Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Probabilistic Forecasting of Blood Glucose in Diabetes Mellitus
We aim to develop a (probabilistic) forecast model for glucose levels in individuals with diabetes mellitus based on historical continuous glucose monitoring (CGM) data and machine learning (ML= under real-world conditions. Ultimately, these forecasts may improve diabetes management systems such as the artificial pancreas.
A traditional point forecast is a single number, say 5.5 mmol/L. In contrast, a probabilistic forecast has the form of a predictive probabilistic distribution, such as a normal distribution with a mean of 5.5 mmol/L and a standard deviation of 0.35 mmol/L. Thus, probabilistic models have the advantage of quantifying uncertainty in the models’ predictions. As a starting point, you will get familiar with our existing ML pipeline for probabilistic forecasts based on historical CGM data. We evaluated our model based on a dataset consisting of N=400 individuals with over 33’000 thousand days of data. Currently, we are increasing the size of the dataset to create the world’s largest CGM dataset. As a second step, we will emphasize the robustness (e.g., against distribution shift) and generalization capability of the evolved ML models. For example, the models should work equally well on data before and after changing an artificial pancreas system.
A traditional point forecast is a single number, say 5.5 mmol/L. In contrast, a probabilistic forecast has the form of a predictive probabilistic distribution, such as a normal distribution with a mean of 5.5 mmol/L and a standard deviation of 0.35 mmol/L. Thus, probabilistic models have the advantage of quantifying uncertainty in the models’ predictions. As a starting point, you will get familiar with our existing ML pipeline for probabilistic forecasts based on historical CGM data. We evaluated our model based on a dataset consisting of N=400 individuals with over 33’000 thousand days of data. Currently, we are increasing the size of the dataset to create the world’s largest CGM dataset. As a second step, we will emphasize the robustness (e.g., against distribution shift) and generalization capability of the evolved ML models. For example, the models should work equally well on data before and after changing an artificial pancreas system.