Despite the importance of maintaining an adequate hydration status, water intake is frequently neglected due to the fast pace of people's lives. For the elderly and people with cognitive deficits, reduced water intake can be even more worrying, not only because of the detrimental effects of dehydration, but also because the mechanisms to regulate hydration in these individuals are usually less effective. It is well established that drinking related habits play a critical role in overall human health. By constantly monitoring fluid intake, we gain information that can be extremely useful in dealing with unhealthy drinking habits and play an important role in preventing many diseases.
This thesis deals with the problem of developing a novel machine learning method for drinking detection in naturalistic settings. The main focus was the development of a drinking detection program that works directly on the microcontroller with the lowest possible power consumption. It 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. To develop robust methods, we designed a special data collection procedure, which collected 135 hours of data, of which 2 hours and 30 minutes corresponds to drinking activities.
To validate the performance of the proposed method, an extensive evaluation was carried out on a computer and directly on the wristband. The best model turned out to be XGBoost with an accuracy of 94.3 % and a recall of 83.5 % with a signal window length of 4.92 s and 50% overlap. We also performed a test in a controlled environment with predefined drinking and non-drinking activities. We achieved a recall score of 76.0 %. Finally, some volunteers wore the bracelet during the day in free living conditions, with a precision of 74.5 % and recall of 89.9 %. Energy efficiency analysis showed that with moderate triggering of the method, we can expect autonomous operation of the wristband for up to 203 days.
The achieved results suggest that the proposed method is capable of detecting drinking events in a free-living scenario using data from wrist-worn inertial sensors and is robust enough to cope with data from participants about whom it had no prior knowledge.
|