Design and Validation of an Open-Source Closed-Loop Testbed for Artificial Pancreas Systems

TL;DR

This paper introduces an open-source, validated testbed for artificial pancreas systems that enables safe, efficient testing of control algorithms and safety features using realistic simulations and fault injection, accelerating diabetes research.

Contribution

The paper presents a novel open-source testbed integrating glucose simulators and fault injection to facilitate validation of APS control algorithms and safety features.

Findings

Advanced algorithms maintained blood glucose in safe range over 93% of the time.

Testbed accurately reproduces patient glucose trajectories from clinical data.

Fault injection demonstrates controller resilience under hazardous conditions.

Abstract

The development of a fully autonomous artificial pancreas system (APS) to independently regulate the glucose levels of a patient with Type 1 diabetes has been a long-standing goal of diabetes research. A significant barrier to progress is the difficulty of testing new control algorithms and safety features, since clinical trials are time- and resource-intensive. To facilitate ease of validation, we propose an open-source APS testbed by integrating APS controllers with two state-of-the-art glucose simulators and a novel fault injection engine. The testbed is able to reproduce the blood glucose trajectories of real patients from a clinical trial conducted over six months. We evaluate the performance of two closed-loop control algorithms (OpenAPS and Basal Bolus) using the testbed and find that more advanced control algorithms are able to keep blood glucose in a safe region 93.49% and 79.46% of the time on average, compared with 66.18% of the time for the clinical trial. The fault injection engine simulates the real recalls and adverse events reported to the U.S. Food and Drug Administration (FDA) and demonstrates the resilience of the controller in hazardous conditions. We used the testbed to generate 25 years of synthetic data representing 20 different patient profiles with realistic adverse event scenarios, which would have been expensive and risky to collect in a clinical trial. The proposed testbed is a valid tool that can be used by the research community to demonstrate the effectiveness of different control algorithms and safety features for APS.