Assessing the Features on Blood Glucose Level Prediction in Type 1 Diabetes Patients Through Explainable Artificial Intelligence

Type 1 diabetes mellitus (T1DM) is an autoimmune disease causing insulin deficiency and impaired blood sugar control. Managing postprandial glucose response (PGR) is a major challenge in T1DM, as it necessitates precise insulin dosing prior to meals. The potential of the artificial pancreas lies in its integration of insulin delivery and glucose monitoring, but accurate incorporation of meal information remains an open challenge. While machine learning methodologies could help in addressing this issue, most existing models lack interpretability and comprehension of the output. This study aims to assess the influence of various metabolic features on the prediction of postprandial Blood Glucose Levels (BGLs) using eXplainable Artificial Intelligence (XAI) methodologies. Two distinct prediction models are developed, considering varying prediction horizons following the meal (15 min and 60 min). Specifically, BGLs prediction models incorporate preprandial blood glucose values, insulin, meal carbohydrate intake, and glycemic index and load as input variable. To evaluate their impact, SHapley Additive exPlanations (SHAP), a method that provides insights into the contribution of each feature to the model predictions, was employed. By leveraging XAI methodologies, the study aims to improve the interpretability and transparency of BGLs prediction models. The findings of this research can contribute to the development of decision-support tools for patients affected by T1DM, helping the management of PGR and reducing the risks of adverse events.