Estimating Glycemia From HbA_1c and CGM: Analysis of Accuracy and Sources of Discrepancy

Regular monitoring of blood glucose levels is crucial to achieve glycemic control, but precisely measuring blood glucose over an extended period is challenging. Both hemoglobin A_1c (HbA_1c) and continuous glucose monitors (CGM) can be used to estimate average blood glucose levels over time.

HbA_1c has been shown to be associated with risk of complications (1,2), and HbA_1c measurement has the advantage of providing a retrospective estimate of glycemia from a single nonfasting blood sample. However, nonglycemic factors, such as acute or chronic alterations in the mean age and turnover of the red blood cell (RBC) population, may affect HbA_1c. While this effect is readily appreciated in cases of acute blood loss or hemolysis, individual differences in RBC life span may be stable over time and affect HbA_1c. The reliability of HbA_1c measurements has been shown to be improved by incorporating an estimate of mean RBC age (M_RBC) (3).

CGM can also be used to estimate average blood glucose levels over time and has proven to be a valuable tool to monitor diabetes (4,5). CGM-derived statistics reflect glycemia during a monitoring period, the length of which can vary substantially. A common CGM monitoring period in clinical practice and in many clinical studies is 10–14 days (6). Most CGM devices use proprietary conversion formulas to estimate blood glucose concentration from measurements of interstitial fluid glucose concentration, but standards for CGM device quality control, traceability, and performance are still in development (7).

HbA_1c and CGM can produce inaccurate estimates of true glycemia for several reasons (Fig. 1), leading to confusion about how best to estimate individuals’ glycemia (8–10). In this study, we compared the ability of different CGM time periods and HbA_1c to estimate mean glycemia accurately over 90 days (AG₉₀), the time period typically reflected by a single HbA_1c measurement. Since there is no reference laboratory-based measurement of AG₉₀, we estimated it (eAG₉₀) using correction methods described in Table 1 (11,12). We analyzed sources of inaccuracy in HbA_1c-based (eAG_A1c) and CGM-based (eAG_CGM) estimates of eAG₉₀: HbA_1c assay analytic errors, CGM sensor bias, CGM period lengths, and M_RBC variation as a proxy for nonglycemic effects. We also investigated whether CGM use over limited monitoring periods and HbA_1c could be combined to improve accuracy.

We reviewed published studies and reports that included measurements of imprecision and bias for HbA_1c assays and CGM sensors. HbA_1c assay error was estimated with the mean coefficient of variation reported for NGSP-certified HbA_1c assays of 2.48% (in dimensionless units, see Supplementary Table 1), with a mean absolute assay bias of 0.1% (in NGSP HbA_1c units) (13). We estimated a representative CGM sensor bias by averaging measurements from published studies of CGM sensor bias. See Supplementary Table 2 and Supplementary Methods for details and citations. Imprecision of single CGM measurements has a negligible effect (typically <0.5 mg/dL) on CGM-based estimates of glycemia when thousands of distinct CGM measurements are averaged, and we did not include it as a consideration in this analysis.

We retrospectively analyzed CGM and HbA_1c results for 315 individuals with diabetes who were regular CGM users at Massachusetts General Hospital (Table 2). We included participants with ≥90 consecutive days of CGM data obtained with the Dexcom G5 or G6, and no more than 10% missing data, and with paired HbA_1c measurements at the end of the 90 days. HbA_1c was measured with an antibody-based assay (Roche cobas c 501) or a boronate affinity method (Trinity Biotech Premier), the results of which were comparable (<0.2% mean absolute difference) based on split samples assayed at regular intervals. Each method is relatively unaffected by hemoglobinopathies and interfering factors (1) and has high precision (coefficients of variation <2% in the Massachusetts General Hospital laboratories). The research was approved by the Mass General Brigham Institutional Review Board, which waived the requirement for informed consent due to minimal risk.

For the analysis of CGM accuracy, eAG₉₀ was estimated as the time-weighted average over 90 days of CGM with the addition of sensor bias. CGM monitoring periods were analyzed with use of all available CGM data, and CGM-based estimates of glycemia (eAG_CGM) were determined for shorter intervals with time-weighted averaging and then compared with eAG₉₀ for determination of accuracy, with mean absolute error (95th percentile) reported in milligrams per deciliter of glucose (Table 1).

The HbA_1c-based point estimate of glycemia (eAG_A1c) was computed with a regression equation derived in the A1c-Derived Average Glucose (ADAG) study (14) (Table 1).

Quantifying and analyzing the impact of all hypothesized nonglycemic factors on the error of eAG_A1c is not feasible, but M_RBC has been shown to be a major contributor to these nonglycemic effects and serves as a useful proxy for understanding the contribution of nonglycemic factors to errors in estimation of glycemia (3,15). M_RBC effects on HbA_1c were estimated with use of a published mechanistic model (3) of the relationship of HbA_1c, average glucose, and M_RBC. For the analysis of eAG_A1c accuracy, mean true glycemia over 90 days (eAG₉₀) was estimated with use of theoretical models that assume M_RBC is known (Table 1). The estimated theoretical errors were compared with empirical estimates provided by estimators of variance derived from data in the ADAG study and other published studies (14,16). See Supplementary Methods for more detail.

The data that support the findings of this study are available from the Massachusetts General Hospital, but restrictions on sharing apply to maintain patient anonymity. Summary data are available from the authors on reasonable request and with permission of the Mass General Brigham Institutional Review Board.

We estimated that HbA_1c measurements between 5.7% and 7.0% (NGSP % units) have an expected absolute assay error (95th percentile) of 0.1% (0.3%), which corresponds to an expected absolute error in estimating eAG₉₀ of 3 (8–10) mg/dL. For CGM, the representative absolute CGM sensor bias (95th percentile) of 3 (10) mg/dL implies that an individual who monitored CGM over a full 90-day period using at least six different sensors with uncorrelated biases should expect that the eAG_CGM calculation would have an absolute error of ∼2 (7) mg/dL in estimating eAG₉₀. This analysis does not include consideration of possible CGM device malfunction or delays (17).

eAG_CGM differs from eAG₉₀ less as the periods become longer (Fig. 2A). A common monitoring length of 14 days at the end of the 90-day period yielded an eAG_CGM that differed from eAG₉₀ with a mean absolute error (95th percentile) of 14 (34) mg/dL (Fig. 2A). Thus, individuals with true mean glycemia of 154 mg/dL could have an eAG_CGM of 120–188 mg/dL (95% CI). Biological variation in glycemia may also cause distinct monitoring intervals of the same length but at different times to result in different eAG_CGM. eAG_CGM calculated for two nonoverlapping 14-day intervals within the same overall 90-day interval had a mean variation (95th percentile) of 9% (20%) (Fig. 2B). For example, when comparing an eAG_CGM of 168 mg/dL calculated from 14 days of CGM with that calculated from a distinct nonoverlapping 14-day period within the same 90-day interval, the second 14-day eAG_CGM is expected to range from 134 to 202 mg/dL (95% CI). While no systematic difference was found in estimating AG₉₀ using different 14-day intervals across the 90-day CGM duration (Supplementary Fig. 1), averaging eAG_CGM from two nonoverlapping 14-day periods reduced the error, with the lowest mean absolute error (95th percentile) of 6 (15) mg/dL obtained by combining the monitoring intervals between 42 and 56 and 14 and 28 days before the end of the 90-day interval (Supplementary Table 6). When a significant percentage of CGM data during the monitoring interval is missing or when comparing shorter periods of time, these inaccuracies increase (see Supplementary Figs. 2–4).

The impact of nonglycemic effects (conceptualized and modeled as M_RBC) is that similar eAG₉₀ levels will correspond to different HbA_1c values (Fig. 2C). Common regression formulas such as the ADAG study empirical regression line will reflect the specific M_RBC and average glucose joint distribution of the population in which they were inferred. For example, when AG₉₀ is ∼125 mg/dL, the ADAG formula is most accurate for individuals with M_RBC at ∼50 days, and for AG₉₀ ∼200 mg/dL the ADAG formula is most accurate for individuals whose M_RBC is ∼53 days. This pattern likely reflects nonrandom associations between M_RBC and AG₉₀ in the ADAG study cohort, and using this formula to map HbA_1c to eAG_A1c in different cohorts might lead to inaccuracies. If we model the theoretical effects of M_RBC we obtain a mean absolute error (95th percentile) in estimating eAG₉₀ of 10 (24), 10 (25), and 12 (29) mg/dL for HbA_1c 5.7%, 6.5%, and 7.0% respectively, corresponding to 117, 140, and 154 mg/dL. Thus, individuals with true mean glycemia of 140 mg/dL can expect to have eAG_A1c between 115 and 165 mg/dL (95% CI). These theoretical errors were compared with errors estimated using published data from the ADAG study and another empirical study where HbA_1c and average glucose were both measured directly (14,16) (Fig. 2D). The similarity between the CIs helps validate the assumption that most HbA_1c variability can be attributed to RBC life span variation and any other highly correlated nonglycemic factors. Additional detail on nonglycemic effects in empirical studies can be found in Supplementary Results and Supplementary Fig. 6.

Figure 2E shows a comparison of the accuracy of the average of eAG_A1c and eAG_CGM with the accuracy of either eAG_A1c or eAG_CGM alone, as a function of the length of the CGM monitoring period. The average of eAG_A1c and eAG_CGM calculated from 3, 10, and 14 days of CGM at the end of the 90-day period differed from eAG₉₀ with a mean absolute error (95th percentile) of 15 (36), 11 (27), and 10 (26) mg/dL respectively. When the CGM monitoring period was <26 days (95% CI 21, 28), the average of these two different measurement methods had a lower mean absolute error as well as a significantly reduced frequency of very large errors. The empirical mean absolute error (95th percentile) for eAG_A1c shown in Fig. 2E (16 [40] mg/dL) is larger than what was found with the theoretical analysis above (12 [29] mg/dL) or in prior empirical studies (12 [31] mg/dL) (14,16) and may be caused by differences between the M_RBC distribution in our cohort and that in the ADAG equation derivation cohort and also by intrapatient fluctuation in M_RBC. See Supplementary Results and Supplementary Figs. 8 and 9 for more detail.

HbA_1c and CGM are both used to estimate average glycemia for people with diabetes, but eAG_A1c and eAG_CGM can differ, with inaccuracies arising from assay error for both measurements, limited duration of CGM monitoring, and nonglycemic factors for HbA_1c. Understanding the sources and magnitude of inaccuracies can help clarify how to determine the most accurate estimate of glycemia given available data.

The mean errors (95th percentile) in eAG_CGM from 14 days of CGM and in eAG_A1c due to typical assay error and normal biological variation are estimated to be similar (14 [34] mg/dL for eAG_CGM and 12 [29] mg/dL for eAG_A1c). For an individual with true mean glycemia of 167 mg/dL there is more than a 5% chance that the mismatch will be >40 mg/dL (∼1.3% in NGSP HbA_1c units). Discrepancies may thus be larger than the difference between common thresholds for the diagnosis of prediabetes and diabetes (e.g., HbA_1c 5.7% and 6.5% differing by 0.8%, the equivalent of 26 mg/dL).

The errors in estimating eAG_A1c and eAG_CGM are likely independent of each other, and averaging them is therefore typically beneficial when CGM data for fewer than ∼26 days (95% CI 21, 28) are available. When more CGM data are available, eAG_CGM is expected to be more accurate on its own assuming sensor bias is low and data are not missing in a systematic way. Longer CGM periods will consistently reduce the error, even if nonoverlapping periods are combined. Previous studies showing that a 4-week CGM period leads to robust CGM metrics (18) are consistent with our results, and longer periods are needed in the setting of data missing not at random (19). In Supplementary Results, we provide specific scenarios that are often seen in clinical practice and offer further suggestions on when to use CGM and when HbA_1c depending on the context.

We found that theoretical models of HbA_1c variability due to M_RBC corresponded to published empirical data (16) (Fig. 2D), supporting the validity of the assumption that M_RBC can explain the majority of nonglycemic variation in HbA_1c in real world settings and that no significant bias is present in our assumptions about AG₉₀. Our estimation of nonglycemic effects (M_RBC) relies on a commonly used regression equation from the ADAG study (12). This and other commonly used regression equations implicitly reflect the joint M_RBC and AG₉₀ distribution of their derivation cohorts. Individuals with a combination of M_RBC and AG₉₀ that is not well represented in the relevant derivation cohort will have more inaccurate glycemia estimates. Similarly, differences between published empirical regression formulas can be explained by differences in this joint distribution in the derivation cohorts. When eAG_A1c and eAG_CGM differ consistently by even greater amounts than expected for a particular individual, there may be a stable nonglycemic source of inaccuracy, like an M_RBC that differs significantly from that assumed by the regression equation used to generate the HbA_1c-based estimate of glycemia, in which case it would be more accurate to use extended CGM periods for estimation of glycemia, or to infer novel regression equations in the population of interest to reduce the impact of nonglycemic factors. If this situation is not recognized, use of HbA_1c to guide management may lead to chronic under- or overtreatment (for further details see Supplementary Results and Supplementary Fig. 5).

This study contributes to the growing literature disentangling the relationship between CGM-based statistics and HbA_1c (4,9,20–25). First, means and 95% CIs are provided for expected errors in eAG_A1c and eAG_CGM (Fig. 2 and Supplementary Table 6). Second, the analysis minimizes the effects of biases due to short-term behavioral changes or to the reported effective therapeutic benefit of CGM (26–29) by focusing on CGM data from individuals with diabetes who were dedicated CGM users. Third, estimates are provided for the impact of M_RBC as a surrogate for nonglycemic effects on HbA_1c, with comparison of theoretical and empirical analysis of this impact. Fourth, this study can help explain the frequent and significant mismatches between HbA_1c and the glucose management indicator (GMI) (23–25,30). GMI is calculated from an individual’s current eAG_CGM and is thus affected by CGM assay errors, monitoring duration, and missing data. Additionally, the GMI equation was derived from eAG_CGM and HbA_1c measurements and thus GMI assessment will be inaccurate for individuals with combinations of M_RBC and AG₉₀ that were not well represented in the GMI derivation cohort (see Supplementary Results for detail).

The analysis has limitations, and the conclusions may not generalize to some patient cohorts. First, there is no available gold standard measurement of glycemia over 90 days, and our study relies on correction methods (11,12). In particular, eAG₉₀ used for CGM analysis relies on estimates of sensor bias. Data on CGM sensor bias are very limited, and errors may depend on brand, model, batch, and placement. Results of the sensitivity analysis of the empirical sensor bias data used to derive the representative sensor bias in this study (Supplementary Table 2) suggest that conclusions are robust (Supplementary Fig. 10), but if real-world sensor bias is different from what was assumed in this study, it may impact overall error estimates and suggestions for decision-making (31). Some recent studies (32,33) suggest that substantially greater sensor biases are possible under some conditions, and if sensor bias exceeds ∼7 mg/dL, averaging HbA_1c and CGM may be more accurate no matter how much CGM is available. It will thus be important to update this analysis when larger and more accurate data sets reporting CGM sensor bias become available. Second, the empirical analysis focused on eAG_CGM from individuals with diabetes who used CGM extensively (>90 days with <10% missing data), and these individuals may differ in important ways from the average person with diabetes. For instance, they may have less variable glycemia during both monitored and unmonitored periods, yielding eAG_CGM errors lower than would be expected for a randomly selected CGM user. Third, we cannot account for the effect of systematic difference in glycemia during times when CGM data are missing, which may have a significant impact on the reliability of eAG_CGM. Identification of real-world patterns of glycemia when data are missing is not currently feasible. Fourth, averaging nonoverlapping periods of CGM will typically lead to a lower error; however, the optimal 14-day periods found in this study may not generalize to other cohorts. Fifth, CGM devices are not standardized (7), and biases for specific devices may be different from what was assumed in this study. Sixth, we do not account for nonglycemic effects that may act independent of M_RBC, and the use of other mechanistic models might be required to further investigate these independent effects. These results do not apply to individuals with hemoglobin variants and will be inaccurate for individuals with rapidly changing hematologic states, e.g., in the setting of hemolysis or transfusion. Lastly, our findings, including the 26-day threshold below which averaging eAG_A1c and eAG_CGM is expected to reduce error, need to be further validated in independent cohorts to increase confidence that they are unbiased and not due to intrinsic fluctuations of our study data.

The findings of this study should assist clinicians in interpreting CGM and HbA_1c and understanding and explaining discrepancies. This study may also help investigators evaluate the reliability of new devices or assays and plan informative clinical trials that involve monitoring glycemia.

Funding. V.T., R.M.C., S.O.O., and J.M.H. were supported by National Institutes of Health (NIH) grant DK123330. V.T., J.M.H., and C.E.P. were supported by NIH grant HD104756.

Duality of Interest. D.J.W. reports serving on data monitoring committees of trials of semaglutide for Novo Nordisk. C.E.P. has received fees and royalties from Mediflix and UpToDate (Wolters Kluwer), respectively, for presentations and articles related to diabetes over which she had full control of content. No other potential conflicts of interest relevant to this article were reported.

Author Contributions. V.T. and J.M.H. conceptualized, designed, and conducted the study as well as interpreted the results. Data collection was performed by V.T., C.M., H.R.P., C.H.P., S.N.H., E.L., and B.A. V.T. and J.M.H. wrote the first draft of the manuscript, and all authors edited, reviewed, and approved the final version of the manuscript. V.T. and J.M.H. are the guarantors of this work and, as such, had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.

C.E.P. is an editor of Diabetes Care but was not involved in any of the decisions regarding review of the manuscript or its acceptance.