, 82.33 ± 12.54% sensitivity, an F1-score of 0.79 ± 0.10, and ROC-AUC of 0.83 ± 0.09. On Warsaw, EEGMamba achieved 78.33 ± 24.90% accuracy and 80.00 ± 14.21% sensitivity. XAI revealed model- and dataset-dependent patterns involving frontal/frontopolar, central, temporal, and parietal regions, with recurrent importance of C3, C4, Cz, Fp1/Fp2, and P3/P4. On Warsaw, BIOT demonstrated strong agreement between IG and occlusion (ρ = 0.744, p < 0.001). EEG foundation models show potential for subject-independent schizophrenia detection, with EEGMamba providing the strongest predictive performance. Complementary XAI analyses revealed spatially distributed discriminative information and highlighted the importance of jointly evaluating predictive performance and interpretability across independent datasets.