Toward clinically interpretable control: Nonlinear MPC for safe and efficient automated insulin delivery

초록

Closed-loop glucose regulation must simultaneously achieve three competing objectives: accurate tracking of target glucose levels, minimal control effort, and robustness to physiological disturbances. This study presents a controlled comparative evaluation framework for automated insulin delivery using a MATLAB-based virtual patient cohort simulation environment. Five controllers, namely Nonlinear Model Predictive Control (NMPC), Adaptive MPC (AMPC), Linear MPC (LMPC), Linear-Quadratic Regulator (LQR), and Proportional-Integral-Derivative (PID), are evaluated under clinically relevant disturbance scenarios, including meal variability, sensor noise, and exercise-induced perturbations, across two dynamic gain configurations. Controller performance is assessed using clinically relevant metrics, including root mean square error (RMSE) relative to a 100 mg/dL target, time in range (TIR), and insulin effort per hour (EffI), along with actuator-related indicators such as constraint satisfaction and saturation behavior. Simulation results over a cohort of 25 virtual patients indicate consistent performance advantages of predictive controllers over classical approaches. In particular, NMPC achieves up to 41% reduction in RMSE, 22% improvement in TIR, and 37% reduction in insulin effort compared to PID and LQR. Predictive controllers also demonstrate stable regulation under noisy and exercise-perturbed conditions, with average computation time of 0.3 s, supporting real-time feasibility. These results highlight NMPC as an effective approach for balancing glycemic accuracy, control effort, and robustness under challenging conditions. More broadly, the proposed framework enables reproducible, cohort-level evaluation of control strategies for automated insulin delivery systems.

키워드