Enhancing Multiclass Dementia Detection with EEG Signals: A Feature-Driven LSTM Approach

Early and accurate differentiation of dementia subtypes remains a major challenge in clinical neurophysiology and is crucial for timely intervention and personalized care. This study introduces a novel EEG-based deep learning framework for classifying Alzheimer's Disease (AD), Mild Cognitive Impairment (MCI), Subjective Cognitive Decline (SCD), and Healthy Controls (HC). Resting-state EEG data from 139 participants (>45 years) were analyzed through spectral and connectivity features, focusing on Theta-Beta activity linked to cognitive dysfunction. A feature-enriched Long Short-Term Memory (LSTM) model integrated Pearson's correlation-based connectivity measures with optimized dimensionality reduction techniques to enhance class separability and computational efficiency. The proposed PCC-PCA-LSTM model achieved 99.47% classification accuracy and an AUC of 1.00 across all classes, significantly outperforming conventional EEG-based diagnostic approaches. These findings demonstrate the model's potential as a clinically viable and scalable tool for early detection, continuous monitoring, and decision support in dementia care, bridging the gap between neurophysiological insight and practical diagnostic utility.