Menu
Recent Research Spotlight
In a recent work, we propose how to incorporate deep evidential regression to construct glucose prediction models that are equipped with uncertainty quantification.
Our recent work Uncertainty Quantification in Neural Network-based glucose prediction for diabetes was presented at ProbML 2026 : Symposium on Probabilistic Machine Learning held at Seoul in July 2026.
In our study, we integrate an uncertainty-quantification framework known as evidential regression into various neural network types and found that an optimal model leveraging a Transformer-based backbone structure leads to a glucose prediction system that yields the best calibrated uncertainty distributions.
The diagram above is a rolling forecast plot with one hour predictive horizon for our model based on combined input sequences of basal insulin, blood glucose, bolus insulin, carbohydrate intake. Dashed ellipses locate points where model would fail to detect hypoglycemia unless the uncertainty bands are used for detection thresholds. This gives an example of how uncertainty estimates such as the shaded 95% confidence interval here can be very useful. We hope that our work contributes towards future developments of robust A.I.-enabled glucose prediction models that can provide warning alerts for diabetic patients or assist in glucose trend monitoring systems.