Glucose Insulin Digital Twin and Artificial Pancreas Using PID, MPC and RL

SIMULATION WITH VIRTUAL PATIENTS, NOT PEOPLE. NOT A MEDICAL DEVICE, NOT TREATMENT ADVICE. How do basal-bolus therapy, PID, model predictive control on a digital twin and a reinforcement-learning agent compare when the comparison is fair? A minimal-model glucose-insulin twin is calibrated per patient from a logged day and all controllers run on the 30 virtual patients of simglucose (UVA/Padova 2008), tuned on 12 patients and tested on 18 others with the same meals and sensor noise. With announced meals, MPC and PPO + meal bolus reach a median time in range of 87.9 % and 93.9 % and cannot be separated. Without meal information PPO reaches 65.8 %, tuned PID 61.7 %, MPC 53.3 %. The median time below range is 0 for all, and it hides up to 46 % in the worst patient, so spread and worst cases are always reported.

Category: Biomedical Engineering

Biomedical Engineering Project BM05 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.