Abstract
Background and Aims: Hillo has developed a Machine Learning (ML) method to predict future blood glucose levels (BGL) with a very good accuracy, but for large prediction horizons the error can lead to bad decisions for the patient. It is necessary to anticipate and mitigate these errors in real time to give a confidence index in the prediction and filter out the riskiest ones.Methods: We have developed a ML method to model the uncertainty around a predicted value. When a prediction is performed, only part of the information that may impact BGL is known (past CGM readings, insulin injections and meal intakes). Future unknown inputs and external disturbance to the system can be modeled as a random noise around an expected value. A BGL predictor and a ML variance estimator are trained consecutively, using the same data, to estimate the probability of having a blood glucose value at a target time, conditioned to our known information at prediction time, thus we build a consistent conditional probability distribution estimator:Results: Performances are compared with a Uniform Error Model computed from the marginal error distribution. The density model methodology can be improved by using a different distribution‐model and by adding a calibration step.Conclusions: We showed a method to anticipate and mitigate risky situations.