Comparing Supervised and Unsupervised Algorithms for the Retrospective Detection of Faults in Continuous Glucose Monitoring Sensors

Continuous Glucose Monitors (CGMs) are minimally invasive sensors that play a key role in managing Type 1 Diabetes (T1D). Among the common issues affecting these devices are pressure-induced sensor attenuations (PISAs), which can lead to falsely low glucose readings, impacting both real-time monitoring and retrospective analysis. This study explores the retrospective detection of PISAs using supervised and unsupervised learning methods. While supervised methods benefit from labeled data to distinguish normal from faulty conditions, their applicability is limited by the challenge of obtaining accurate patient-specific labels. In contrast, unsupervised methods require no labels, making them particularly valuable in T1D, where personalization is crucial. The algorithms are evaluated on a dataset generated in silico with a state-of-the-art T1D simulator. Among the tested algorithms, Random Forest and Isolation Forest demonstrate promising results, with the latter achieving a recall of 74% and less than two false alarms throughout the monitoring period. These findings in a simulated setting suggest potential for further investigation on real-world data and improvements in detecting additional sensor malfunctions.