Abstract
Background and Aims
Recent advancements of closed-loop insulin delivery for T1D therapy aim to achieve better control while simultaneously reducing the need of patient interventions. As patients become increasingly confident in new technologies, it is important to develop automated methods that can detect possible malfunctioning in real-time. This work analyzes an innovative approach to detect insulin pump faults (IPF) based on advanced machine learning algorithms.
Methods
Using the latest version of the T1D Padova/UVA simulator, we generated a 30 days dataset in which we simulated IPF occurring at night, during fasting, and during the day, in correspondence of meals. From the data, we extracted a large pool of numerical attributes (features) capable of describing the patient status over time and highlighting suspicious portions. Then, a selection procedure was performed to determine the most effective feature set for detecting IPFs. Finally, we compared the performance of several state of art unsupervised anomaly detection algorithms on the data obtained.
Results
The best performance overall is obtained by the Histogram Based Outlier Score algorithm, which detected 87% of the IPF with 0.08 False Positives per day on average. Considering only the overnight period, we obtained a recall of 0.79 with 0.03 FP/day. In the diurnal portion, we obtained a higher recall of 0.95 with 0.05 FP/day.
Conclusions
In-silico data show that the proposed method can detect IPF and improve the safety of CL systems. In the future, we aim to test it inside dedicated clinical trials.
