Attention-Based Multi-Task Learning and PPO Reinforcement Learning for Explainable Blood Glucose Prediction

Accurate blood glucose prediction is critical for effective diabetes management, yet existing models struggle with individual variability, data sparsity, and non-linearity.We propose an attention-based multi-task learning (MTL) model integrated with proximal policy optimization (PPO) reinforcement learning (RL) to enhance forecasting accuracy and adaptability.MTL captures shared patterns across multiple prediction tasks, while PPO dynamically refines predictions based on patientspecific glucose trends.By incorporating explainability techniques such as SHAP analysis and Monte Carlo dropout, our approach not only achieves state-of-the-art predictive accuracy but also enhances trust in AI-driven decision support systems for diabetes care.Evaluated on the BrisT1D blood glucose dataset, our model achieves a R 2 score of 0.85, MSE of 0.0017, RMSE of 0.0419, and MAE of 0.0310, significantly surpassing conventional methods.This work advances personalized real-time glucose forecasting, offering a promising step toward AI-powered glycemic management in clinical settings.