Personalized Multi-Horizon Glucose Forecasting in Type 1 Diabetes via Deep Learning

This study presents a Deep Learning framework for fore- casting blood glucose levels at a time horizon of 45 and 60 minutes using ElectroCardioGraphy (ECG)–derived features alongside baseline glucose readings. Using data from the D1NAMO dataset, we extracted and normalized Heart Rate (HR) and Heart Rate Variability (HRV) met- rics and categorized future glucose levels into clinically meaningful bins. A feedforward neural network was trained to classify these categories. The resulting model demonstrated promising classification performance on both prediction horizons, with balanced predictive capability across most glucose categories. To enhance transparency and clinical relevance, we employed SHapley Additive exPlanations (SHAP) to assess feature importance and interpret the model’s predictions. SHAP analysis re- vealed that HR and HRV metrics—particularly hr_max, hr_min, hr_std, and nni_mean—provide valuable complementary information to current glucose levels for predicting short-term glycemic trends. This highlights the potential of ECG–derived features as predictive biomarkers for short- term glycemic changes. Overall, the proposed framework offers a promis- ing and explainable approach to glucose forecasting, supporting trans- parency, physiological insight, and trust–key factors for the adoption of predictive models in real-time clinical monitoring.