1008 Applications of Machine Learning to Improve Time in Bed Detection by Leg-Worn Actigraphy

• Published in: Sleep (New York, N.Y.) Vol. 42; no. Supplement_1; pp. A405 - A406
• Main Authors: Ferree, Thomas, Moynihan, Martin, Gozani, Shai
• Format: Journal Article
• Language: English
• Published: Westchester Oxford University Press 13.04.2019
• Subjects: Algorithms; Decision trees; Machine learning; Sleep
• ISSN: 0161-8105; 1550-9109; 1550-9109
• DOI: 10.1093/sleep/zsz067.1005

Introduction Actigraphy is used to analyze human sleep. Among the metrics reported, Time in Bed (TIB) is essential as it sets the temporal frame for sleep/wake classification. Determining TIB automatically in real-world settings is challenged by behaviors that mimic sleep. We report here a novel algorithm that uses three signal attributes available in leg-worn actigraphy. We use machine learning techniques to optimize algorithm parameters. Methods This work is based upon data from 42 subjects collected over 155 nights. Subjects wore the actigraphy device afternoon-to-morning for an average 710 minutes and recorded TIB Start and TIB End in their diary. We developed a three-tiered algorithm. Tier one uses body orientation and step counting to identify upright active intervals. We define the time between these intervals as segments. Tier two is a Bayesian classifier that combines activity, orientation, and time to compute the probability a segment is within TIB. Tier three is a decision tree for gathering sequences of segments that comprise TIB. We fit the Bayesian classifier and applied simulated annealing to the decision tree to minimize the errors in TIB. Results The 155 nights of data included 843 segments. The Bayesian model alone classified 92.9% of segments correctly (Sensitivity = 93.6%, Specificity = 92.5%). After optimizing the decision tree, the complete algorithm classified 97.3% of segments correctly (Sensitivity = 94.0%, Specificity = 98.9%). Accuracy increased partly because because the tree can include segments within TIB that have low probability based on their attributes alone. Visual inspection of residual errors showed that restful behaviors right before sleep remain a challenge. Conclusion In this study we applied machine learning to develop algorithms that combine activity, orientation, and time to estimate TIB for leg-worn actigraphy. The Bayesian model and decision tree were fit to the training population and yielded 97% accuracy for classifying segments within TIB. Future work will explore new segment features, different classifiers, and using within-night and within-subject data to further improve estimates of TIB. Support (If Any) This study was funded by NeuroMetrix, Inc.