AI for Health

🇬🇧 Error anticipation of blood glucose predictions for people with diabetes

Nov 17 10:25 - 10:37

Présentée par


Modeling the diabetes mellitus system is extremely complex in nature: it is dynamic, non-linear, interactive, and patient-specific. Mastering it allows to simulate accurately blood glucose levels (BGL) trends and anticipate excursions (hypo-/hyperglycemia), which will help improve the lives of tens of millions of patients worldwide. To achieve this, scientists have focused on very complex nonlinear models to try to describe the system. At Hillo, we focused on a patient-based machine learning approach trained on a few weeks of historical patient data with unmatched accuracy and performance. However, the task of predicting a BGL value one hour ahead is very ambitious, and the model lacks crucial information such as the patient’s behavior over the prediction horizon. We will show how we developed a probabilistic framework that focuses on predicting the BGL probability density rather than a BGL value. Based on this, we build a confidence interval predictor, a below-range probability estimator, and a method to filter out the most hazardous predictions to secure the system.