The glucose management indicator (GMI) estimates HbA1c from continuous glucose monitoring (CGM) profiles. The formula was developed with real-time CGM (rtCGM) sensors (1). As previous data indicated discrepancies between observed laboratory HbA1c and CGM-derived estimates of HbA1c for some individuals and sensor types (1,2), we aimed to compare observed HbA1c and GMI using both rtCGM and intermittent scanning CGM (iscCGM) profiles collected during routine care in people with type 1 diabetes.
We analyzed 132,361 CGM days from a total of 1,973 individuals with type 1 diabetes for ≥1 year from the German/Austrian/Swiss/Luxembourgian Prospective Diabetes Follow-up Registry (DPV) (3). As measurement ranges of the CGM devices differed, we truncated glucose values to the same range (40–400 mg/dL). We calculated the GMI from up to 90 CGM days per individual (median [interquartile range] 77 [46–89] days per individual) as GMI (%) = 3.31 + 0.02392 ⋅ [mean glucose (mg/dL)] (1). We compared GMI and observed HbA1c at the end of the 90-day period overall and by age-group. Absolute differences between GMI and observed HbA1c were illustrated for rtCGM vs. iscCGM and stratified by glucose variability, normal weight vs. overweight, and HbA1c <7.5% vs. ≥7.5% using boxplots. Low/high glucose variability was defined as coefficient of variation (CV = SD divided by the mean) <36/≥36%. Overweight was defined based on German Health Interview and Examination Survey for Children and Adolescents (KiGGs) reference as >90th percentile for individuals aged <18 years and as BMI >25 kg/m2 for adults.
Median age and duration of type 1 diabetes were 14 [10, 17] and 5 [3, 9] years, respectively; 52% (n = 1,031) were boys/men, and 68% (n = 1,329) were pump users. Mean (± SD) GMI estimated from CGM data were slightly higher than mean observed HbA1c in the overall cohort (7.8 ± 0.9% vs. 7.6 ± 1.2%) and within each age-group (<6 years: 7.5 ± 0.7% vs. 7.2 ± 1.0%, 6 to <12 years: 7.7 ± 0.7% vs. 7.3 ± 0.9%, 12 to <18 years: 8.0 ± 0.9% vs. 7.8 ± 1.3%, ≥18 years: 7.7 ± 1.0% vs. 7.6 ± 1.3%). Overall, 11% (n = 224) had an absolute difference between GMI and observed HbA1c of less than ± 0.1%, and 46% (n = 910) had an absolute difference of at least ± 0.5%.
Stratification by sensor type revealed that mean GMI and HbA1c were similar in rtCGM users (n = 341 CareLink Pro/Personal, n = 64 Dexcom G5; 7.6 ± 0.7% vs. 7.6 ± 1.1%), whereas iscCGM users (n = 1,568 FreeStyle Libre) had higher mean GMI than HbA1c (7.9 ± 0.9% vs. 7.6 ± 1.2%). Similar patterns were observed after stratification by glucose variability or weight (Fig. 1): while absolute differences between GMI and laboratory HbA1c were almost symmetrically distributed around 0 in rtCGM users, GMI was higher than the laboratory HbA1c value in almost three-fourths of iscCGM users. This finding was still observed after additional stratification by age-group (data not shown). Stratification by HbA1c level revealed higher GMI than HbA1c in most individuals with HbA1c <7.5%, with GMI-HbA1c differences being higher in iscCGM users. In contrast, three-fourths of all rtCGM users with HbA1c ≥7.5% had lower GMI than HbA1c, whereas differences in iscCGM users with HbA1c ≥7.5% were almost symmetrically distributed around 0 (Fig. 1).
Difference between CGM-derived estimated HbA1c and observed laboratory HbA1c. Difference between laboratory HbA1c and GMI by sensor type, stratified by age-group, glucose variability, and HbA1c (based on truncated sensor values). Boxes represent the interquartile range (IQR = Q3 − Q1), whiskers represent the 1.5-fold IQR, and dots represent 5th and 95th percentiles. Solid and dashed lines within boxes indicate median and mean, respectively.
Difference between CGM-derived estimated HbA1c and observed laboratory HbA1c. Difference between laboratory HbA1c and GMI by sensor type, stratified by age-group, glucose variability, and HbA1c (based on truncated sensor values). Boxes represent the interquartile range (IQR = Q3 − Q1), whiskers represent the 1.5-fold IQR, and dots represent 5th and 95th percentiles. Solid and dashed lines within boxes indicate median and mean, respectively.
In conclusion, our analysis of glucose profiles collected during routine care from people with type 1 diabetes revealed discrepancies between CGM-derived GMI and laboratory HbA1c in a considerable subset of individuals. This finding is in line with results from Leelarathna et al., who analyzed data from three different rtCGM sensors. They found an absolute difference between GMI and laboratory HbA1c of ≥0.5% in one-third of all participants but found that only 20% of all participants had differences of <0.1%. It has been suggested that the difference between laboratory HbA1c and CGM-derived GMI is clinically meaningful and should be considered when therapeutic goals are set (1).
Our analysis, however, found that discrepancies between GMI and HbA1c differed between iscCGM and rtCGM. Measurement ranges as well as distribution of sensor glucose values differed by sensor type. CGM systems typically have a higher accuracy in the euglycemic range, whereas accuracy often decreases in the hypoglycemic and/or hyperglycemic range (4). Moreover, different modes of calibration lead to different sensitivities and specificities in the detection of biochemical hypoglycemia (5). This indicates that it is necessary to adjust the GMI formula for sensor type.
As both measures, CGM-derived GMI and laboratory HbA1c, may affect therapeutic goals and diabetes management, further research is needed to examine potential explanatory factors, for example, duration of CGM usage, indication for usage, sensor type/version, ethnicity, subcutaneous adipose tissue, glucose variability, life span of red blood cells, or hemoglobinopathies that may affect glycosylation. Formulas specific for sensor type that account for the respective measurement range and type of calibration may be needed to accurately estimate GMI.
Article Information
Acknowledgments. The authors thank all individuals who contribute to the DPV database as well as the numerous participating DPV centers in Germany, Austria, Switzerland, and Luxembourg. Special thanks to A. Hungele and R. Ranz for support and development of the DPV documentation software and to K. Fink and E. Bollow for the DPV data management (all clinical data managers, Ulm University).
Funding. This work was supported by the German Center for Diabetes Research (DZD) funded by the Federal Ministry of Education and Research (grant no. 82DZD0017G) and by a grant from the German Diabetes Association (DDG).
The funders had no role in study design, data collection and analysis, preparation of the manuscript, or decision to publish.
Duality of Interest. Additional funding was provided by Abbott and Sanofi. J.M.G. and R.W.H. report grants to their institution from Abbott, Lilly, Sanofi, Boehringer Ingelheim, and Novo Nordisk, outside the submitted work. S.v.S. is a consultant for Abbott, Dexcom, Lilly, Medtronic, and Novo Nordisk. She received speakers honoraria from Abbott, Berlin-Chemie, Lilly, Merck-Serono, Medtronic, and Novo Nordisk. T.D. reports grants and personal fees from AstraZeneca, Boehringer Ingelheim, Lilly, Dexcom, Medtronic, Novo Nordisk, Sanofi, and Ypsomed, outside the submitted work, and is a shareholder in DreaMed, Ltd. No other potential conflicts of interest relevant to this article were reported.
Author Contributions. J.M.G. performed the statistical analysis, performed literature research, and wrote and edited the manuscript. S.v.S., M.F., U.E., K.P., T.D., E.H., and R.W.H. researched data, contributed to data interpretation and discussion, and reviewed and edited the manuscript. All authors reviewed and approved the final manuscript. R.W.H. is the guarantor of this work and, as such, had full access to the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.
Prior Presentation. Parts of this work were presented at Advanced Technologies & Treatments for Diabetes (ATTD), Madrid, Spain, 19–22 February 2020.
