Monitoring Statistical Properties of Kalman Filter Residuals in an Artificial Pancreas to Detect Overnight Pump Malfunctions Causing Insulin Suspension

Continuous Subcutaneous Insulin Infusion (CSII) pumps are essential components in a closed-loop insulin delivery system (the so-called artificial pancreas), a recent therapeutic option for type 1 diabetes management. The efficacy of glucose regulation obtained by such a system might be severely compromised if the pump fails, possibly putting the patient’s safety at risk. Therefore, reliable and timely detection of these malfunctions is of critical practical importance. For this reason, in this study we present a novel algorithm specifically designed to detect overnight pump malfunctions resulting in insulin suspension. This algorithm adopts a real-time model-based approach and builds upon our team’s previous work, which relied solely on the magnitude of prediction residuals. In contrast, our current contribution extends the fault detection strategy by incorporating a comprehensive analysis of various statistical properties associated with the residuals. These properties encompass not only the magnitude but also the sample mean, median, variance, power, and other statistics related to the whiteness of the residuals. The considered linear black-box patients’ models have been individualized by estimating the model parameters on seven days of each patient’s data. Subsequently, the model is used in a Kalman filter to generate the predicted glucose levels with their confidence intervals. The effectiveness of the new strategy is compared to the previously proposed approach on an in-silico dataset generated by the UVa/Padova Type 1 Diabetes simulator, which has been accepted by the US Food and Drug Administration as a substitute for animal testing prior to human clinical trials with an artificial pancreas. The best performing detection strategy is based on the sign-sum test and achieves a recall of 91.3%, while generating less than 1 false alarm every 10 days. This improves the performance of our previous contribution that exhibits a comparable number of false alarms, but an overall lower recall metric, equal to 82.7%. The proposed algorithm is designed to complement other dedicated modules within the system, aiming to detect malfunctions related to glucose sensors (such as missed samples and pressure-induced sensitivity losses) as well as human errors (including missed meal announcements). A robustness analysis has been conducted to examine the performance of the algorithm in the presence of these additional anomalies. The results of these analyses demonstrated that the proposed algorithm maintains its sensitivity to pump malfunctions while exhibiting only a marginal increase in false alarms.