An Interpretable EEG Machine Learning Prototype for Alzheimer’s Disease Classification

Abstract

Resting-state electroencephalography (EEG) can capture the slowing of neural oscillations associated with Alzheimer’s disease (AD), but many machine-learning studies remain difficult to inspect, reproduce, or test. This study developed an interpretable, subject-level AD versus healthy-control classifier from the dataset OpenNeuro ds004504 and deployed it as a public research prototype. Preprocessed eyes-closed EEG recordings from 36 people with AD and 29 healthy controls were analysed. Welch power spectral density estimates were used to generate absolute and relative bandpower summaries, theta/alpha and delta/alpha ratios, a slow/fast ratio, and signal-level descriptors. Five-fold stratified cross-validation compared two Random Forest configurations, EEG Logistic Regression, an age-only baseline, and an EEG-plus-age model. EEG Logistic Regression gave the strongest mean fold-wise performance: accuracy 0.846, balanced accuracy 0.848, F1 score 0.849, and ROC-AUC 0.948. Age alone was near chance, while adding age did not improve the EEG-only model. The feature pattern was consistent with AD-related EEG slowing, including higher slow-wave-related ratios and lower relative alpha power. The trained pipeline was deployed as a Gradio application on Hugging Face Spaces. The result is a reproducible research and educational prototype, not a clinical diagnostic device, and requires external validation before any clinical interpretation.