Enhancing Continuous Glucose Monitoring-based Eating Detection with Wearable Biomarkers

Proper diet monitoring is a cornerstone of preventing and treating Type 2 Diabetes. However, this usually relies on burdensome manual meal logging. Continuous glucose monitors (CGMs), which have recently gained popularity as a tool to help Type 2 Diabetics with their treatments, may allow for a burden-free, sensor-based approach to logging periods of eating through monitoring the glucose dynamics and attempting to identify periods of post-prandial glucose response. However, CGMs-alone may not be sufficient in properly detecting periods; periods such as those present in gastric emptying may result in false positives for eating detection, given the sharp rise in glucose response. This work seeks to augment CGM-captured signals with that of other wearable biomarkers, captured from smartwatches, to aid in the detection of eating periods. These signals have been shown to detect eating motions. We explore a hierarchical model approach to augmenting CGM-based eating detection with additional sensing modalities. We test our model data collected from healthy participants eating in free-living conditions. We find that CGM-based eating detection can be improved by retrospectively reviewing wearable sensing data for confirmation, improving our model performance of eating detection, as measured by the area under the receiver operating characteristic curve, by 0.15 (from 0.64 to 0.79), and similarly across additional performance metrics.