Toward interpretable schizophrenia detection from EEG using autoencoder and EfficientNet

Schizophrenia (SZ) is a complex mental disorder that necessitates accurate and timely diagnosis for effective treatment. Traditional methods for SZ classification often struggle to capture transient EEG features, face high computational complexity, and lack interpretability. This study proposes two complementary pipelines, one using a convolutional autoencoder (CAE) combined with an extreme gradient boosting (XGB) classifier. Alternatively, we introduce a unique approach employing spectral scalograms (SS) combined with the EfficientNet (ENB) architecture. The SS, obtained through continuous wavelet transform, reveals temporal and spectral information of EEG signals, aiding in the identification of transient features, aiding in accurate SZ classification. Experimental evaluation on a comprehensive dataset demonstrates the efficacy of our approach, achieving a five-fold mean cross-validation accuracy of 95.3% using CAE with XGB and 97% utilizing SS with the ENB0 model. Grad-CAM was then applied on ENB0 to highlight the time-frequency bands key to each decision, and SHAP on the CAE-XGB pipeline to rank the most essential EEG channels. These complementary views clarify both where and why the models work, paving the way for more transparent clinical EEG analysis.