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

Area of Science:

• Metabolic Health
• Endocrinology
• Biomedical Data Science

Background:

• Current Type 2 diabetes (T2D) and prediabetes classification relies on fasting glucose or HbA1c, overlooking underlying pathophysiological heterogeneity.
• Understanding metabolic subphenotypes (muscle/hepatic insulin resistance, β-cell dysfunction, impaired incretin action) is crucial for targeted diabetes treatment and prevention.
• Existing diagnostic methods do not fully capture the diverse mechanisms contributing to glucose dysregulation.

Purpose of the Study:

• To identify and characterize distinct metabolic subphenotypes contributing to early glucose dysregulation and T2D risk.
• To develop and validate a machine learning framework for predicting these subphenotypes using glucose time-series data.
• To assess the utility of at-home continuous glucose monitoring (CGM) for identifying metabolic subphenotypes and stratifying risk.

Main Methods:

• Gold-standard metabolic tests were conducted in individuals with early glucose dysregulation.
• A machine learning framework was developed to predict metabolic subphenotypes from oral glucose tolerance test (OGTT) glucose time-series data.
• Predictions were validated using an independent cohort and tested with CGM data from at-home OGTTs.

Main Results:

• Substantial inter-individual heterogeneity in metabolic subphenotypes was revealed, with significant proportions exhibiting dominant muscle/liver insulin resistance or β-cell/incretin deficiency.
• The machine learning model accurately predicted insulin resistance (auROC 95%), β-cell deficiency (auROC 89%), and incretin defect (auROC 88%) from OGTT glucose curves.
• At-home CGM data successfully predicted muscle insulin resistance (auROC 88%) and β-cell deficiency (auROC 84%), demonstrating clinical applicability.

Conclusions:

• The prediabetic state is characterized by significant metabolic heterogeneity that can be defined by glucose curve dynamics during OGTT.
• Machine learning analysis of glucose curve shapes offers a superior method for identifying specific metabolic defects compared to current estimates.
• At-home CGM provides a practical and scalable approach to risk stratify individuals and guide targeted interventions for T2D prevention.