Supervised Learning-based Detection of Pressure-induced Failures in Continuous Glucose Sensors

Continuous Glucose Monitors (CGMs) are mini- mally invasive sensors that are crucial for the management of Type 1 Diabetes (T1D). Pressure-induced sensor attenuations (PISAs) are common faults that occur in the every-day use of the devices. PISAs result in false low blood glucose readings that can impact the reliability in both real-time and retrospective use of the sensors. In this work, we investigate the detection of PISAs failures using supervised learning techniques. Specifically, 6 algorithms, namely Support Vector Machine, Logistic Regression, k-Nearest Neighbors, Decision Tree, Random Forest and Adaptive Boosting are tested and compared to recognize the failures on a dataset generated in-silico using a state-of-art T1D simulator. Encouraging performance are achieved by some of the algorithms tested. In particular, Random Forest recognized up to 90% of the failures with less than 1 false alarms every 10 days. The results obtained in the simulated scenario paves the way for further analysis on real-data and possible improving for detecting new types of fault of the sensors.