# Machine Learning–Based Sleep Electroencephalographic Brain Age Index and Dementia Risk: An Individual Participant Data Meta-Analysis

## TL;DR

A machine learning model using sleep EEG data can predict dementia risk, with higher brain age index linked to a 39% increased risk.

## Contribution

Introduces a machine learning–based sleep EEG brain age index as a novel digital marker for dementia risk.

## Key findings

- Each 10-year increase in BAI was associated with a 39% higher dementia risk after adjusting for age, sex, and other factors.
- The association remained significant after adjusting for comorbidities and apnea-hypopnea index scores.
- Findings were consistent across sex and age groups, suggesting broad applicability.

## Abstract

Is a higher brain age index (BAI) derived from sleep electroencephalography (EEG) using machine learning associated with a higher risk of future dementia in community-dwelling older adults?

In this individual participant data meta-analysis of 7105 adults from 5 longitudinal cohorts, every 10-year increase in BAI was associated with a 39% higher risk of incident dementia, independent of age, sex, apolipoprotein E ε4 status, and global cognition and comorbidities at the sleep study.

These findings suggest that sleep EEG-based BAI may serve as a promising early digital marker for dementia risk stratification.

This individual participant data meta-analysis explores the association between a machine learning–based sleep electroencephalography (EEG) brain age index and dementia risk among community-dwelling adults from 5 longitudinal cohorts.

Microstructures of sleep electroencephalography (EEG) are closely related to cognition and undergo age-dependent changes. However, their multidimensional nature makes them challenging to interpret using conventional approaches. The machine learning–based EEG brain age index (BAI) measures the deviation between sleep EEG-based brain age and chronological age.

To determine the association between sleep BAI and incident dementia in community-dwelling populations.

For this individual participant data (IPD) meta-analysis, sleep study data from 5 community-based longitudinal cohorts were pooled. These cohorts included the Multi-Ethnic Study of Atherosclerosis (MESA; 2010-2013), the Atherosclerosis Risk in Communities (ARIC) study (1987-1989), the Framingham Heart Study–Offspring Study (FHS-OS; 1995-1998), the Osteoporotic Fractures in Men Study (MrOS; 2003-2005), and the Study of Osteoporotic Fractures (SOF; 2002-2004).

Adults (aged ≥18 years) without dementia at the time of polysomnography were included.

The BAI was computed using interpretable machine learning, incorporating sleep EEG features extracted from central channels in overnight, home-based polysomnography. Fine-Gray models were used 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-analysis. Analyses were performed between March 2024 and September 2025.

Incident dementia or probable dementia was determined in each cohort, with death as a competing risk.

This meta-analysis included 7105 participants from the MESA (n = 1802; mean [SD] age, 69.3 [9.0] years; 956 females [53.1%]), ARIC (n = 1796; 62.5 [5.7] years; 918 females [51.1%]), FHS-OS (n = 617; 59.5 [8.9] years; 318 females [51.5%]), MrOS (n = 2639 males [100%]; 76.0 [5.3] years), and SOF (n = 251 females [100%]; 82.7 [2.9] years) cohorts. The median (IQR) time to dementia was 4.8 (4.2-5.6) years in the MESA cohort (n = 119 [6.6%]), 16.9 (14.9-19.8) years in the ARIC cohort (n = 354 [19.7%]), 13.1 (8.5-16.2) years in the FHS-OS cohort (n = 59 [9.6%]), 3.6 (1.3-7.1) years in the MrOS cohort (n = 470 [17.8%]), and 4.6 (4.2-5.2) years in the SOF cohort (n = 86 [34.3%]). Across the cohorts, each 10-year increase in BAI was associated with a 39% higher risk of incident dementia (hazard ratio [HR], 1.39 [95% CI, 1.21-1.59]; P < .001) after adjustment for covariates. These associations remained after additional adjustment for comorbidities and apnea-hypopnea index scores (HR, 1.31 [95% CI, 1.14-1.50]; P < .001) and apolipoprotein E ε4 (HR, 1.22 [95% CI, 1.02-1.45]; P = .03), and they were consistent across sex and age groups.

In this IPD meta-analysis, a higher sleep EEG-based BAI was associated with a higher risk of incident dementia. These findings highlight the need to evaluate the predictive value of the BAI as a noninvasive digital marker for early detection of dementia in community settings.