Abstract
This study extends the evaluation of a personalized machine learning approach to predicting blood glucose (BG) levels in patients with Type 1 Diabetes (T1DM) conducted during the CDDIAB study. The primary aim was to assess the relevance of these predictions in supporting patients' therapeutic decision-making processes. The extended study involved 15 additional T1DM patients who recorded BG levels, meal intake, and insulin doses over 30 days, with no alterations to their standard treatment regimens. Predictive models were developed using pharmacokinetic modeling and machine learning algorithms to forecast BG fluctuations up to 90 minutes in advance.Two analyses were performed: the first compared patient decisions made with predictions versus their actual actions, while the second compared predictions with conventional treatment decisions. Results showed that in 85% of cases, patients made better therapeutic decisions when provided with BG predictions, such as adjusting insulin dosing or taking preventative actions for low BG. The study concluded that BG predictions offer valuable insights for patients, improving decision-making over simple alerts. Future efforts will focus on integrating these predictions into a decision support system that can provide direct therapeutic recommendations to patients.