The alarmingly sharp increase in type 1 diabetes mellitus incidence in recent years has been coupled with a marked rise in the development of closed-loop systems aimed at continuously monitoring blood glucose levels, predicting glucose behavior and delivering adjusted insulin doses. However, to date, these systems still fall short of ensuring tight glycemic control in many patients, particularly in patients with high glucose variability. The presented clinically optimized blood glucose prediction technology, developed at the Weizamnn Institute of Science, provides for improved prediction accuracy, regardless of patient age and disease severity, and outperforms standard models, particularly in high-risk patient populations. The technology can be easily integrated into existing artificial pancreas systems and bears the potential of enhancing disease management and patient quality of life, translating to the long-term clinical benefits of stable glycemia.
This invention can be integrated in standard closed-loop artificial pancreas systems comprised of a continuous glucose monitoring device and an insulin pump.
- Trained and tested using massive volumes of real-life data
- Accurate in broad, heterogeneous patient populations, including high-risk populations
- Prediction horizon of up to 60 minutes
- Compatible with standard closed-loop system components
- Easily implementable
The technology developed by this research group involves a clinically optimized blood glucose prediction algorithm with a reduced error rate. Its development leveraged machine-learning computational tools and over 1.5 million glucose recordings collected from 141 type 1 diabetes mellitus (T1DM) patients simultaneously using a CGM and insulin pump. The model integrated a novel gradually connected network (GCN) in place of the standard artificial neural networks, which proved to be highly suitable for continuously acquired data. The algorithm demonstrated high accuracy, with similar performance for patients of different age groups and from a range of degrees of glycemic control. Moreover, it outperformed the autoregressive model in cases of high-risk patients spending a large percentage of time in hypoglycemia and in patients with high blood glucose variability.
Comparison of performance of GCN3 model (orange) compared to the autoregressive model (blue) in patients of (A) different age groups, (B) different baseline hemoglobin A1c (HbA1c) values, (C) glucose variability and (D) percent of time in hypoglycemia. Shown is the percentage of predictions falling in clinically hazard zones (C-E) on the Clarke Error Grid (CEG) for all tested patients. From: Diabetes Technol Ther. 2020, 22.