Machine learning analysis of year-long continuous glucose monitoring data identified month-to-month variability in time spent within target glucose ranges as a key marker associated with concurrent diabetic kidney disease in patients with type 1 diabetes, according to a study published in Diabetes, Obesity and Metabolism. 

Researchers analyzed cross-sectional data from a prospective observational cohort of 282 adults with type 1 diabetes who had 12 consecutive months of personal continuous glucose monitoring (CGM) data collected through the Korean National Home Care Pilot Program. Diabetic kidney disease (DKD) was defined as a urine albumin-to-creatinine ratio of at least 30 mg/g or an estimated glomerular filtration rate below 60 mL/min/1.73 m², with the abnormality documented on at least two occasions during the one-year observation period. Forty patients met criteria for DKD and 242 did not. 

The investigators developed four machine learning models: light gradient boosting machine (LightGBM), extreme gradient boosting (XGBoost), random forest, and logistic regression. The models used glucose management indicator (GMI) features, broader CGM-derived features, clinical covariates, or combinations of these variables. They evaluated whether longitudinal CGM metrics improved discrimination of concurrent DKD beyond GMI alone. Feature importance was evaluated using Shapley Additive Explanations (SHAP). 

LightGBM was the best-performing model overall when combining CGM-derived metrics with clinical covariates, achieving an area under the receiver operating characteristic curve (AUROC) of 0.91 and an F1 score of 0.65. All three tree-based models (LightGBM, XGBoost, and random forest) outperformed logistic regression across reported performance measures. The authors said this finding supported the use of machine learning to identify nonlinear glycemic patterns that conventional linear approaches may not capture. 

The standard deviations of monthly time in range (TIR) and time in tight range (TITR) were the most influential CGM-derived features in the LightGBM model, ranking behind only hypertension and duration of diabetes. For each patient, the standard deviation reflected how much TIR or TITR varied across the 12 monthly measurements, with higher values indicating greater month-to-month instability. 

In contrast, the conventional coefficient of variation of sensor glucose did not differ significantly between patients with and without DKD. SHAP analysis showed that greater month-to-month variability in both TIR and TITR was associated with a higher predicted probability of concurrent DKD, with TITR exhibiting a more consistent positive relationship. 

Among 3,384 monthly CGM observations—12 for each of the 282 patients—median TIR was 65% among patients with DKD vs 69% among those without DKD; median TITR was 38% vs 44%, respectively. Patients with DKD had 44% greater variability in monthly TIR and 23% greater variability in monthly TITR, based on the standard deviation of 12 monthly measurements. Mean glycated hemoglobin (HbA1c) at baseline did not differ significantly between groups, nor did the overall coefficient of variation of sensor glucose. 

The investigators also evaluated whether longitudinal CGM metrics improved classification beyond GMI alone. Compared with GMI-only models, adding broader CGM-derived features produced a 22% net reclassification improvement with XGBoost. The corresponding improvements were 18% with random forest and 17% with LightGBM, although neither met the study’s threshold for statistical significance. 

A subgroup analysis of 18 patients with early DKD—defined as albuminuria with an estimated glomerular filtration rate of at least 60 mL/min/1.73 m²—mean GMI, TIR, and TITR did not differ significantly from values among patients without DKD. However, variability across monthly TIR and TITR measurements remained higher in the early-DKD group. 

Findings remained generally consistent across multiple sensitivity analyses, including repeated feature selection within each cross-validation fold, linear interpolation of missing CGM intervals, adjustment for CGM device type, and exclusion of patients with transient renal laboratory abnormalities.  

The study had several limitations. Its cross-sectional design precluded assessment of temporal or causal relationships between CGM variability and DKD. Because feature selection was initially performed using the full analytical cohort, optimistic estimates of model performance due to information leakage could not be excluded. The study was conducted at a single center, limiting generalizability, and the findings require validation in independent external cohorts. Other limitations included residual differences among CGM devices, difficulty interpreting nonlinear interactions within the high-dimensional models, and heterogeneity introduced by patients with transient renal abnormalities. 

"Our results demonstrate that the SD [standard deviation] of monthly TIR and TITR are the predominant CGM-derived markers associated with DKD status," the authors wrote. Assessment of these measures "may provide additional information for DKD risk stratification beyond conventional mean CGM metrics," they added.

The researchers declared no conflicts of interest.