From Stability to Variability: Classification of Healthy Individuals, Prediabetes, and Type 2 Diabetes Using Glycemic...

Area of Science:

• Endocrinology
• Metabolic Health
• Data Science in Medicine

Background:

• Glucose control assessment typically relies on HbA1c, but continuous glucose monitoring (CGM) offers dynamic insights.
• Understanding the continuum from normoglycemia to dysglycemia (HbA1c ≥ 5.7%) is crucial for early detection and management.
• Machine learning (ML) presents an opportunity to leverage complex CGM data for improved classification.

Purpose of the Study:

• To investigate glucose control continuum using CGM metrics across healthy, prediabetes, and type 2 diabetes mellitus (T2DM) populations.
• To develop and evaluate ML-based classification models for identifying dysglycemia using CGM data patterns.

Main Methods:

• Pooled CGM data from five studies (n=836: 282 healthy, 133 prediabetes, 432 T2DM).
• Extracted and compared various CGM indices across participant groups.
• Trained and tested XGBoost and Logistic Regression models (70/30 split) to classify dysglycemia versus healthy individuals.

Main Results:

• Significant progressive shifts in CGM indices observed from healthy to T2DM groups (P < 0.001).
• Key differences noted in mean glucose, time below/above range, and complexity indices between healthy and prediabetes groups (P < 0.01).
• XGBoost models achieved high accuracy: AUC 0.91 for prediabetes and 0.97 for dysglycemia identification.

Conclusions:

• CGM metrics demonstrate a gradual decline in glucose homeostasis and increased variability across the normo- to dysglycemia spectrum.
• ML models utilizing CGM data show strong potential for classifying individuals with prediabetes and diabetes.