Detection of drinking via a wrist-worn inertial sensor.

Abstract

Alcohol addiction is the third leading lifestyle-related cause of death in the United States. There are not enough support tools for alcoholics who want to quit alcohol consumption. The detection of drinking in a free-living environment may improve the just-in-time adaptive intervention to this behavior. Traditional methods to detect alcohol consumption suffer from long response time that hinders prompt intervention and prevention. This paper proposes to employ inertial sensors to automatically detect drinking of alcohol by leveraging the hand gesture characteristics that are specific to drinking. Due to the lack of publicly available sensor dataset of alcohol drinking, this paper focused on the detection of general beverage drinking by exploiting the hand gestures. A public dataset containing seven daily activities (including hand-related activities such as eating, drinking, smoking, etc.) collected from 11 subjects in a controlled environment was adopted for this study. The detection model was developed using deep neural networks containing both convolutional and recurrent neural networks. The proposed approach achieved an F1-score accuracy of 87% in the Leave-One-Subject-Out (LOSO) cross-validation. We argue that contributions of this paper would be useful when an alcohol specific dataset becomes available.

Drinking Gesture Detection Using Wrist-Worn IMU Sensors with Multi-Stage Temporal Convolutional Network in Free-Living Environments

“…We first applied the MS-TCN model on DX-I and DX-II separately using LOSO cross-validation to evaluate the performance in semi-controlled environments and free-living environments. Then the CNN-LSTM approach from [13] was applied to our datasets as the benchmark. presents the performance of segment-wise evaluation with three thresholds (k=0.1, 0.25, and 0.5).…”

“…This approach detected HtM with an F1-score of 85% in free-living environments (Eating activities were excluded in their dataset). Senyurek et al [13] proposed a convolutional neural network and longshort-term-memory network combined model (CNN-LSTM) for detecting drinking gesture using IMU data acquired from smartwatches. This approach achieved an F1-score of 87% on a publicly available dataset collected from 11 participants in the Leave-One-Subject-Out (LOSO) scheme.…”