The Glucose Management Indicator (GMI) is the accepted method to estimate HbA1c from mean CGM glucose. We hypothesize that HbA1c estimation could be improved with machine learning (ML) methods that account for additional covariates.
CGM, HbA1c, and demographic data were aggregated from five studies (3,550 HbA1c measurements, 1,095 participants). For each HbA1c, the following statistics were derived from all CGM data recorded up to 30 days before and 5 days after measuring HbA1c: mean, SD, CV, and % time in various ranges. Using HbA1c as the outcome and all CGM glucose statistics and demographics as covariates, the ML models LASSO, LASSO with all two-way interactions, and random forest were tuned via 5-fold cross validation (CV). Top performing models were compared to GMI via 5-fold CV and the paired Wilcoxon signed-rank test.
Random forest performed best, with 20% lower estimation error than GMI (p < 0.001). After restricting HbA1c measurements to those for which an additional HbA1c had been measured for the same patient at least 70 days prior (1,833 HbA1c measurements, 755 participants), the models were re-trained with prior HbA1c as a covariate. LASSO performed best, with 27% lower estimation error than GMI (p < 0.001). An ordinary least squares (OLS) model accounting for mean glucose, race, and prior HbA1c had an estimation error only 0.3% higher than LASSO, suggesting that augmenting GMI with just two additional covariates could substantially improve estimation.
J. Grossman: None. A.T. Ward: None. D.M. Maahs: Advisory Panel; Self; Eli Lilly and Company, Insulet Corporation, Medtronic, Novo Nordisk A/S. Consultant; Self; Abbott, Sanofi. Research Support; Self; Bigfoot Biomedical, Dexcom, Inc., Roche Diabetes Care, Tandem Diabetes Care. P. Prahalad: None. D. Scheinker: None.