Incorporating Prior Information in Adaptive Model Predictive Control for Multivariable Artificial Pancreas Systems

Area of Science:

• Biomedical Engineering
• Control Systems Engineering
• Computational Physiology

Background:

• Adaptive model predictive control (MPC) shows promise for automated artificial pancreas systems.
• Current models lack explicit incorporation of prior knowledge, limiting accuracy.
• Integrating prior information on meals and activity can enhance glycemic prediction and control.

Purpose of the Study:

• To develop an adaptive MPC algorithm using regularized partial least squares (rPLS) for improved glucose prediction.
• To enhance artificial pancreas control by incorporating prior knowledge into the glucose prediction model.
• To assess the algorithm's robustness against disturbances like meals and physical activity.

Main Methods:

• Developed a glucose prediction model using regularized partial least squares (rPLS), encoding prior information as a regularization term.
• Implemented an adaptive MPC with dynamic setpoint and insulin dosing constraints based on estimated plasma insulin concentration (PIC).
• Validated the adaptive MPC using the multivariable glucose-insulin-physiological variables simulator (mGIPsim) with in silico subjects.

Main Results:

• The proposed adaptive MPC achieved 81.9% ± 7.4% time in range (TIR) for virtual subjects.
• This compares favorably to 73.9% ± 7.6% TIR for a standard MPC using an autoregressive exogenous (ARX) model.
• The adaptive MPC demonstrated no hypoglycemia or severe hypoglycemia, outperforming the ARX-based MPC.

Conclusions:

• The adaptive MPC algorithm effectively incorporates prior knowledge for improved glucose prediction model updating.
• This approach contributes to developing fully automated artificial pancreas systems.
• The system shows potential for mitigating disturbances from meals and physical activity in type 1 diabetes management.