dc.description.abstract | Eating and exercise detection using continuous glucose monitor (CGM) signals is key to provide recommendations for a healthy lifestyle. However, this can be challenging given imbalanced data and other contexts. Previous works have used accelerometers, gyroscopes, glucose monitors, and other sensors but not necessarily all three plus others combined. Therefore, I aim to build a model by testing various techniques and testing glucose along with different statistical body measurements, such as electrodermal activity, heart rate, blood volume, accelerometer, gyroscope, etc. A sliding window is used to extract statistical measures from each body measurement, such as standard deviation, mean, and range to look for patterns correlated to eating and exercise. I select an extreme gradient boosted decision tree algorithm with Synthetic Minority Oversampling Technique. I compare the performance of just solely using glucose and then adding more sensory data and discovered that there is not consistent change in performance. I also adjusted the window and overlap to compare eating detection performance and found that there is not a concrete impact on the performance. Furthermore, I performed exercise detection and compare with and without CGM. There appears to be no significant performance difference with or without glucose. In addition to eating detection, I also examine for correlation between glucose variation and exercise moments. I finally conclude that it is not feasibly possible to detect eating with my current methods. However, for exercise detection, I can produce better detection results compared to eating, but my current method for detecting correlations between glucose levels and exercise moments can be later improved. | |