Predicting Type 2 Diabetes Metabolic Phenotypes Using Continuous Glucose Monitoring and a Machine Learning Framework.

Abstract

The study challenges the conventional classification of type 2 diabetes (T2D) and prediabetes based solely on glycemic levels. Instead, the results highlight the heterogeneity of underlying physiological processes that represent separate pathways to hyperglycemia. Individuals with normoglycemia and prediabetes can be classified according to the relative contribution of four distinct metabolic subphenotypes: insulin resistance, muscle and hepatic, β-cell dysfunction, and incretin defect, which comprise a single dominant or codominant physiologic process in all but 9% of individuals.Use of multiple time points during OGTT generates time-series data to better define the shape of the glucose curve: the application of a novel machine learning framework utilizing features derived from dynamic patterns in glucose time-series data demonstrates high predictive accuracy for identifying metabolic subphenotypes as measured by gold-standard tests in the clinical research unit. This method predicts insulin resistance, β-cell deficiency, and incretin defect better than currently-used estimates, with auROCs of 95%, 89%, and 88%, respectively.The muscle insulin resistance and β-cell deficiency prediction models above were validated with an independent cohort and then tested using glucose data series derived from OGTT performed at home with a continuous glucose monitor (auROC of at-home prediction of insulin resistance and β-cell deficiency is 88% and 84%, respectively). This approach offers a practical and scalable method for metabolic subphenotyping and risk stratification in individuals with normoglycemia or prediabetes, with potential to inform targeted treatments to prevent progression to T2D.