An Assessment of the Accuracy of an Automated Bite Counting Method in a Cafeteria Setting

Advances in body-worn sensors and mobile health technology have created new opportunities for empowering people to take a more active role in managing their health. Obesity has been recognized as a target of opportunity that could particularly benefit from this approach. Selfmonitoring of dietary intake is critical for weight loss/management, but currently used tools such as food diaries require users to manually estimate and record energy intake, making them subjective, prone to error, and difficult to use for long periods of time. Our group is developing a new tool called the “bite counter” that automates the monitoring of caloric intake. The device is worn like a watch and uses sensors to track wrist motion during a meal. Previous studies have shown that our method accurately counts bites during controlled and uncontrolled meals in the lab. This thesis describes a study to evaluate the accuracy of the method in a cafeteria setting. A cafeteria booth that can seat 1 to 4 people was instrumented with tethered wrist motion trackers, embedded scales, and video cameras, to enable recording of wrist motion, changes in food weight, and actual activities during eating. A total of 276 subjects were recorded eating uncontrolled meals. The data was manually reviewed and the times of all actual bites taken were recorded as “ground truth”. The wrist motion data was then analyzed using the automated bite counting method to determine the times of automated bite detections. These were compared against the ground truth to evaluate the accuracy of the bite counting method. In total, 22,383 bites were evaluated, consisting of 380 different foods, eaten using 4 different utensils from 4 different containers, across a variety of subject demographics. Results show that the method varied in accuracy from 39 % (for ice cream cones) to 88% (for salad bar) across the 39 most commonly eaten foods (>=100 bite occurrences in the data set). The average accuracy found across all bites was 76% with a positive predictive value of 87%. A second test of the bite counting method using modified timing thresholds resulted in 82% accuracy with a 82% positive predictive value. These results indicate that the method works well across a