Abstract
This study presents a machine learning (ML) method designed to predict and mitigate real-time blood glucose (BGL) prediction uncertainty in patients with Type 1 Diabetes (T1DM). While accurate BGL predictions can assist patients in managing their insulin needs, erroneous predictions may lead to incorrect dosing decisions. The proposed approach aims to provide reliable BGL predictions for up to 2 hours, while minimizing risky prediction errors. The method uses a Gaussian density model to estimate the conditional probability distribution of future BGL levels, based on continuous glucose monitoring (CGM) data, carbohydrate intake, and insulin doses from 14 T1DM patients.The study trains both a mean and standard deviation predictor to form a conditional density estimator, which can predict BGL values and their associated uncertainty. By computing probabilities for different zones of the Parkes Error Grid Analysis (EGA), the system can filter predictions with a high likelihood of falling in zones associated with greater clinical risk (zones C, D, and E). The methodology shows promising results, with prediction uncertainty decreasing when the algorithm filters high-risk predictions. Future improvements include optimizing the density model and incorporating a calibration step to further enhance prediction accuracy and clinical safety.