Enriched resting-state EEG prediction of cognitive decline in prodromal Alzheimer's disease: a machine-learning approach

Objective: We evaluated the accuracy of standard machine learning (ML) algorithms in predicting 1-year cognitive decline in Alzheimer's disease patients with mild cognitive impairment (ADMCI) using resting-state electroencephalographic (rsEEG) biomarkers enriched with APOE genotype, sex, age, and educational attainment data. Methods: The study analyzed datasets from 63 ADMCI patients obtained from an international archive. The ML algorithms included Simple Logistic Regression, Model Trees, Logistic Regression, K-nearest neighbor, and Support Vector Machine. Input features comprised lobar rsEEG source activities across delta (<4 Hz) to alpha (≈10-12 Hz) bands, cerebrospinal fluid (CSF Aβ1-42/p-tau), and structural magnetic resonance imaging (sMRI) biomarkers. Cognitive decline was assessed over a 1-year follow-up ("stable" vs. "decliner") based on Mini-Mental State Examination (MMSE) scores. Results: The four independent ML algorithms accurately predicted changes in the MMSE score over a 1-year follow-up, with accuracies of 77-78% in ADMCI participants aged ≥ 70 years and 74-77% in those aged < 70 years. Conclusions and significance: These findings suggest that rsEEG biomarkers in ADMCI patients may not only reveal underlying pathophysiological mechanisms affecting cortical arousal and vigilance but also hold predictive value for cognitive outcomes.