Presenter of 1 Presentation
ACCURACY OF BGL PREDICTION FROM A PERSONALIZABLE PHYSIOLOGICAL MODEL OF BLOOD GLUCOSE DYNAMICS USING REAL-WORLD DATA
Abstract
Background and Aims
Glycemic control is challenging due to the complex blood glucose (BGL) regulation dynamics. A system generating personalized models to mimic the individual BGL regulation physiology was developed, enabling BGL predictions up to 24 hours and personalized insulin dose optimization.
Methods
96 subjects (30 T1DM and 66 T2DM, 53 female, age 47.9±15.5 years, BMI of 29.1±5.0 kg/m2) included in a clinical study carried out in partnership with Grupo Fleury and Medtronic were monitored during 144 hours with CGM (required 2 or more calibrations/day), activity tracker and our own mobile app to collect BGL, meal, insulin and physical activity data. The model was individually trained using 96-hour data, and tested using 24-hour data by calculating 6-hour and 24-hour BGL prediction curves that were evaluated using MARD and Consensus Error Grid (CEG).
Results
The median MARD for all, T1DM and T2DM patients for 6-hour BGL predictions was 16.0%, 29.8% and 13.9% respectively, and for 24-hour was 16.8%, 37.1% and 14.5% respectively. The median CEG points in zones AB (CEGAB) for 6-hour and 24-hour BGL predictions for all and T2DM patients was 100%, while T1DM patients showed 91.6% and 89.7% respectively. The median CEG points in zones DE (CEGDE) was 0% for all groups and prediction horizons.
Conclusions
Our model accurately predicted BGL up to 24-hour ahead, showing potential on BGL excursion risk identification, insulin dose optimization, preventive actions recommendation, among others. We thank Grupo Fleury for clinical, financial, protocol design, structure, and follow-up support, and Medtronic for providing the iPro2 CGMs and protocol design support.