Abstract
This study presents the initial assessment of a personalized machine learning approach to predict blood glucose (BG) levels in patients with Type 1 Diabetes (T1DM), conducted as part of the CDDIAB study. The goal was to evaluate the efficacy of a predictive system using individual BG measurements combined with contextual data, helping patients improve glycemic control by anticipating fluctuations in blood glucose. Fourteen patients with T1DM participated in the study, tracking their BG levels, meal intake, and insulin doses over 30 days under real-life conditions without any specific intervention on their usual diabetes treatment. The study developed predictive algorithms based on each patient's data, using pharmacokinetic modeling and machine learning techniques to provide 30-, 60-, and 90-minute predictions. These models were evaluated using Parkes error grid analysis, which measures the clinical accuracy of BG predictions. Results showed that 99.9%, 98.6%, and 96.3% of predicted BG values fell within the clinically acceptable zones (A and B) for 30-, 60-, and 90-minute prediction horizons, respectively. This promising approach demonstrates the potential to assist patients in making better insulin dosing decisions, particularly during periods of rapid glucose changes, and could be integrated into mobile applications for real-time decision support.