Scholarlink

<p>Type 1 diabetic patients require exogenous insulin to maintain normal blood glucose levels. Recent advances, such as artificial pancreas (AP) systems, have improved disease management. These systems enable closed-loop continuous subcutaneous insulin infusion (CSII), reducing patient burden and lowering hypoglycemia risk. Traditional control algorithms, such as PID (proportional–integral–derivative) and MPC (model predictive control), are commonly used in insulin pumps. While they improve glycemic control, they often lack adaptability to the patient’s changing physiological state. In this study, we propose a reinforcement learning (RL)-based deep Q-learning agent to control insulin delivery in type 1 diabetic patients. Traditional algorithms struggle with the significant delay in insulin’s physiological effects. To address this, we introduce a novel insulin action replay (IAR) mechanism by adding a look-ahead reward based on future outcomes, improving learning and control accuracy. We also utilize a dual-hormone therapy to automate the delivery of basal insulin and glucagon, which significantly reduces the risk of hypoglycemia and contributes to enhanced glucose management. Compared to the standard basal-bolus treatment with low glucose suspend (LGS) and RL-based random replay (RR), our IAR framework shows superior time-in-range (TIR) performance. In adults, IAR achieved 87.17% TIR, exceeding RR by 4.4% and LGS by 14.4%. In adolescents, it achieved 87.16% TIR, outperforming RR by 2.0% and LGS by 15.3%. These gains were statistically significant at the 0.05 level (adults: <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(p=0.016\)</EquationSource> </InlineEquation>; adolescents: <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(p=0.008\)</EquationSource> </InlineEquation>). Overall, the RL-based IAR shows superior adaptability and offers a promising supplement to existing glycemic control methods for Type 1 diabetes.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Dual hormone controller for type 1 diabetes based on dueling deep Q-network and insulin action replay

  • 期刊论文

  • 发表日期：

    2026年3月31日

查看全文
  • Namit Arora,
  • Gorseet Paul Singh,
  • Bhomic Kaushik,
  • Sheetal Rajpal,
  • Virendra Kumar,
  • Ankit Rajpal

摘要

Type 1 diabetic patients require exogenous insulin to maintain normal blood glucose levels. Recent advances, such as artificial pancreas (AP) systems, have improved disease management. These systems enable closed-loop continuous subcutaneous insulin infusion (CSII), reducing patient burden and lowering hypoglycemia risk. Traditional control algorithms, such as PID (proportional–integral–derivative) and MPC (model predictive control), are commonly used in insulin pumps. While they improve glycemic control, they often lack adaptability to the patient’s changing physiological state. In this study, we propose a reinforcement learning (RL)-based deep Q-learning agent to control insulin delivery in type 1 diabetic patients. Traditional algorithms struggle with the significant delay in insulin’s physiological effects. To address this, we introduce a novel insulin action replay (IAR) mechanism by adding a look-ahead reward based on future outcomes, improving learning and control accuracy. We also utilize a dual-hormone therapy to automate the delivery of basal insulin and glucagon, which significantly reduces the risk of hypoglycemia and contributes to enhanced glucose management. Compared to the standard basal-bolus treatment with low glucose suspend (LGS) and RL-based random replay (RR), our IAR framework shows superior time-in-range (TIR) performance. In adults, IAR achieved 87.17% TIR, exceeding RR by 4.4% and LGS by 14.4%. In adolescents, it achieved 87.16% TIR, outperforming RR by 2.0% and LGS by 15.3%. These gains were statistically significant at the 0.05 level (adults: \(p=0.016\) ; adolescents: \(p=0.008\) ). Overall, the RL-based IAR shows superior adaptability and offers a promising supplement to existing glycemic control methods for Type 1 diabetes.