 The paper discusses the importance of using interpretable models for decision support systems, DSS, in type 1 diabetes, T1D, management, specifically for corrective insulin boluses, CIB, suggestion. It presents a case study that compares two long short-term memory neural networks, LSTM, with similar prediction accuracy, and shows that only one of them learned the physiological relationship between inputs and glucose prediction, making it the preferred choice for integration in the DSS. The paper also uses SHAP to explain the output of black-box models and verifies the effectiveness of the selected model in improving patient's glycemic control. This article was authored by Francesco Prendin, Giacapo Pavan, Giacomo Cappan, and others.