Early disease detection using multiscale EEG graph fusion and bioinspired deep ensemble learning

Abstract Early detection of Alzheimer’s disease (AD) is essential for timely clinical intervention, yet remains challenging due to subtle neural changes in its initial stages. This study proposes a non-invasive EEG-based framework that integrates multiscale signal processing, dynamic brain connectivity analysis, nonlinear feature learning, and ensemble classification for improved diagnosis. Empirical Wavelet Transform (EWT) is employed to decompose nonstationary EEG signals into informative frequency subbands. Dynamic Functional Connectivity Graph Fusion (DFCG) is then used to model time-varying inter-channel synchronization patterns across multiple EEG rhythms. Discriminative representations are extracted using Nonlinear Manifold Statistical Feature Encoding (NMSFE), which preserves the intrinsic geometric structure of connectivity features. Finally, a Bio-Inspired Deep Ensemble Neural Network (BIDENN) performs classification by adaptively combining multiple deep learners to enhance robustness and generalization. Experimental evaluation on publicly available EEG datasets, using cross-validation and external validation protocols, demonstrates consistent performance with an accuracy of 96.2%, precision of 95.8%, recall of 96.5%, and F1-score of 96.1%. The results, supported by confusion matrix and ROC analysis, indicate that the proposed approach provides a reliable and efficient tool for early-stage Alzheimer’s disease detection.