Risk-Aware Multi-Horizon Glucose Prediction for Type 2 Diabetes Using Temporal Attention and Conformal Prediction

Reliable glucose forecasting is an important tool for proactive and safety-oriented diabetes supervision. However, at long prediction horizons and in heterogeneous type 2 diabetes (T2D) populations, point forecasts alone are insufficient for risk-aware alerting. We propose a multi-horizon glucose prediction framework that employs a lightweight recurrent forecaster along with a temporal attention uncertainty quantification via split conformal prediction. Prediction intervals calibrated separately for each horizon are integrated into a supervisory layer that triggers alerts based on clinically relevant thresholds. Experiments on real-world CGM data show that temporal attention consistently improves forecasting accuracy, reducing the root mean square prediction error (RMSE) at a 120min horizon from 33.4 to 28.3,mg/dL. When uncertainty is incorporated into supervision, event recall exceeds 0.92 for both hypoglycemia and hyperglycemia, approaching 0.95 at conservative operating points. By tuning the miscoverage level α, daily alarm rates can be reduced from above 90 to approximately 21-31 per day, depending on the event type, while maintaining recall above 0.90.