Frequent and long-term exposure to hyperglycemia increases the risk of chronic complications, including neuropathy, nephropathy, and cardiovascular disease. Existing continuous subcutaneous insulin infusion (CSII) and continuous glucose monitoring (CGM) technologies model only specific aspects of glycemic regulation, such as predicting hypoglycemia and administering small insulin boluses. Similarly, current digital twin approaches in diabetes management primarily focus on predicting glucose responses to human behavior and insulin therapy. As a result, these technologies lack the ability to provide alternative treatment scenarios that could guide proactive behavioral interventions for optimal diabetes management. To address this gap, we propose GlyTwin, a novel computational framework that enhances digital twin technologies by integrating counterfactual explanations to simulate optimal behavioral treatments for glucose control. GlyTwin generates counterfactual treatments by recommending adjustments to behavioral choices, such as carbohydrate intake and insulin dosing, to significantly reduce the occurrence and duration of hyperglycemic events. In addition, GlyTwin incorporates stakeholder preferences into its intervention-generation process, ensuring that the tool is personalized and user-centric. We evaluate GlyTwin on AZT1D, a new dataset constructed by collecting longitudinal data from 50 individuals living with type 1 diabetes (T1D) on automated insulin delivery (AID) systems, each monitored for 26 days. Results show that GlyTwin outperforms state-of-the-art methods for generating counterfactual explanations, with 85.8\% valid explanations and 87.3\% effectiveness in preventing hyperglycemia compared with historical data.

翻译：摘要：长期频繁暴露于高血糖环境会增加神经病变、肾病及心血管疾病等慢性并发症风险。现有持续皮下胰岛素输注（CSII）与持续血糖监测（CGM）技术仅能模拟血糖调控的特定环节（如预测低血糖及实施小剂量胰岛素推注）。同样，当前糖尿病管理领域的数字孪生方法主要聚焦于预测人体行为及胰岛素治疗对血糖的响应。因此，这些技术缺乏提供替代治疗方案的能力，难以指导实现最优糖尿病管理的前瞻性行为干预。针对这一空白，我们提出GlyTwin——一种通过整合反事实解释模拟血糖控制最优行为治疗的新型计算框架。GlyTwin通过推荐碳水化合物摄入量及胰岛素剂量等行为选择调整方案生成反事实治疗方案，显著降低高血糖事件的发生频率与持续时间。此外，该工具在干预生成过程中融入了利益相关者的偏好，确保其具备个性化与用户中心化特性。我们在AZT1D数据集（新构建的50名使用自动胰岛素输注（AID）系统的1型糖尿病（T1D）患者纵向数据，每人监测26天）上评估了GlyTwin性能。结果表明，与历史数据相比，GlyTwin在反事实解释生成方面优于现有最优方法，解释有效率达85.8%，预防高血糖有效性达87.3%。