Blood Glucose Prediction Using a Lightweight Sequential Transformer with Heart Rate Integration

Abstract

Diabetes mellitus remains one of the most prevalent chronic conditions worldwide, demanding continuous and accurate monitoring of blood glucose levels to prevent life-threatening complications such as hypoglycemia and hyperglycemia. Traditional glucose prediction systems rely solely on historical glucose readings, overlooking the physiological relationship between cardiovascular activity and glycemic fluctuations. In this work, a lightweight Sequential Transformer model is proposed that integrates heart rate signals alongside conventional diabetes management inputs — including basal insulin, bolus insulin, and carbohydrate intake — to achieve more physiologically informed blood glucose forecasting across multiple prediction horizons ranging from 5 to 30 minutes. The model is trained and evaluated on the OhioT1DM dataset comprising six Type-1 diabetic patients. To address the clinically critical problem of hypoglycemia detection, a Combined Focal-Asymmetric Huber Loss is introduced alongside a hypoglycemia-aware oversampling strategy. A post-training threshold calibration further tunes the decision boundary by maximising the F2-score on the validation set. The proposed system achieves a root mean square error of 5.81 mg/dL, a mean absolute percentage error of 1.28%, and an R² of 0.877, with 98.89% of predictions falling within the clinically safe Zone A of the Clarke Error Grid. Hypoglycemia sensitivity improved from 0% in the baseline to 66.7% at the critical 5-minute prediction horizon, demonstrating that targeted loss design and sampling strategies can transform a clinically unsafe model into a practically deployable glucose forecasting system.