Transformer Models Enable Accurate Age Prediction From Sleep Physiology

Biological age estimation, derived from physiological signatures such as brain activity, is emerging as a valuable biomarker for health and well-being. Discrepancies between biological and chronological age have been linked to multiple physical, mental, and cognitive health outcomes. However, current approaches primarily focus on MRI-based measurements, which are costly, challenging to obtain, and contraindicated for certain populations. This study explores polysomnographic (PSG) sleep signals, which capture activity from multiple physiological systems, as an accessible alternative for biological age prediction. Sleep serves as an ideal platform for age prediction due to its standardized data collection protocols, abundant public data resources, and the presence of well-documented age-related changes in sleep architecture. Additionally, the proliferation of consumer sleep monitoring tools offers potential for widespread application and longitudinal analysis. We trained transformer-based neural network models on over 10,000 nights of PSG data and performed rigorous internal and external validation. Our best models achieved age predictions with an absolute error of 5-10 years from just a single physiological time series input and were especially accurate with respect to certain stages of sleep (specifically, N2). Electroencephalography (EEG) signals were essential for capturing sleep architecture changes that correlate with age, while 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 beyond the training dataset remains a challenge (age prediction errors increase between internal validation and external data by at least 3 to 5 years). These findings show that noninvasive sleep-derived electrophysiological signals, particularly EEG, can rival MRI-based age prediction models in accuracy—while offering lower cost, greater accessibility, and broader applicability.