Insulin Dose Recommendation System for Type-1 Diabetes Patients Using Reinforcement Learning and Continuous Glucose Monitoring Data

Abstract
This research proposes a high-performance reinforcement learning (RL)-based insulin dose recommendation system for Type-1 Diabetes Mellitus (T1DM), integrating continuous glucose monitoring (CGM) data, temporal state encoding, and hybrid reward optimization. Utilizing PPO and SAC as policy backbones and incorporating LSTM-driven glucose sequence modeling, the framework was validated across synthetic and real-world datasets including UVA/Padova simulations and OhioT1DM records. The system consistently achieved Time-in-Range (TIR) performance exceeding 96.8%, with mean absolute error (MAE) reduced to 12.9 mg/dL and zero critical hypoglycemic episodes, confirming its dosing precision under variable physiological conditions. Compared to rule-based and conventional RL methods, the proposed model demonstrated a 4.5-5.6% improvement in glycemic outcomes and a 26% reduction in insulin dosing variance. The inclusion of hybrid reward shaping and noise-robust policy regularization enabled safe learning across diverse patient profiles, including meal-induced spikes and sensor noise conditions. With inference latency below 150 ms and online adaptation capability, the system is well-suited for real-time closed-loop insulin therapy applications. Its data-driven design supports deployment in intelligent artificial pancreas systems, without reliance on handcrafted dosing heuristics or fixed clinical protocols.

Author
Sazan Kamal Sulaiman

DOI
https://www.google.com/search?q=https://doi.org/10.1109/ICCR64138.2025.11292565

ISSN
2836-8142

Publish Date: 19-Dec-2025