Machine Learning Diagnostic Models of Prediabetes and Type 2 Diabetes using Continuous Glucose Monitoring data

Background and Aims: increasing evidence supports continuous glucose monitoring (CGM) sensors adoption to monitor glycemia in populations beyond people with type 1 diabetes. Interestingly, analyzing collected data, CGM offers a cheap and minimally invasive way to diagnose states of dysglycemia as prediabetes and type 2 diabetes (T2D). To this aim, in this work we developed new machine learning tools to classify prediabetes and T2D from CGM data. Methods: the dataset comprised ten days of glucose measurements recorded using Dexcom G4 and G6 CGM sensors from 325 individuals, including healthy subjects and people diagnosed with prediabetes or T2D, based on measured A1c levels. None of the participants were under pharmacotherapy. CGM-derived features including glucose variability indices, time in ranges and time series complexity metrics, and demographics were extracted as feature set. This dataset was then split in a stratified 80-20 ratio for training and testing. We built a linear regression model with a posteriori classification (model A) and a cascade of two binary logistic regression classifiers (model B) to predict whether subjects were healthy, prediabetic, or T2D. For both models, feature selection was performed using Variance Inflation Factor (VIF) and Recursive Feature Elimination (RFE). Results: both models demonstrated satisfactory performance. Model A achieved balanced accuracy of 0.72 (VIF) and 0.69 (RFE), while Model B scored 0.69 (VIF) and 0.71 (RFE). In terms of weighted F1-score, Model A obtained 0.76 (VIF) and 0.74 (RFE), while Model B achieved 0.74 (VIF) and 0.76 (RFE). Conclusions: this study presents promising results regarding the use of CGM sensors for diagnosing prediabetes and T2D.