EENet-RLA: An Explainable Prediction Learning Framework for Alzheimer's Disease Classification from EEG Signals

Alzheimer's disease (AD) is a prevalent neurodegenerative disorder affecting millions worldwide. Electroencephalography (EEG), a non-invasive, cost-effective, and safe diagnostic tool, is widely used for detecting neurological conditions. Existing EEG-based classification methods for AD diagnosis have limitations, particularly in adequately considering causal relationships between channels and implementing optimal feature selection, creating a need for highly interpretable feature screening mechanisms. This study presents EENet-RLA, a framework that integrates dynamical system theory with deep learning for AD classification, validated on the BrainLat EEG dataset. The framework operates in two stages, feature extraction and EEG classification, with the deep learning architecture serving primarily as a feature mapping and representation extractor. The core methodological contribution lies in the causal, stability-driven EEG channel selection strategy based on embedding entropy (EE), which quantifies nonlinear directional interactions between EEG channels. This strategy combines bootstrap resampling, multiple random seeds, and minimum connectivity thresholds to identify reproducible, informative channels under limited sample conditions. For classification, spatial and temporal EEG features are extracted using ResNet and LSTM respectively, then fused via a Multi-Head Attention mechanism to capture discriminative patterns. The proposed approach achieves 98.54% segment-level classification accuracy and perfect individual-level performance, demonstrating the discriminative potential of causality-informed feature selection in small-sample settings. While ensuring high accuracy, the method streamlines the analytical process and demonstrates the feasibility of causal-based EEG channel selection in AD characterization, with potential applicability to studying other neurological conditions with similar signal characteristics.