Intelligent Dual Basal–Bolus Calculator for Multiple Daily Insulin Injections via Offline Reinforcement Learning

Managing blood glucose levels through multiple daily injections (MDIs) of insulin presents a challenge in daily diabetes care, necessitating frequent self-dosing adjustments to avoid glycemic extremes. While essential in MDI treatment, bolus calculators need significant usability improvements due to meal counting and parameter optimization complexities. Recent advances in insulin dosing calculators based on reinforcement learning (RL) offer potential solutions, as they can adapt to varying insulin–glucose dynamics. However, many RL approaches present safety risks during training, where suboptimal actions can lead to dangerous glycemic events. Hence, this study aimed to introduce a novel dosing algorithm for MDIs, termed ‘Dual-BB (basal–bolus) agent’, by combining offline RL with an online fine-tuning technique. The offline phase utilized historical insulin treatment data to train the model, reducing the risks associated with real-time learning. Meanwhile, online fine-tuning allows for real-time patient-specific adjustments. The Dual-BB agent demonstrated improved safety and optimized insulin dosing compared to conventional methods in trials conducted in silico using virtual type 1 diabetes patients. By analyzing the control performance metrics of time-in-range (TIR) and time-below-range (TBR) before and after implementing the proposed algorithm, we observed a TIR improvement of 9.7% and a TBR improvement of 1.1% when comparing the proposed algorithm with conventional basal–bolus calculation algorithms. These results suggest that our offline learning approach effectively mitigates early-stage risks while providing stable, personalized treatment optimization. The Dual-BB algorithm presents a practical solution for MDI treatment in clinical settings, ensuring safer and more effective glucose management.