Explainable AI Assessment of Meal-Related Features Impact in Predicting Basal Insulin for Type I Diabetes

Type 1 diabetes management using artificial pan-creas (AP) technology faces challenges in accurately regulating postprandial glucose levels. Current AP systems, based on heuris-tic control algorithms for basal insulin delivery, still require user input for mealtime insulin delivery. In the context of personalized patient-centered medicine, this study explores the impact of meal-related features on predicting postprandial basal insulin requirements using explainable artificial intelligence (XAI). Machine learning (ML) models were trained using preprandial blood glucose, mealtime insulin, preprandial basal insulin, and features from 15 T1D patients. The performance in terms of the root-mean-square-error (RMSE) and the mean-absolute-error (MSE) were (0.42 ± 0.1) insulin unit and (0.32 ± 0.07) insulin unit, respectively. Employing SHapley Additive exPlanations (SHAP) reveals significant influences of mealtime insulin bolus, carbohy-drates, and glycemic load, aligning with physiological knowledge on postprandial glycemia response (PGR).