GlucoNet-MM: A multimodal attention-based multi-task learning framework with decision transformer for personalised...

Accurate and personalized blood glucose prediction is critical for proactive diabetes management. Conventional machine learning (ML) models often struggle to generalize across patients due to individual variability, nonlinear glycemic dynamics, and sparse multimodal input data. This study aims to develop an advanced, interpretable deep learning (DL) framework for patient-specific, policy-aware blood glucose forecasting.

We propose GlucoNet-MM, a novel multimodal DL framework that combines attention-based multi-task learning (MTL) with a Decision Transformer (DT), a reinforcement learning paradigm that frames policy learning as sequence modeling. The model integrates heterogeneous physiological and behavioral data, continuous glucose monitoring (CGM), insulin dosage, carbohydrate intake, and physical activity, to capture complex temporal dependencies. The MTL backbone learns shared representations across multiple prediction horizons, while the DT module conditions future glucose predictions on desired glycemic outcomes. Temporal attention visualizations and integrated gradient-based attribution methods are used to provide interpretability, and Monte Carlo dropout is employed for uncertainty quantification.

GlucoNet-MM was evaluated on two publicly available datasets, BrisT1D and OhioT1DM. The model achieved R² scores of 0.94 and 0.96 and mean absolute error (MAE) values of 0.031 and 0.027, respectively. These results outperform single-modality and conventional non-adaptive baseline models, demonstrating superior predictive accuracy and generalizability.

GlucoNet-MM represents a promising step toward intelligent, personalized clinical decision support for diabetes care. Its multimodal design, policy-aware forecasting, and interpretability features enhance both prediction accuracy and clinical trust, enabling proactive glycemic management tailored to individual patient needs.