Detection of health-related behaviours using head-mounted devices
[bi:thesis]
Shengjie Bi. Detection of health-related behaviours using head-mounted devices. PhD thesis, Dartmouth Computer Science, Hanover, NH, May 2021. ©Copyright the author. PhD Dissertation.Abstract:
The detection of health-related behaviors is the basis of many mobile-sensing applications for healthcare and can trigger other inquiries or interventions. Wearable sensors have been widely used for mobile sensing due to their ever-decreasing cost, ease of deployment, and ability to provide continuous monitoring. In this dissertation, we develop a generalizable approach to sensing eating-related behavior.
First, we developed Auracle, a wearable earpiece that can automatically detect eating episodes. Using an off-the-shelf contact microphone placed behind the ear, Auracle captures the sound of a person chewing as it passes through the head. This audio data is then processed by a custom circuit board. We collected data with 14 participants for 32 hours in free-living conditions and achieved accuracy exceeding 92.8% and F1 score exceeding77.5% for eating detection with 1-minute resolution.
Second, we adapted Auracle for measuring children’s eating behavior, and improved the accuracy and robustness of the eating-activity detection algorithms. We used this improved prototype in a laboratory study with a sample of 10 children for 60 total sessions and collected 22.3 hours of data in both meal and snack scenarios. Overall, we achieved 95.5% accuracy and 95.7% F1 score for eating detection with 1-minute resolution.
Third, we developed a computer-vision approach for eating detection in free-living scenarios. Using a miniature head-mounted camera, we collected data with 10 participants for about 55 hours. The camera was fixed under the brim of a cap, pointing to the mouth of the wearer and continuously recording video (but not audio) throughout their normal daily activity. We evaluated performance for eating detection using four different Convolutional Neural Network (CNN) models. The best model achieved 90.9% accuracy and 78.7%F1 score for eating detection with 1-minute resolution. Finally, we validated the feasibility of deploying the 3D CNN model in wearable or mobile platforms when considering computation, memory, and power constraints.
Citable with [BibTeX]
Projects: [auracle]
Keywords: [mhealth] [sensors] [wearable]
Available from the publisher: [page]
Available from the author:
[bib]
Please obtain a copy from the publisher.