75-LB: Machine-Learned, Biophysical Prediction of Glucose Response for 1,000 Subjects

Background: Reducing glycemic excursions following oral intake decreases the risk for microvascular disease in those with type 2 diabetes (T2D) and prevents conversion to T2D in those with prediabetes. Currently there are no accurate methods by which to predict a real world subject’s glycemic response to a particular meal. We developed a novel machine learning tool to infer personalized physiological parameters using consumer sensors for the accurate prediction of glycemic responses to food and activity.

Method: We prospectively recruited 1,000 healthy, prediabetic and type 2 diabetic subjects instructed to wear a Freestyle Libre Continuous Glucose Monitor (CGM) and Heart Rate Monitor (HRM) and logged foods for 10 days. A novel biophysical, machine-learned model was trained on CGM, HRM data and food logs to predict and tune glucose responses to food and activity. The training used a separate deep-learning network trained to predict the infusion of glucose into subject bloodstream using known glycemic index (GI) values.

Results: 715 nondiabetic, 35 prediabetic, and 246 T2D subjects were enrolled. Our novel biophysical, machine-learned model predicted glucose excursions over a database of many million foods and activity rates and exceeded 99.9% on the Clarke Error Grid regions A and B for a wide variety of individuals. Many correlations exist between caloric intake and composition, fiber intake, household income, sleep quality, and inferred physiological parameters.

Conclusion: Precise prediction of glycemic responses will allow accurate counterfactual analysis of lifestyle choices and allow all individuals to better plan and select foods and activities to match their lifestyle, physiology, and psychological predispositions. This novel technology has potential novel applications for both diabetes treatment and prevention.