CM-EEGNet: Multiscale Contrastive Learning with Prototype-Based Channel Contribution Analysis for Dementia Diagnosis

Abstract

Electroencephalography (EEG) is a cost-effective and noninvasive modality for identifying neurophysiological alterations associated with Alzheimer’s disease (AD). To robustly extract disease-specific features, this study presents contrastive multiscale EEGNet (CM-EEGNet), a domain-specific deep learning framework that systematically integrates multi-scale temporal convolutions with supervised contrastive learning. By explicitly coupling parallel temporal convolutions of varying kernel sizes—calibrated to capture both spectral slowing (4–12 Hz) and transient dynamics—with a joint contrastive and cross-entropy optimization objective, the model effectively maximizes interclass separability within the latent representation space. Furthermore, we introduce channel prototype attribution (CPA), which causally evaluates channel-wise contributions based on shifts in contrastive evidence scores following channel occlusion. Evaluated under a strictly subject-independent protocol, CM-EEGNet achieved a segment-level accuracy of 96.26% for AD classification, significantly outperforming conventional baselines. Moreover, CPA highlighted the occipital (O2) and parietal (P4) channels as key AD discriminators, where occlusion of the top three channels reduced accuracy to 77.11%. This specialized integration establishes a highly discriminative and interpretable framework for EEG-based diagnosis, demonstrating both robust classification performance and clinical transparency.