Explainable Multi-Task Deep Learning for Blood Glucose Forecasting: A Lightweight, Interpretability Focused Approach

Accurate and transparent blood glucose forecasting is crucial for effective diabetes management. This paper presents an interpretable deep learning (DL) framework based on multi-task learning (MTL) for forecasting glucose levels at multiple prediction horizons. In contrast to complex hybrid systems, we isolate the MTL core and integrate explainability methods such as Shapley Additive Explanations (SHAP) and permutation-based feature importance (PFI). Our approach enables a clear understanding of model behavior while achieving strong predictive performance. Evaluated on the BrisT1D dataset, the model achieves an R^2 score of 0.956, RMSE of 0.045, and MAE of 0.033, while highlighting the critical influence of historical glucose readings. This focused study provides insights into how interpretable AI can support reliable decision-making in diabetes care.