Dementia Risk and Machine Learning-Derived Brain Age Index from Sleep Electroencephalography: A Pooled Cohort Analysis of Over 7,000 Individuals Across Five Community Cohorts

Abstract

Importance: Sleep electroencephalographic (EEG) microstructures are closely related to cognition and undergo age-dependent changes. However, their multidimensional nature makes them challenging to interpret using conventional approaches. Machine learning-computed EEG brain age index (BAI) represents the difference between the sleep EEG-based brain age and chronological age, quantifying deviations in sleep EEG microstructures from normative patterns. Objective: To determine the association between sleep BAI and incident dementia in community-dwelling populations. Design: Five individual cohorts and random-effects meta-analysis. Setting: This study pooled data from five community-based, methodologically consistent, longitudinal cohorts: MESA, ARIC, FHS-OS, MrOS, and SOF. We used Fine-Gray models to assess the association between BAI and incident dementia within each cohort, accounting for death as a competing risk. Cohort-specific estimates were then pooled using random-effects meta-analyses. Participants: 7,071 participants (MESA 54-94 years old, ARIC 52-75, FHS-OS 40-81, MrOS 67-96, SOF 79-93) without dementia at the time of polysomnography were included. Exposure: The sleep EEG-based BAI was computed using interpretable machine learning, incorporating 13 age-dependent features extracted from central EEG channels in overnight, home-based sleep polysomnography. Main Outcomes: and Measures Incident dementia or probable dementia was determined in each cohort, with death as a competing risk. Results: Across the five cohorts, dementia incidence ranged from 6.6% to 34.3% over a median follow-up of 3.5 to 17.0 years. Across cohorts, each 10-year increase in BAI was associated with a 39% increased risk of incident dementia (hazard ratio: 1.39 [95% confidence interval=1.21-1.59], p<0.001) after adjustment for age, sex, race, education, body mass index, current smoking, sleep medications, and physical activity level. The top feature underlying BAI was waveform kurtosis in N2 with a negative association with incident dementia (p<0.001). The associations remained after additional adjustment for multiple comorbidities, APOE e4 status, and apnea-hypopnea index, and were consistent across sex and age groups. Conclusions and Relevance: A higher sleep EEG-based BAI was associated with a higher risk of incident dementia across five community-based longitudinal cohorts. Future studies are warranted to evaluate the predictive value of BAI as a non-invasive digital biomarker for the early detection of dementia in community settings.