GAT-BiGRU: explainable multi-task temporal graph learning for glucose forecasting, hypoglycemia risk, and...

People with type 1 diabetes mellitus (T1DM) show glucose variability driven by insulin dosing, meals, activity, and circadian rhythms. Many deep learning approaches treat glucose forecasting and hypoglycemia detection as separate tasks and provide limited transparency.

We developed an explainable, multi-task temporal graph framework that jointly predicts glucose trajectories and hypoglycemia risk at 30 and 60 minute horizons, and provides bounded, patient-specific insulin adjustment recommendations.

Temporal GAT-BiGRU transforms multimodal continuous glucose monitoring (CGM) time series into a temporal k-neighborhood graph, with each time point represented as a feature-enriched node. A graph-attention encoder performs multi-head message passing over history edges, while an attention-based BiGRU captures longer dependencies. We evaluated OhioT1DM and BrisT1D using hypoglycemia and predicted Time in Range (TIR) metrics. A prediction-driven counterfactual module retrospectively generates bounded basal/bolus adjustments using patient-specific Insulin-to-Carbohydrate Ratio (ICR) and Insulin Sensitivity Factor (ISF); interpretability is supported via GNNExplainer.

On OhioT1DM, Temporal GAT-BiGRU achieved hypoglycemia precision-recall area under the curve (PR-AUC) 0.93, mean absolute error (MAE) 9.40 mg/dL, root mean square error (RMSE) 15.8 mg/dL, mean absolute relative difference (MARD) 6.01%, and predicted TIR 71.39%. On BrisT1D, performance remained strong with PR-AUC 0.98, MAE 9.36 mg/dL, RMSE 15.3 mg/dL, MARD 7.55%, and predicted TIR 64.78%. The insulin module generated bounded, subject-specific recommendations, typically suggesting ∼10% basal increases with individualized meal-bolus updates.

Temporal GAT-BiGRU provides accurate glucose prediction through temporal graph reasoning, sequence modeling, and interpretable explanations. It supports personalized decision support and closed-loop glucose management systems.