Transformer Models Enable Accurate Age Prediction From Sleep Physiology

Abstract Biological age estimation based on physiological signatures such as brain activity has emerged as a valuable biomarker; the gap between an individual’s predicted biological age and their chronological age is increasingly being linked to diverse health and cognitive outcomes. This study explores polysomnographic (PSG) recordings of sleep to evaluate how well diverse physiological signals—including EEG, ECG, EOG, and EMG—can support accurate age estimation, both individually and in combination. PSG serves as an ideal platform for age estimation due to its standardized data collection protocols, abundant public data resources, and its capture of well-documented age-related changes in sleep architecture that also relate to chronic health outcomes. Accordingly, we trained transformer-based neural network models on over 10,000 nights of public PSG data and performed rigorous internal and external validation. The best models achieved age estimates with an absolute error of 5-10 years on held out data from a single channel of electroencephalography (EEG). Interestingly, accuracy in these models was stage-dependent, with higher concentrations of N2 sleep in a given input sequence yielding more precise estimates than sequences dominated by other sleep stages. Age overestimation was associated with worse depression and cognitive performance outcomes. Electrocardiogram (ECG) signals, although less accurate overall, tended to overestimate age in association with health conditions such as elevated blood pressure, higher body mass index, and sleep apnea. Despite strong performance, generalization to external data remains a challenge (age estimation errors increase between internal validation and external data by at least 3 to 5 years). These findings show that non-invasive sleep-derived electrophysiological signals, particularly EEG, can support age estimation with accuracy comparable to functional MRI-based (5 to 11 years estimation error), yet at a fraction of the cost. The proliferation of consumer sleep monitoring devices, coupled with these highly accurate electrophysiological models, makes population-scale, at-home assessment of biological brain aging increasingly feasible.1 Competing Interest Statement The authors have declared no competing interest. Funding Statement This study was funded by an internal Research & Development grant by the Johns Hopkins Applied Physics Laboratory. No other funding, payment, or services were received in support of this work. Author Declarations I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained. Yes The details of the IRB/oversight body that provided approval or exemption for the research described are given below: Study used ONLY openly available human data hosted at the National Sleep Research Resource (NSRR) at sleepdata.org I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals. Yes I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance). Yes I have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable. Yes Footnotes will.coon{at}jhuapl.edu We updated various parts of this manuscript. ↵1 Model checkpoints along with training, evaluation and analysis scripts are available: https://osf.io/qw2d5 Data Availability Study used ONLY openly available human data hosted at the National Sleep Research Resource (NSRR) at sleepdata.org