EEG-based schizophrenia detection using handcrafted biomarkers and a TOA-optimized hybrid multi-branch CNN–Transformer framework

Schizophrenia is a chronic psychiatric disorder for which electroencephalography (EEG) offers a low-cost, non-invasive window into abnormal neural dynamics. However, many EEG-based computer-aided diagnosis (CAD) pipelines still rely on a single feature representation, heuristic tuning of model components, and window-level evaluation that may not translate to clinically meaningful subject-level decisions. We therefore developed a hybrid, calibration-aware CAD framework that fuses multi-branch deep embeddings with interpretable handcrafted EEG biomarkers and jointly optimizes feature selection and classifier hyperparameters via a Treble Opposite Algorithm with feature-search hybridization (TOA-FSH). Resting-state multi-channel EEG from two public cohorts-MSU-SZ (84 adolescents; 45 schizophrenia/39 healthy controls) and RepOD-SZ (28 adults; 14 schizophrenia/14 healthy controls)-was preprocessed under strict subject-wise protocols and converted into continuous wavelet transform (CWT) scalogram images. Deep features from EfficientNetV2-S, MobileNetV3-Large, ResNet-50, and a Swin Transformer branch were concatenated with handcrafted spectral, complexity, and connectivity features. TOA-FSH selected a sparse hybrid subset and tuned a k-nearest neighbor classifier, followed by probability calibration and subject-level fusion. On MSU-SZ, the final model achieved approximately 97.6% window-level balanced accuracy (BAC) and 100% subject-level BAC. When trained on MSU-SZ and tested without retraining on the unseen RepOD-SZ cohort, it again achieved approximately 97.6% window-level BAC and 100% subject-level BAC, indicating strong cross-cohort generalization under a harmonized preprocessing and evaluation scheme. Ablation analysis further showed incremental gains from ensembling, TOA-FSH-based selection, and handcrafted-feature fusion beyond single-backbone baselines. These findings indicate that calibration-aware hybrid modeling can provide reliable and transferable subject-level EEG markers for clinician-guided schizophrenia screening and assessment.