LAUSR

Introduction: Many CGM users find the magnitude and complexity of their data challenging. We developed an AI method for detecting and classifying discernable self-management events reflected in CGM data. Methods: Machine learning and signal detection techniques were used to detect CGM patterns which we call “CGM events.” The key features used for detection include raw glucose values, smoothed values, numeric derivatives of smoothed values, and features based on time and date of readings. For training, a 10-fold cross validation was used to tune the necessary parameters. We used mean square error, sensitivity, and specificity to evaluate the models’ performance. After detection, each event is classified according to clinical significance based on glucose levels, time above target, and a severity score. Results: We trained different models on 17280 data points from 20 patients over 60 days and then evaluated their performance using separate test data. The system accurately detected and classified CGM events from actual data. The severity score used in classifying events is significantly negatively correlated with the standard time in range measure. Conclusions: Advanced machine learning and signal detection techniques can be applied to accurately detect CGM events. The classification of detected events may give CGM users and providers more insights into glucose data and can be used for self-management or clinical decision support. Disclosure S. Liu: None. M. Shomali: Employee; Self; WellDoc, Inc. A. Kumbara: Employee; Self; WellDoc, Inc. A. K. Iyer: Employee; Self; WellDoc, Inc. M. Peeples: Employee; Self; WellDoc, Inc. M. A. Dugas: None. K. Crowley: None. G. Gao: None.