Comparing Different Individualized Black-Box Models for Insulin Pump Faults Detection on Artificial Pancreas Data

Objective: Prompt and accurate detection of insulin pump faults could be key in preventing sustained hyperglycemia and possibly ketoacidosis. Model-based fault detection techniques relay on patient models to warn the patient of a possible malfunctioning. Here, we want to compare four individualized models to understand if there is a preferable choice in terms of fault detection ability. Method: Individualized, linear, black-box parametric models (ARX, ARMAX, ARIMAX, BJ) are identified with BIC-based optimal order on 7 days of fault-free closed-loop data for 100 virtual subjects, using one of the most recent versions of the UVA-Padova T1 Diabetic Patient Simulator, that accounts for intra/inter-patient variability. An online prediction of up to PH=3h of glucose concentration, along with its confidence interval, is calculated through a Kalman filter based on the individualized model, running on data of past infused insulin (possibly affected by an unknown 6h insulin suppression), ingested meals and CGM values for 10 days. The real time fault detection algorithm raises an alarm if the predicted residual portion stays out its confidence interval for more than 15min. Result: The performance of the detection method is evaluated in terms of false positives per day and recall (FP/days, RE%), also taking into consideration the detection time. The performance is (0.12, 70%), (0.13, 85%), (0.17, 80%), (0.16, 78%), while the detection time is 236min, 242min, 237min, 238min for ARX, ARMAX, ARIMAX and BJ, respectively. If we compute the Euclidian distance from the optimal point (0, 100%), this metric will suggest ARMAX as the best model. Conclusion: Although ARMAX appears to be the best choice, the use of the other models only slightly impacts the fault detection performance and detection time.