0426 Sleep Prediction Algorithm Based On Deep Learning Technology

Abstract

Introduction: Clinicians believe that maladaptive sleep-related factors such as inactivity during daytime, excessive time spent in bed, poor exposure to light, and high stress level contribute development of chronic insomnia. However, there's paucity of evidences how exactly one's behavior during daytime affect one's sleep at that night. Traditional assessment methods, such as polysomnography, sleep diary, and actigraphy, cannot integrate sleep outcome with sleep-related factors by themselves. That is, prediction of sleep is wholly dependent to a researcher. We need to find the way which can integrate sleep-related factor and predict one's sleep automatically. In this study, we tried to develop an algorithm that can predict sleep quality by using deep learning technology. Methods: Sixty-nine healthy participants were enrolled to our study. We measured their sleep and sleep-related factors including daytime activity, exposure to light, and heart rate variability by using actigraphy and heart rate sensor (ActiGraph GT3X® and Polar H7®) for two weeks. The dependent variable of the analysis was sleep quality (sleep efficiency ≥90%). We used raw data for predicting sleep with deep learning technology, multi-input one-dimensional convolutional neural network (1D CNN). Additionally, we used traditional analysis and machine learning technology to aware which factors affect sleep quality. For this, we used subdivided data into 3 part: wake-up to noon (P1), noon to 18:00 (P2), and 18:00 to bedtime (P3). We performed logistic regression analysis for traditional analysis, and random forest(RF) as machine learning technology. Results: P1 vigorous activity, P1, P2, P3 exposure to light, and P1, P2, P3 exposure to outside light showed significant correlations with sleep quality in weighted logistic regression. The effect of other variables on sleep quality was below significance level. Logistic regression model's accuracy rate was 79.2%, RF model's was 83.6%, 1D CNN model's was 87.2% without sleep diary information. Conclusion: The algorithm made by 1D-CNN model can predict sleep quality accurately which is comparable to the traditional logistic regression and machine learning technology. This result can enable more objective and accurate sleep prediction by automation of integration and interpretation of the data.