Sleep disturbances are linked to age and mortality. Polysomnography (PSG), the gold standard for sleep diagnostics, provides detailed physiological data during sleep. Currently, PSG analysis focuses on basic metrics like sleep stages and apnea-hypopnea index (AHI), scored manually and subject to variability. While AHI is crucial, other aspects of sleep also influence health outcomes. Studies have shown associations between all-cause mortality and increased arousal burden, decreased sleep efficiency, and reduced REM sleep. Deep learning offers a potential solution by analyzing the entirety of PSG data, going beyond the limitations of manual scoring. Age is a strong predictor of morbidity and mortality, with sleep architecture changing with age. A previous study demonstrated that an age estimation model using EEG data was associated with mortality; this study aims to improve upon this by using deep learning models analyzing a wider range of PSG signals to model age as a proxy for mortality risk, interpret learned features, and examine associations between age estimation error (AEE) and lifestyle factors and mortality.

Existing literature highlights the relationship between sleep disturbances and mortality. Manual scoring of PSGs, while standard practice, is time-consuming and prone to error. Deep learning methods have shown promise in analyzing PSG data, offering more detailed insights. Prior work has demonstrated the use of EEG-based features to estimate brain age, with the resulting error correlated with mortality. This study builds on this research by using deep learning and a wider range of PSG signals to improve age estimation accuracy and explore the relationship between age estimation error and mortality risk more comprehensively.

This study utilized a large dataset of 13,332 PSGs from seven cohorts (STAGES, Stanford Sleep Cohort, Wisconsin Sleep Cohort, SHHS, SofM Sleep Study, CFS, and HomePAP). The data were split into training (2500 PSGs), validation (200 PSGs), and test (10,699 PSGs) sets. Deep neural networks were trained to estimate age from various combinations of PSG signals (EEG, EOG, EMG, ECG, respiratory). An ensemble model averaging the predictions from several individual models was also developed. Model performance was evaluated using mean absolute error (MAE). Gradient SHAP was used to interpret the models, attributing relevance scores to input signal samples. The association between AEE and mortality was analyzed using Cox proportional hazards models in a subset of the data (9386 subjects with mortality data from SHHS, MRoS, and WSC cohorts), adjusting for demographics, lifestyle, and health covariates. Sensitivity analyses were conducted to examine the effects of hypertension and sleep apnea.

The ensemble model achieved a MAE of 8.16 ± 3.75 years in the test set. Basic sleep measures resulted in a significantly higher MAE (12.5 ± 4.06 years). Gradient SHAP analysis revealed that the model's age estimates were influenced by clinically relevant waveforms, such as sleep-stage transitions and apnea. A 10-year increase in AEE was associated with a 29% increased risk of all-cause mortality (HR = 1.29, 95% CI: 1.20–1.39) and a 40% increased risk of cardiovascular mortality (HR = 1.40, 95% CI: 1.21–1.62). For a 60-year-old, a difference of ±10 years in AEE translated to an 8.7-year difference in life expectancy. Sensitivity analyses showed that the association between AEE and mortality remained significant even after excluding subjects with hypertension or sleep apnea, although the effect size was reduced. The model showed some bias towards underestimating age in older subjects and overestimating age in younger subjects. Night-to-night variability in AEE was also observed but not significant.

This study demonstrates the potential of deep learning to extract clinically relevant information from PSGs that predicts mortality beyond basic sleep metrics. The model's interpretability using SHAP analysis validates its reliance on known physiological patterns associated with aging. The strong association between AEE and mortality highlights its potential as a biomarker for overall health and life expectancy. The findings suggest that PSG data holds much richer information than conventionally extracted, warranting further exploration of its potential in predicting health outcomes.

Deep learning models can accurately estimate age from PSG data, with the age estimation error significantly associated with mortality risk. This approach provides a novel, non-invasive, and relatively inexpensive method for assessing health and predicting life expectancy. Future research could focus on refining the models, validating them in diverse populations, and exploring the underlying biological mechanisms driving the observed associations.

The study's limitations include potential biases due to the non-uniform age distribution in the dataset, the reliance on a single PSG recording per subject, and variations in PSG recording equipment and procedures across different cohorts. The interpretability analysis using SHAP assumes linearity, which might not fully capture the complexities of the deep learning model. Future studies should address these limitations to enhance the generalizability and robustness of the findings.