Explainable machine learning for Alzheimer's disease characterization using small-sample EEG data

Abstract

Alzheimer's disease (AD) is associated with progressive cognitive decline and altered brain functional activity, yet objective and interpretable electrophysiological indicators remain insufficiently established. Resting-state electroencephalography (EEG) offers a low-cost and clinically accessible candidate, provided that the analysis is validated at the subject level and remains interpretable. This study evaluated an interpretable resting-state EEG framework for distinguishing AD patients from healthy control (HC) subjects. A total of 63 participants (36 AD and 27 HC) from a publicly available dataset were included. Subject-level spectral and nonlinear complexity features were extracted from 19 preprocessed scalp channels; missing-value imputation, feature screening, redundancy pruning, scaling, and model fitting were carried out within each fold of leave-one-subject-out cross-validation. Four linear classifiers were compared, and Ridge Logistic regression was retained for out-of-fold SHAP interpretation because of its balanced hard-label performance and direct compatibility with Linear SHAP. Ridge Logistic regression achieved an exploratory AUC of 0.912 (accuracy = 0.857, sensitivity = 0.778, specificity = 0.963) under a non-nested validation design. Across 30 independently balanced epoch resamples, mean AUC was 0.887 ± 0.020; using all accepted epochs yielded AUC = 0.917. SHAP analysis indicated that the classifier drew jointly on posterior α activity, frontal and temporal θ power, the θ/α ratio, and slow/fast ratio features. A classifier-independent microstate analysis revealed reduced putative Class B occurrence, increased putative Class C and Class D duration, and six FDR-corrected off-diagonal transition differences in AD; the three temporal effects persisted across repeated K = 4 initializations and matched K = 3–6 solutions. These findings suggest that resting-state EEG can provide non-invasive and interpretable information about AD-related functional alterations, pending validation in larger, independent cohorts.