Development of a method for compliance detection in wearable sensors

Authors: FARD, SIAVASH - University Of Alabama; GHOSH, TONMOY - University Of Alabama; HOSSAIN, DELWAR - University Of Alabama; MCCRORY, MEGAN - Boston University; THOMAS, GRAHAM - Brown University School Of Medicine; HIGGINS, JANICE - University Of Colorado; JIA, WENYAN - University Of Pittsburgh; BARANOWSKI, TOM - Children'S Nutrition Research Center (CNRC); STEINER-ASIEDU, MATILDA - University Of Ghana; ANDERSON, ALEX - University Of Georgia; SUN, MINGUI - University Of Pittsburgh; FROST, GARY - Imperial College; LO, BENNY - Imperial College; SAZONOV, EDWARD - University Of Alabama

Submitted to: IEEE Access
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 9/17/2023
Publication Date: 1/22/2024
Citation: Fard, S.E., Ghosh, T., Hossain, D., McCrory, M.A., Thomas, G., Higgins, J., Jia, W., Baranowski, T., Steiner-Asiedu, M., Anderson, A.K., Sun, M., Frost, G., Lo, B., Sazonov, E. 2024. Development of a method for compliance detection in wearable sensors. 2023 International Conference on Electrical, Computer and Energy Technologies (ICECET). https://doi.org/10.1109/ICECET58911.2023.10389483.
DOI: https://doi.org/10.1109/ICECET58911.2023.10389483

Interpretive Summary: Studies relying on wearable sensors to measure physical activity or food intake must assess wear time (compliance with the directions for wearing the sensor). This paper reports a novel method for measuring wear time using an accelerometer in the Automatic Ingestion Monitor v2 (AIM-2), a set of sensors attached to an eyeglass frame. The method was developed using data from 30 participants for two days each (US dataset) and tested with an independent dataset (Ghana dataset) on 10 households (30 Participants, 3 days for each, a total of 90 days). The signals from the accelerometer provided extract features used to train an artificial intelligence algorithm. The accuracy of the algorithm for the US dataset was 95.37%, and for the Ghana dataset was 95.86%, showing satisfactory performance. The trained algorithm can be used to monitor compliance with device wear in real-time applications.

Technical Abstract: One of the crucial elements in studies relying on wearable sensors for quantification of human activities (like physical activity or food intake) is the assessment of wear time (compliance). In this paper, we propose a novel method based on the Automatic Ingestion Monitor v2 (AIM-2), deployed for measuring nutrient and energy intake. The proposed method was developed using data from a study of 30 participants for two days each (US dataset) and tested with an independent dataset (Ghana dataset) on 10 households (30 Participants, 3 days for each, a total of 90 days). The signals from the accelerometer sensor of the AIM-2 were used to extract features and train the gradient-boosting tree classifier. To reduce the error in the classification of non-compliance in situations where the sensor changes its position with respect to gravity, a two-stage classifier followed by post-processing was introduced. Previously, we developed an offline compliance classifier, and this work aimed to develop a classifier for a cloud-based feedback system. The accuracy and F1-score of the developed two-phase classifier based on K-fold validation for the training and validation dataset were 95.37% and 96.93%, and for the Ghana dataset, were 95.86% and 92.56%, respectively, showing satisfactory performance results. The trained classifier can be deployed to monitor compliance with device wear in real-time applications.