Sleep-Derived Features From Multi-Night In-ear EEG Identify Patterns Linked To Mild Cognitive Impairment

INTRODUCTION: We investigated whether sleep features from multi-night, at-home in-ear EEG could distinguish mild cognitive impairment (MCI) from cognitively normal (CN) older adults, independent of demographic confounds. METHODS: Forty-three older adults (24 MCI, 19 CN) completed 361 overnight in-ear EEG recordings at home (mean: 4.44 and standard deviation: 2.29) nights per participant). Deep learning-derived embeddings were compared to bandpower features. We controlled for age and sex to identify sleep-specific features of MCI beyond demography. RESULTS: Deep embeddings outperformed bandpower features (F1: 0.76 vs. 0.67) in classifying MCI from CN. Multi-night aggregation improved AUROC from 0.61 to 0.77, reduced within-subject variability 1.6-fold, and enhanced group separability based on Cohen's d, improving from 0.65 to 0.96. While some features correlated with age, MCI discrimination remained significant after controlling for demographics. DISCUSSION: Multi-night in-ear EEG captures sleep features that reliably distinguish MCI from CN individuals beyond age-related effects, offering a scalable approach for at-home sleep monitoring.