Exploring Nutritional Influence on Blood Glucose Forecasting for Type 1 Diabetes Using Explainable AI

Area of Science:

• Endocrinology and Metabolism
• Artificial Intelligence in Medicine
• Computational Biology

Background:

• Type 1 diabetes mellitus (T1DM) requires precise blood glucose management, often aided by artificial pancreas (AP) systems.
• Current AP systems struggle with postprandial glucose response (PGR) due to incomplete understanding of meal impacts.
• Accurate insulin dosing for meals necessitates better prediction of postprandial blood glucose levels (BGLs).

Purpose of the Study:

• To quantify the influence of meal-related factors on predicting BGLs at 15, 60, and 120 minutes post-meal.
• To utilize deep neural network (DNN) models for BGL prediction incorporating nutritional data.
• To enhance the interpretability of BGL prediction models using eXplainable Artificial Intelligence (XAI).

Main Methods:

• Developed DNN models to predict BGLs using preprandial glucose, insulin dose, and meal nutritional factors (energy, carbs, protein, lipids, fiber, GI, GL).
• Applied SHapley Additive exPlanations (SHAP) to assess the impact and contribution of each input feature on model predictions.
• Validated model performance and feature importance against clinical literature hypotheses.

Main Results:

• Identified specific meal components and their quantities significantly impacting BGLs at various postprandial time points.
• Quantified the contribution of each nutritional factor to BGL prediction accuracy using SHAP values.
• Demonstrated the effectiveness of DNNs and XAI in understanding complex PGR determinants.

Conclusions:

• Meal composition significantly influences postprandial glucose excursions in T1DM.
• XAI methods provide crucial insights into the factors driving BGL predictions, enhancing model transparency.
• Findings support the development of advanced decision-support tools and improved AP technology for T1DM management.