A comparative study of explainability methods for time-series forecasting of blood glucose levels

Abstract

Deep learning models offer significant potential for blood glucose forecasting in diabetes management, yet their clinical adoption is limited by interpretability challenges. This study presents a benchmark of explainable AI (XAI) methods—including Attention, Saliency Maps, Integrated Gradients, SHAP, and LIME—applied to a Bidirectional LSTM model for multivariate clinical time-series forecasting. Using the OhioT1DM dataset and a Bidirectional LSTM forecasting model, we conduct a benchmark of explainable AI (XAI) methods for multivariate clinical time-series forecasting. Results demonstrate significant performance variations among XAI methods, with attention mechanisms and gradient-based approaches showing high faithfulness for temporal clinical data. The study provides practical guidelines for selecting appropriate explanation methods in clinical settings, advancing the development of transparent and trustworthy AI systems for diabetes care that can enhance clinician confidence and support informed decision-making.