Explainable cluster-based learning for prediction of postprandial glycemic events and insulin dose optimization in type 1 diabetes.

Abstract

Effective management of postprandial glycemic excursions in type 1 diabetes requires accurate prediction of adverse events and personalized insulin adjustments informed by interpretable models. This study presents an explainable dual-prediction framework that simultaneously forecasts postprandial hypoglycemia and hyperglycemia within a 4-hour window using cluster-personalized ensemble models. Glycemic profiles were identified through a hybrid unsupervised approach combining self-organizing maps and k-means clustering, enabling the training of specialized random forest classifiers. The system outperformed baseline models on both real-world and simulated datasets, achieving high performance (AUC = 0.84 and 0.93; MCC = 0.47 and 0.73 for hypo- and hyperglycemia, respectively). Model interpretability was addressed using global (SHAP) and local (LIME) explanations, while interaction analysis revealed the non-linear effects of carbohydrate intake and insulin bolus combinations. An insulin adjustment module further refined pre-meal bolus recommendations based on predicted risk. Simulated evaluations confirmed improved postprandial time-in-range and reduced hypoglycemia without excessive hyperglycemia. These results underscore the potential of profile-driven and explainable machine learning approaches to support safer, individualized diabetes care.