Predicting future risk of developing cognitive impairment using ambulatory sleep EEG: Integrating univariate analysis and multivariate information theory approach

Background: Early identification of individuals at risk for cognitive impairment is crucial, as the preclinical phase offers an opportunity for interventions to slow disease progression and improve outcomes. Objective: While sleep electroencephalography (EEG) has shown significant promise in detecting cognitive impairment, this study aims to 1) develop and validate overnight EEG biomarkers for the prediction of future cognitive impairment risk, 2) assess their predictive performance within 5 years, and 3) explore the feasibility of using wearable, low-density EEG devices for convenient at-home monitoring. Methods: Overnight polysomnography was performed on 281 cognitively normal women in the Study of Osteoporotic Fractures (SOF). Cognitive reassessments were conducted approximately five years later. Features such as relative EEG power across different frequency bands and channel interactions, quantified using generalized mutual information measures, were extracted and used as inputs for machine learning models. Binary classification models distinguished participants who developed cognitive impairment from those who remained cognitively normal. Optimal feature subsets and frequency bands for classiffiation were identifed, with additional analyses testing the contribution of demographic data, sleep macrostructure, and APOE genotype. Results: The optimal model, utilizing univariate and multivariate EEG features, achieved an AUC of 0.76. Features from the N3 sleep stage and gamma band exhibited the largest effect sizes. Adding demographics, sleep macrostructure, and APOE genotype did not enhance performance. Conclusions: Overnight EEG analyses demonstrate a promising, cost-effective approach for early cognitive impairment risk assessment. Larger studies with more diverse populations are required to validate and expand these findings in diverse populations.