An Explainable Machine Learning Model that Uses Continuous Glucose Monitoring to Screen for Hypertension in Patients with Type 1 Diabetes

Background Patients with Type 1 diabetes (T1D) routinely wear continuous glucose monitoring (CGM) devices for glycemic management. Beyond glucose tracking, CGM data may contain patterns associated with comorbidity risk. Objective We hypothesized that CGM metrics could enhance classification of hypertension when added to standard biochemical and demographic data. Methods We analyzed data from 689 T1D patients with a median of 286 days of CGM monitoring. We engineered 36 CGM-derived features including time in range quartiles, glycemic variability metrics, and temporal patterns. Using an XGBoost model with 10-fold cross-validation and SMOTE for class imbalance, we compared biochemical-and-demographic versus combined (biochemical + demographic + CGM) models for classifying seven diagnosed comorbidities: hypertension, hypothyroidism, retinopathy, lipid metabolism disorders, airway disease, nephropathy, and neuropathy. Results CGM data significantly improved classification of diagnosed hypertension (biochemical-and-demographic ROC-AUC: 0.748 ± 0.092; combined ROC-AUC: 0.770 ± 0.088; ΔROC-AUC = + 0.022, P  = 0.0191 via DeLong test). CGM features such as TBR, CV, rolling mean, and slope of glucose emerged among the top 15 discriminators in the SHAP analysis. No significant improvements were observed for other comorbidities in terms of ROC-AUC. Decision curve analysis confirmed the net clinical benefit of the combined model across threshold probabilities consistent with the population prevalence. Conclusions CGM data provides significant added value for hypertension risk classification in T1D patients without additional patient burden. This represents an opportunistic screening approach leveraging existing monitoring infrastructure.