Integrating Neural Differential Forecasting with Safe Reinforcement Learning for Blood Glucose Regulation

2025-11-15

AI Summary

Addressing the challenge of balancing personalization, safety, and robustness in automated insulin delivery (AID) systems for type 1 diabetes under dynamic meal and physiological variations, this paper proposes TSODE—a novel controller integrating uncertainty-aware Neural Ordinary Differential Equations (NeuralODEs) for glucose prediction with Thompson sampling–based reinforcement learning. Prediction uncertainty is rigorously quantified via conformal calibration, and a safety-aware action suppression mechanism prevents hazardous dosing decisions—specifically, pre-meal insulin overdosing and repeated corrections. Evaluated on the FDA-accepted UVa/Padova simulator, TSODE achieves 87.9% time-in-range (TIR) and <10% time-below-range, significantly outperforming existing baselines. The framework ensures clinical interpretability, safety compliance, and practical deployability while maintaining rigorous uncertainty quantification and adaptive personalization.

Technology Category

• Humans and AI: Human-Aware Planning and Behavior Prediction
• Machine Learning: Calibration & Uncertainty Quantification
• Search and Optimization: Sampling/Simulation-based Search

Application Category

• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

Automated insulin delivery for Type 1 Diabetes must balance glucose control and safety under uncertain meals and physiological variability. While reinforcement learning (RL) enables adaptive personalization, existing approaches struggle to simultaneously guarantee safety, leaving a gap in achieving both personalized and risk-aware glucose control, such as overdosing before meals or stacking corrections. To bridge this gap, we propose TSODE, a safety-aware controller that integrates Thompson Sampling RL with a Neural Ordinary Differential Equation (NeuralODE) forecaster to address this challenge. Specifically, the NeuralODE predicts short-term glucose trajectories conditioned on proposed insulin doses, while a conformal calibration layer quantifies predictive uncertainty to reject or scale risky actions. In the FDA-approved UVa/Padova simulator (adult cohort), TSODE achieved 87.9% time-in-range with less than 10% time below 70 mg/dL, outperforming relevant baselines. These results demonstrate that integrating adaptive RL with calibrated NeuralODE forecasting enables interpretable, safe, and robust glucose regulation.

Problem

• Balancing glucose control and safety under meal uncertainty
• Guaranteeing safety while enabling adaptive personalization in diabetes
• Preventing insulin overdosing risks through predictive uncertainty quantification

Innovation

• Integrates Thompson Sampling RL with NeuralODE forecaster
• Uses conformal calibration to quantify predictive uncertainty
• Rejects or scales risky insulin actions for safety