Achieving Accurate Ubiquitous Sleep Sensing with Consumer Wearable Activity Wristbands Using Multi-class Imbalanced Classification

• Published in: 2019 IEEE Intl Conf on Dependable, Autonomic and Secure Computing, Intl Conf on Pervasive Intelligence and Computing, Intl Conf on Cloud and Big Data Computing, Intl Conf on Cyber Science and Technology Congress (DASC/PiCom/CBDCom/CyberSciTech) pp. 768 - 775
• Main Authors: Liang, Zilu, Chapa Martell, Mario Alberto
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01-08-2019
• Subjects: Classification algorithms; data quality; Fitbit; Heart rate; machine learning; Machine learning algorithms; Sleep; Training; Vegetation; wearable
• DOI: 10.1109/DASC/PiCom/CBDCom/CyberSciTech.2019.00143

Consumer activity wristbands such as Fitbit provide an affordable method for ubiquitous sleep sensing in daily settings. These devices are also increasingly used in scientific studies as measurement tools of sleep outcomes. Nevertheless, the accuracy of Fitbit has raised wide concern. In this paper, we explore the feasibility of applying machine learning to improve the quality of Fitbit sleep data. The problem of interest was formulated into a multiclass imbalanced classification problem. We examined the performance of different combinations of seven machine learning algorithms and three resampling techniques. The preliminary results showed that the accuracy in detecting wakefulness and light sleep was improved by up to 43% and 44% respectively compared to the proprietary algorithm of Fitbit. Our future work will focus on improving the overall accuracy of the classification models in detecting all sleep stages.