Bolus Insulin calculation without meal information. A reinforcement learning approach

In continuous subcutaneous insulin infusion and multiple daily injections, insulin boluses are usually calculated based on patient-specific parameters, such as carbohydrates-to-insulin ratio (CR), insulin sensitivity-based correction factor (CF), and the estimation of the carbohydrates (CHO) to be ingested. This study aimed to calculate insulin boluses without CR, CF, and CHO content, thereby eliminating the errors caused by misestimating CHO and alleviating the management burden on the patient. A Q-learning-based reinforcement learning algorithm (RL) was developed to optimise bolus insulin doses for in-silico type 1 diabetic patients. A realistic virtual cohort of 68 patients with type 1 diabetes that was previously developed by our research group, was considered for the in-silico trials. The results were compared to those of the standard bolus calculator (SBC) with and without CHO misestimation using open-loop basal insulin therapy. The percentage of the overall duration spent in the target range of 70-180 mg/dL was 73.4% and 72.37%, <70 mg/dL was 1.96 and 0.70%, and >180 mg/dL was 23.40 and 24.63%, respectively, for RL and SBC without CHO misestimation. The results revealed that RL outperformed SBC in the presence of CHO misestimation, and despite not knowing the CHO content of meals, the performance of RL was similar to that of SBC in perfect conditions. This algorithm can be incorporated into artificial pancreas and automatic insulin delivery systems in the future.