MPC-based active fault detection in artificial pancreas for T1D treatment

Type 1 diabetes (T1D) is characterized by the inability to produce insulin, requiring lifelong exogenous administration to maintain euglycaemia and avoid the severe short- and long-term consequences of hypo- and hyperglycaemia. The Artificial Pancreas (AP) has emerged as a leading therapeutic solution, automating insulin delivery through closed-loop algorithms based on continuous glucose monitoring. However, the safety of AP systems relies heavily on the reliability of hardware components; faults such as insulin pump occlusions or delivery failures can lead to prolonged hyperglycaemia and keto-acidosis if not promptly detected. Current fault detection methodologies typically operate as passive monitoring layers that analyse system behaviour reactively. The objective of this work is to overcome these limitations by proposing an Active Fault Detection strategy integrated directly into the control algorithm. The core idea is to shape the control action such that it not only regulates glucose but also maximizes the information available to distinguish between nominal operation and fault conditions. To achieve this, a modified Model Predictive Control (MPC) formulation is proposed. The standard cost function is modified to include a "likelihood of detection" term, designed to amplify the divergence between nominal and faulty output predictions. To ensure that this "informative" control action does not compromise patient safety or induce hypoglycaemia, the optimization problem is formulated including soft output constraints (OC). To evaluate the performance of the proposed system, the UVA/Padova T1D Simulator (FDA-accepted model for preclinical testing) was used, employing a cohort of 100 virtual adult subjects. Scenarios of complete pump failure were simulated and compared between the proposed MPC formulation and a baseline MPC formulation (Pavia/Padova) alongside standard passive detection methods, testing different activation strategies. The results demonstrate that introducing the active detection component significantly enhances the sensitivity of external detectors. Both tested detectors reflected an improvement in true positive rate and detection delay. Furthermore, the analysis confirms that the inclusion of output constraints is crucial: while the unconstrained active controller degrades glycaemic control, the proposed constrained formulation successfully preserves safety (maintaining a stable Glucose Risk Index and Time in Range) while enabling superior fault detection capabilities.