Efficient EEG-based schizophrenia classification via a lightweight hybrid CNN-BiLSTM model

Schizophrenia (SCH) is a severe neuropsychiatric disorder characterized by abnormalities in perception, cognition, and behavioral functioning, making early and objective diagnosis a major challenge in clinical practice. Electroencephalography (EEG) has gained significant attention as a non-invasive and cost-effective modality for identifying neural biomarkers associated with SCH. In this study, a novel hybrid deep learning framework is proposed for automated SCH classification using raw EEG recordings. The proposed methodology employs a CBiL-Net framework that integrates one-dimensional Convolutional Neural Networks (1D-CNNs) and Bidirectional Long Short-Term Memory (Bi-LSTM), networks to simultaneously learn discriminative spatial features and complex temporal dependencies inherent in EEG signals. Initially, EEG data undergo preprocessing involving numerical standardization and elimination of invalid samples, followed by segmentation into fixed-length temporal sequences to preserve sequential neural dynamics. The stacked convolutional layers perform multi-level feature extraction, while bidirectional recurrent layers capture long-range contextual dependencies from both forward and backward temporal directions. Fully connected layers with dropout regularization are incorporated in the final classification stage. The proposed framework was evaluated on two benchmark EEG datasets for SCH, namely the IBIB PAN dataset and the Moscow dataset. Experimental results demonstrated strong generalization performance, achieving 89.15% classification accuracy, 83.78% precision, 89.86% F1-score, 96.88% sensitivity, and 81.54% specificity on the IBIB PAN dataset, while yielding significantly superior performance on the Moscow dataset with 99.60% accuracy, 97.66% precision, 98.81% F1-score, 100.00% sensitivity, and 98.92% specificity. Additionally, to assess robustness in a stricter evaluation setting, 5-fold stratified cross-validation was performed, during which the proposed model achieved an average accuracy of 61.80% and an AUC of 68.03%.