Predicting Sleep and Sleep Stage in Children Using Actigraphy and Heartrate via a Long Short-Term Memory Deep...

Area of Science:

• Biomedical Engineering
• Machine Learning in Healthcare
• Pediatric Sleep Medicine

Background:

• Traditional actigraphy algorithms have limitations in accurately detecting wakefulness and predicting sleep stages in children.
• Consumer wearables offer a more accessible method for sleep monitoring compared to laboratory polysomnography.

Purpose of the Study:

• To evaluate the agreement of Long Short-Term Memory (LSTM) algorithm's sleep estimates with polysomnography (PSG) in children.
• To assess the performance of LSTM using actigraphy and heart rate (HR) data from both research-grade and consumer wearables.

Main Methods:

• Utilized Long Short-Term Memory (LSTM), logistic regression, and random forest models on actigraphy and HR data from 238 children (5-12 years).
• Compared LSTM's sleep/wake and sleep stage predictions against criterion polysomnography (PSG) using 10-fold cross-validation.
• Assessed performance using sensitivity, specificity, and accuracy metrics.

Main Results:

• LSTM significantly outperformed traditional methods, achieving 94.1-95.1% accuracy for sleep/wake classification.
• LSTM demonstrated high sensitivity (94.9-95.9%) and improved specificity (84.5-89.6%) compared to older algorithms.
• Incorporating heart rate data enhanced sleep stage prediction but did not improve binary sleep/wake detection.

Conclusions:

• LSTM shows significant promise for accurate sleep and sleep staging prediction using actigraphy data in pediatric populations.
• The integration of heart rate data holds potential for refining sleep stage prediction accuracy.
• This approach could enhance the utility of wearable devices for pediatric sleep assessment.