Multi-view EEG schizophrenia detection using novel scattering-based patch encoding and a deep attention mechanism

Abstract

Schizophrenia (SZ) is an intricate psychiatric disorder characterized by hallucinations, delusions, and disorganized cognition. Early diagnosis and intervention are crucial for improved prognosis. This paper proposes a deep-learning model for diagnosing SZ using multi-channel electroencephalogram (EEG) signals. In the pipeline, we extract rhythms (delta, theta, alpha, beta, and gamma) from multichannel EEG signals for multi-view operation. Features are extracted from each multichannel EEG rhythm (MCER) through novel two-dimensional wavelet scattering-based patch encoders and self-attention mechanisms, encapsulated to form an architecture similar to that of a vision transformer. The features computed from each MCER are concatenated and then binary classified for the detection of SZ. The proposed architecture is studied on two widely used public datasets, namely, the IBIB-PAN and Mental Health Research Center (MHRC) datasets. Several experiments have been carried out for a comprehensive evaluation of the model, including analyses across different non-overlapping time windows, oscillatory levels, statistically validated multi-view analyses (p<0.05), various optimizers, and ablation studies. The performance of the model is validated using multiple evaluation strategies, including ten-fold cross-validation, percentage-based train–test–validation splitting, and the leave-one-subject-out (LOSO) validation approach. The model achieved accuracies of 98.04% on the IBIB dataset and 99.07% on the MHRC dataset using ten-fold cross-validation. Furthermore, t-distributed stochastic neighbor embedding (t-SNE) is used for feature visualization, while class-wise attention maps across different MCERs illustrate the attention distribution of the proposed framework. The proposed architecture has been compared with existing state-of-the-art methods and shows strong potential for SZ detection. © 2026 Elsevier B.V.