Sleep is crucial for health, yet accurately measuring it outside of laboratory settings is challenging. While polysomnography is the gold standard, its cost and complexity limit its use in large-scale studies. Wrist-worn accelerometers offer a more practical alternative, but existing sleep classification algorithms often rely on heuristic methods validated in small cohorts, limiting their generalizability. This study aimed to develop and validate a robust sleep classification model using accelerometer data from a large, diverse population and to investigate the association between objectively measured sleep parameters and mortality risk. The researchers hypothesized that a deep learning approach could improve the accuracy of sleep classification using accelerometer data and that insufficient sleep duration would be associated with increased mortality risk, regardless of sleep efficiency.

The literature review highlighted the limitations of existing sleep assessment methods. Self-reported sleep diaries are subjective and poorly correlated with objective measures. Polysomnography, while accurate, is impractical for large-scale studies. Wrist-worn accelerometers are a promising alternative but current algorithms often rely on handcrafted features and lack validation in diverse populations. The need for data-driven approaches like deep learning to improve sleep classification accuracy from accelerometer data was emphasized. Existing actigraphy-based sleep studies on large populations were also reviewed, highlighting their limitations and the potential for improvement using advanced machine learning techniques.

The study used a three-step pipeline. First, a self-supervised deep recurrent neural network (SleepNet) was developed to classify 30-second epochs of accelerometer data into wake, REM, and NREM sleep stages. SleepNet was pre-trained on a large dataset of unlabeled accelerometer data from the UK Biobank to learn features of human motion dynamics, then fine-tuned using labeled polysomnography data. Three different accelerometer devices were used (ActiGraph GT3X, Activpal X32, and GENEActiv Original), with data preprocessed to remove non-wear time and artifacts. Internal validation was performed using five-fold cross-validation across multiple cohorts, and external validation used independent datasets. Second, the performance of SleepNet was compared to other baseline models using metrics such as F1 score and Bland-Altman plots. Third, the validated SleepNet was applied to data from approximately 100,000 participants in the UK Biobank to investigate the association between sleep duration and sleep efficiency (calculated as the proportion of time asleep while in bed) and all-cause mortality. Cox proportional hazards regression was used to analyze the relationship between sleep parameters and mortality risk, adjusting for multiple potential confounders.

SleepNet demonstrated fair to moderate agreement with polysomnography in both internal and external validation. The model tended to underestimate REM sleep and short sleep durations while overestimating NREM sleep duration. In the UK Biobank analysis, short sleepers (≤6 hours) had a significantly increased risk of all-cause mortality compared to those with normal sleep duration (6-7.9 hours), with hazard ratios of 1.36 (95% CI: 1.18-1.58) for low sleep efficiency group and 1.29 (95% CI: 1.04-1.61) for high sleep efficiency group. This association was observed irrespective of sleep efficiency. The risk of all-cause mortality appeared to decrease linearly as sleep efficiency increased. Subgroup analyses revealed variations in sleep patterns across different demographic and lifestyle characteristics. For instance, older participants generally slept longer with higher efficiency, and females tended to have better sleep parameters compared to males. Individuals with self-reported insomnia symptoms displayed distinct sleep architecture. Descriptive statistics regarding sleep parameters across different subgroups (age, sex, ethnicity, physical activity, smoking, alcohol, education, deprivation, BMI, employment, self-rated health) were presented, showing significant differences in sleep durations and efficiency.

The findings support the use of accelerometer-based sleep classification using deep learning for large-scale epidemiological studies. The study's large sample size and longitudinal design increase the generalizability and robustness of the findings. The association between short sleep duration and increased mortality risk is consistent with previous research, reinforcing the importance of adequate sleep for overall health. The consistent finding across different sleep efficiency groups highlights that the duration of sleep is a critical factor for mortality, irrespective of sleep continuity. Further research is needed to explore the mechanisms underlying this association, and this study provides a valuable tool to investigate these relationships more thoroughly.

This study successfully demonstrated a deep learning approach for accurate sleep stage classification using accelerometer data. The analysis revealed a strong association between short sleep duration and increased mortality risk, regardless of sleep efficiency. Future research could focus on exploring the biological mechanisms driving this relationship and on further refining the sleep classification algorithm to improve its accuracy and generalizability to even more diverse populations.

The study's reliance on wrist-worn accelerometers might not capture all aspects of sleep architecture. The use of a proxy for time in bed in the UK Biobank analysis introduces potential inaccuracies. Some participants might have had undiagnosed sleep disorders, despite efforts to exclude individuals with known sleep problems based on self-reported information. The cross-sectional nature of the mortality analysis means that causality cannot be definitively established.