A deep learning-derived brain health score from overnight EEG predicted lower mortality (HR 0.65-0.69 per SD increase, p<0.0001) and outperformed classical models across diverse cohorts.
Observational (n=42,483)
Yes
Does a deep learning-derived brain health score from overnight EEG predict cognition, disease risk, and mortality across diverse populations?
A deep learning-derived brain health score from a single overnight EEG channel can generalize across diverse cohorts to predict cognition, disease risk, and mortality.
Effect estimate: HR 0.65-0.69 per SD increase
p-value: p=<0.0001
Abstract Introduction Sleep EEG encodes rich information about brain and systemic health, yet most predictive models rely on hand-engineered features. We developed an end-to-end deep learning framework that learns a data-driven latent representation of overnight EEG and derives a single “brain health score.” We evaluated performance across three large population cohorts and conducted a fully independent external validation in a BIDMC clinical sample. Methods We analyzed 36,000 PSGs from six cohorts (FHS, MESA, MrOS, SOF, KoGES, MGH) to train a multi-task model using the C4–M1 EEG channel as 1D time series and 2D spectrograms. The encoder generated a 1024-dimensional latent space; downstream heads predicted cognitive test scores (e.g., fluid and crystallized intelligence), disease status (e.g. dementia, major depression), and a distilled brain health score. Training used cross-validated splits with strict subject-level separation. Independent evaluation included: (1) held-out population cohorts and (2) a new BIDMC clinical cohort (N≈6,483) containing all-cause mortality information. Performance metrics included correlation for cognition, ROC-AUC for disease, and Cox models for mortality. Results Across cross-validated training cohorts, the deep learning–derived score outperformed demographic baselines and classical machine-learning models using expert-defined EEG features. Cognitive correlations reached R ≈ 0.35–0.40; disease classification AUCs ranged 0.65–0.75. In mortality analyses (MrOS, MGH), each SD increase in the score predicted 31–35% lower hazard (HR 0.65–0.69, p 0.0001). In the independent BIDMC validation cohort, the score generalized without retraining; a 1-SD increase predicted a 24% lower age-adjusted mortality rate (p 0.0001). Effect sizes were comparable to or exceeded those observed in development cohorts, demonstrating transportability to real-world clinical EEG. Conclusion A single overnight EEG, analyzed via end-to-end deep learning, yields a generalizable biomarker of brain health that predicts cognition, disease risk, and mortality across diverse populations. Independent validation at BIDMC confirms performance in a clinically heterogeneous setting, supporting its potential for scalable deployment in research and clinical practice. Support (if any) NIH grants R01AG073410 and R01HL161253.
Ganglberger et al. (Fri,) conducted a observational in Brain health and mortality risk (n=42,483). Deep learning-derived brain health score from overnight EEG vs. Demographic baselines and classical machine-learning models was evaluated on All-cause mortality (HR 0.65-0.69 per SD increase, p=<0.0001). A deep learning-derived brain health score from overnight EEG predicted lower mortality (HR 0.65-0.69 per SD increase, p<0.0001) and outperformed classical models across diverse cohorts.