Towards proactively improving sleep: machine learning and wearable device data forecast sleep efficiency 4-8 hours...

Area of Science:

• Computational physiology
• Machine learning in healthcare
• Sleep science

Background:

• Wearable devices offer sleep tracking but lack predictive capabilities for proactive sleep improvement.
• Forecasting sleep parameters before sleep onset is crucial for preventing poor sleep.
• Existing models do not leverage accelerometer data for pre-sleep prediction.

Purpose of the Study:

• To develop and validate machine learning models for forecasting low sleep efficiency prior to sleep onset.
• To identify specific pre-sleep physical activity patterns associated with sleep efficiency.
• To explore the relationship between sleep variability and subsequent sleep efficiency.

Main Methods:

• Utilized accelerometer data from 80,811 UK Biobank participants.
• Developed predictive models using gradient boosting (CatBoost) and deep learning (CNN-LSTM).
• Performed cross-validation and functional data analyses to assess predictive performance and activity associations.

Main Results:

• Both CatBoost and CNN-LSTM models achieved excellent predictive performance (AUCs > 0.90).
• U-shaped relationships were found between pre-sleep activity and sleep efficiency.
• Higher activity within 4 hours before sleep onset was detrimental, while 4-6 hours prior was beneficial.

Conclusions:

• Accurate prediction of sleep efficiency is feasible hours before bedtime using wearable accelerometer data.
• Specific pre-sleep activity timing (4-6 hours prior) can be targeted for interventions to improve sleep.
• Machine learning models can enable proactive, personalized sleep health management.