Machine Learning-based Health-related Behavior Detection using Wearables

Zhang, Shibo

doi:https://doi.org/10.21985/n2-r1wc-2y92

Work

Machine Learning-based Health-related Behavior Detection using Wearables

Public

Download PDF

Wearable-based human activity recognition is well-studied in the machine learning and pervasive computing community. A large corpus of studies focused on using wearable sensors to recognize health-related behaviors that involve high periodicity in the sensed signal, such as sitting, walking, and running. Other activities that occur less frequently throughout the day, such as eating gestures pose an even more significant challenge. One challenge in recognizing these activities or gestures within continuous sensor signals is that of spotting these activities amid irrelevant data in a user-independent manner. Many existing machine learning models are trained to spot these activities, yet yield a large number of false positives, impacting the feasibility of these wearable devices in longitudinal studies. Moreover, many machine learning models are designed without consideration to the limited computational resources of the wearable device, compromising the utility of the models in detecting behavior in real-time. In this research, I focus on spotting sporadic activities that are infrequent throughout the day and considered health-related behaviors, namely, chewing to understand problematic eating behaviors and coughing for respiratory disease diagnostics and assessment, which were further magnified risk behaviors during the Covid-19 pandemic. To address the detection of these human behaviors, I propose a wrist-worn sensor-based feeding gesture detection system that detects overeating behavior using a novel motif-based data fusion algorithm. I then present a neck-worn multi-sensor wearable system equipped with a peak-based segmentation machine learning pipeline to allow reliable chewing detection in free-living settings. Finally, I propose and demonstrate a novel multi-centroid classifier to resolve the on-device challenge of cough detection using a computationally limited earbud platform.

Creator