Development of autoencoder-guided attention-LSTM models for predicting nocturnal hypoglycemia risk in Type 1 Diabetes

OBJECTIVE: Nocturnal hypoglycemia (NH) is a major, often undetected risk for individuals with Type 1 Diabetes (T1DM). Current prediction models lack sufficient lead time for proactive, pre-sleep interventions. This study develops a novel, interpretable deep learning approach for predicting NH events up to 12 hours in advance. METHODS: We propose a hybrid model that infuses physiological knowledge (simulated glucose absorption, insulin kinetics, and subcellular insulin signaling) into an attention-based Long Short-Term Memory (LSTM) network. An Autoencoder-based Attention Weight Mapping (A2WM) framework is introduced to trace model attention back to the original 24-hour inputs (CGM, meals, insulin) for interpretability. The model was developed using the OhioT1DM dataset (open-loop pump users) and externally validated on the SMARTDIAB dataset (open-loop pump users). RESULTS: On the OhioT1DM test set, the model achieved an AUC of 0.9387, recall of 0.9333, and specificity of 0.8548. On the external SMARTDIAB cohort, it demonstrated robust generalizability, achieving an AUC of 0.8684, recall of 0.6316, and specificity of 0.9615. Decision Curve Analysis confirmed a high net clinical benefit, and the A2WM interpretations identified clinically relevant temporal features. CONCLUSION: The proposed physiologically-informed, attention based LSTM model can accurately and interpretably predict 12 hour NH risk from historical data and scheduled basal insulin. SIGNIFICANCE: This work provides a clinically valuable tool for proactive NH management, offering an extended lead time that empowers T1DM patients to implement preventive strategies (e.g., pre-sleep meal or insulin adjustments) well before hypoglycemia occurs.