Safe and Explainable AI for Personalized Insulin Dosing in Type 1 Diabetes

Abstract

People with Type 1 Diabetes (T1D) make frequent insulin dosing decisions, where both over- and under-dosing carry serious health consequences. Existing digital tools are often reactive, relying on current glucose measurements while under-using key physiological factors such as insulin-on-board, carbohydrate absorption, and physical activity. This paper presents a unified AI-driven framework for personalized insulin dosing and safety-aware diabetes management, integrating three components: (1) a multimodal glucose prediction model using CGM history, insulin-on-board, carbohydrate intake, physical activity, and circadian features; (2) a safety-aware insulin dosing module grounded in ADA guidelines and established bolus calculation principles; and (3) a retrieval-augmented explanation layer providing patient-specific clinical reasoning. The framework also includes meal logging and longitudinal safety monitoring. Unlike prior work addressing prediction, dosing, or explainability in isolation, the framework combines them into a proactive decision-support pipeline. On two independent real-world T1D datasets, the prediction models achieve MAE as low as 0.2042 mmol/L and Clarke Zone A+B accuracy up to 99.88%, with minimal degradation under cross-dataset transfer. The dosing module remains within clinically accepted safety bounds, and a 14-day retrospective patient replay demonstrates accurate prediction, interpretable reasoning, and effective detection of high-risk glucose events. These results support safer, explainable AI-assisted diabetes management.