Adaptive regulation of blood glucose levels: A triadic methodology incorporating super twisting and deep policy gradient

作     者：Mohammadzadeh, Ardashir Alattas, Khalid A. 

作者机构：Sakarya Univ Fac Engn Dept Elect & Elect Engn Sakarya Turkiye Univ Jeddah Coll Comp Sci & Engn Dept Comp Sci & Artificial Intelligence Jeddah Saudi Arabia 

出 版 物：《BIOMEDICAL SIGNAL PROCESSING AND CONTROL》 (Biomed. Signal Process. Control)

年 卷 期：2025年第103卷

核心收录：

学科分类：0831[工学-生物医学工程（可授工学、理学、医学学位）] 10[医学] 

基　　金：University of Jeddah, Jeddah, Saudi Arabia [UJ-24-DR-423-1] University of Jeddah 

主　　题：Pancreas system Diabetes patients Closed-loop controller Supertwisting sliding mode control (SSMC) Deep reinforcement learning algorithm (DRL) Sine-Cosine algorithm (SCA) 

摘      要：The insufficiency of insulin secretion by the pancreas leads to elevated blood glucose levels (BGL) in individuals diagnosed with diabetes. Addressing this challenge, the artificial pancreas (AP) has emerged as a viable solution for the autonomous regulation of BGL through the continuous infusion of insulin. The proposed methodology employs a triadic approach to develop an intelligent closed-loop pancreatic system. First, a nonlinear controller, specifically the supertwisting sliding mode control (SSMC), is applied for the regulation of blood glucose. To enhance the quality and efficacy of the controller, a deep reinforcement learning (DRL) algorithm is then integrated to facilitate online updates of the SSMC coefficients. Finally, a sine-cosine algorithm (SCA) is formulated to optimize the hyperparameters of the DRL approach, addressing the challenges associated with the calibration of reinforcement learning algorithms. The developed system has been evaluated on six diabetic patients, involving various tests with differing BGL references, meal consumption, and the management of uncertainties. A comparative analysis with conventional methodologies further substantiates the superiority of the proposed SSMC-based DRL-SCA system, demonstrating an overall average superiority of approximately 61.13% in tracking performance compared to traditional control methods across various cases.