Residual Reinforcement-Enhanced Deep Backbone Basis Network For Multimodal Eeg-Driven Schizophrenia Prediction

Abstract

Deep learning models have shown promising results in schizophrenia prediction using EEG, MRI, and fMRI data, yet most existing models relied on small and single-site datasets, limited modalities, and non-interpretable architectures. High-accuracy studies often used fewer than 50 subjects, which led to overfitting and weak clinical translation. In addition, most neural models were designed for binary diagnosis rather than long-term disease prediction or treatment guidance, and few incorporated explainability or adaptive learning. Existing models operated in constrained conditions where performance dropped significantly during cross-site validation, proving that generalizability remained a critical issue. EEG-based works also processed signals in isolated temporal or spectral space, without accounting for dynamic brain connectivity. Clinical adoption further remained restricted because many high-performing networks behaved as black boxes, preventing psychiatrists from understanding model-inferred biomarkers. The study proposed a Residual Reinforcement-Enhanced Deep Backbone Basis Function Network (RR-DBBFN), which combined a deep backbone architecture with reinforcement-driven residual learning. EEG signals from the Moscow and Warsaw public schizophrenia datasets were preprocessed using a Mutation-boosted Archimedes Optimization (MAO) filter. Multivariate empirical mode decomposition, entropy signatures, and Markov Transition Field (MTF) encodings were extracted and passed through a hybrid spatial–temporal graph module. A reinforcement unit dynamically updated residual layers according to reward signals based on validation accuracy, allowing the network to self-adapt during deeper training cycles. SHAP-based explainability was integrated to identify discriminative brain rhythms. The experimental evaluation showed that the proposed model achieved 98.41% accuracy, 98.22% precision, and 98.64% recall on the Moscow EEG dataset that performs better than the best existing method by 2.39%. The proposed model also achieved an accuracy of 97.88% on the Warsaw dataset and this shows an improvement of 4.96% over the strongest baseline.