Drinking event detection on a sensing wristband using machine learning

Adequate hydration is important for one’s health, but many people do not consume sufficient fluids. By constantly monitoring fluid intake, we gain information that can be extremely useful in dealing with unhealthy drinking habits. This paper deals with the problem of developing a machine learning method for drinking detection, intended for use on an edge device, with a specific focus on power consumption. The proposed approach is based on data from inertial sensors built into a practical, non-invasive wrist-worn device that monitors wrist movement throughout the day and automatically detects drinking events. It ensures low energy consumption by triggering the machine learning only when the probability of drinking is high, as well as by other energy saving measures. To develop and validate our methods, we collected data from 19 participants, which resulted in 135 hours of data, of which 2 hours and 30 minutes correspond to drinking activities. The algorithm was thoroughly assessed through both offline testing and by running the algorithm directly on the wristband in real life. During the offline evaluation, we obtained a precision of 94.5 %, a recall of 84.9 %, and an F1 score of 89.4 %. Testing in real life demonstrated a precision of 74.5 % and a recall of 89.9 %. Additionally, the energy efficiency analysis showed that our proposed technique for triggering the drinking detection method reduced the battery power consumption during the periods of inactivity by a factor of 5.8 compared to continuously monitoring for drinking events.