A MACHINE LEARNING ADAPTIVE CLASSIFIER FOR PREDIABETES AND TYPE 2 DIABETES DIAGNOSIS BASED ON CONTINUOUS GLUCOSE MONITORING DATA

Background and Aims: Continuous Glucose Monitoring (CGM) devices have become the standard for glucose monitoring in diabetes. Beyond glucose tracking, CGM data offers a unique opportunity to support diagnostic applications for prediabetes (PD) and type 2 diabetes (T2D). This study explores adaptive real-time Machine Learning (ML) algorithms for diabetes classification solely based on CGM data. The goal is to provide a less invasive and more cost-effective alternative to traditional diagnostic tests, while keeping limited the number of monitoring days required for accurate classification. Methods: CGM data collected from 325 insulin-naïve subjects using Dexcom G4 and G6 sensors for up to 10 days of monitoring were used to extract demographics (e.g., age, height, weight …) and CGM-derived glycemic features (e.g., mean, SD, GMI, …). The resulting dataset was used to train and evaluate a cascade of two binary logistic regression classifiers: the first discriminating between healthy vs. pathological subjects, the second distinguishing pathological subjects between people with PD and T2D. After each day, the classifier formulates a probable diagnosis if the assigned classification probability exceeds a predefined threshold. Otherwise, the algorithm incorporates the next day’s data iteratively until a decision is reached. Results: The proposed classifier, evaluated on an independent test set, achieved a balanced accuracy of 0.69 and a weighted F1-score of 0.78. Notably, approximately 70% of traces were classified using only one day of CGM data. Conclusions: This study provides promising evidence for CGM use to diagnose PD and T2D. Future work will compare our approach with more sophisticated data-driven methodologies based on artificial intelligence.