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

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Alzheimer's disease (AD) is a leading cause of dementia, necessitating improved diagnostic tools.
• Electroencephalography (EEG) is a safe, non-invasive, and cost-effective method for neurological assessment.
• Current EEG-based AD classification methods struggle with causal relationship analysis and optimal feature selection.

Purpose of the Study:

• To develop and validate EENet-RLA, a deep learning framework integrating dynamical system theory for accurate AD classification using EEG.
• To introduce a novel, causality-driven EEG channel selection strategy based on embedding entropy (EE) for enhanced feature screening.
• To demonstrate the effectiveness of this approach in small-sample settings for AD characterization.

Main Methods:

• The EENet-RLA framework utilizes a two-stage process: feature extraction and EEG classification.
• Causal, stability-driven EEG channel selection is performed using embedding entropy (EE), bootstrap resampling, and minimum connectivity thresholds.
• Deep learning models (ResNet, LSTM) extract spatial and temporal features, fused by a Multi-Head Attention mechanism for classification.

Main Results:

• The proposed framework achieved 98.54% segment-level accuracy and perfect individual-level performance on the BrainLat EEG dataset.
• Causality-informed feature selection proved effective in identifying discriminative EEG channels, particularly in limited sample scenarios.
• The method demonstrated high accuracy while streamlining the analytical process for EEG-based AD diagnosis.

Conclusions:

• EENet-RLA offers a highly accurate and interpretable method for AD classification using EEG, driven by causal feature selection.
• The framework highlights the potential of embedding entropy for identifying informative EEG channels in neurological studies.
• This causality-based approach shows promise for AD characterization and can be adapted for other neurological conditions with similar signal properties.