Machine-Learning-Based Detection of Pressure-Induced Faults in Continuous Glucose Monitors

Abstract

Continuous glucose monitors (CGMs) are prone to faults termed pressure-induced sensor attenuations (PISAs), particularly when the user rolls over on the sensor during sleep. PISAs result in false, low blood glucose readings, leading to undesirable pump shutoffs and an increased risk of hyperglycemia. Data from an outpatient trial with PISA glucose readings labeled for 1125 nights was used. Machine learning methods such as decision trees, support vector machines, neural networks, random forest, and gradient-boosted machine models are compared against each other and with a previously reported rules-based algorithm for PISA detection on the same data set. The best-performing gradient-boosted machine model is further improved using the developed start-PISA model. Model interpretation methods are used to confirm that PISA behavior is well-captured and provides insights into decision-making.