Factors Associated With Discordant A1C-Estimated and Measured Average Glucose Among Hospitalized Patients With Diabetes

This was a retrospective analysis of 17,903 unique adult patients hospitalized at five hospitals within the Johns Hopkins Health System who were discharged between 1 December 2014 and 30 May 2019. We were interested in factors associated with discordance between mAG and eAG at the patient level, so we selected the last admission per patient during the study period, removing all prior admissions to exclude duplicates. The study flowchart is shown in Figure 1. Data were abstracted from charts of hospitalized adults ≥18 years of age who had at least four blood glucose measurements during their hospitalization. Included patients had diabetes based on at least one of the following criteria: International Classification of Diseases, 10th revision, code for diabetes mellitus on their problem list, medical history, or discharge diagnosis; at least two blood glucose measurements ≥200 mg/dL during admission; or an A1C measurement ≥6.5% obtained during the admission. Admissions in which there was no A1C measurement obtained were excluded. The study was approved by the institutional review board of the Johns Hopkins School of Medicine.

Demographics, vital signs, admission and discharge diagnoses, discharge service, laboratory data, diet orders, and medications were extracted from our electronic health record (EHR) system (EpicCare) by Clarity-certified analysts.

A1C was measured from venous samples in the Johns Hopkins Hospital laboratory using the Bio-Rad Variant II hemoglobin testing system. Both serum and POC measurements were used to determine blood glucose levels. The Roche Cobas hexokinase method was used for serum glucose measurements, and POC measurements were performed using the Nova StatStrip glucose meter. The source (venous, arterial, or capillary fingerstick) used for POC measurements was not recorded. POC test results are corrected for hematocrit by the Nova StatStrip device, and there is excellent correlation between the Nova StatStrip and plasma hexokinase methods (27,28), so POC measurements were deemed to be equivalent to serum measurements. POC measurements have also demonstrated accuracy and reliability in critically ill patients (28). The same devices and analytic methods were used at all hospitals in our dataset during the study period.

Exposure variables were selected based on clinical knowledge and published literature. Patient admissions with data missing for key exposure variables were excluded, as shown in the study flowchart in Figure 1. Definitions and data sources for these variables are listed in Supplementary Table S1.

This formula was derived from a weighted combination of continuous glucose monitoring (CGM) and seven-point self-monitoring of blood glucose (SMBG) (26); however, in the hospital setting, blood glucose values are typically measured four times daily (before meals and at bedtime) or every 4–6 hours for patients with NPO status. Thus, this formula was not expected to perfectly correlate with average glucose in the hospital because of differences in sampling frequency.

The primary outcome variable was the glycemic gap, defined as the difference (⁠$Δ$⁠) between mAG and eAG. We defined the glycemic gap in two discordant phenotypes based on the SD of the mean of the difference between the mAG and eAG. Compared with the concordant phenotype (defined as $Δ$ within 1 SD), patients with $Δ$ in the ≤2.5th percentile were determined to have a discordant-low glycemic gap (mAG < eAG; lower-than-expected glucose), and patients with $Δ$ in the ≥97.5th percentile were determined to have a discordant-high glycemic gap (mAG > eAG; higher-than-expected glucose).

Descriptive statistics were used to summarize characteristics of the overall population and to demonstrate differences among the three phenotypic groups (concordant, discordant-low, and discordant-high). Normality of data were assessed using histograms and tests for skewness and kurtosis. All continuous variables were nonnormally distributed, so medians and interquartile ranges (IQRs) are reported.

Univariate logistic regression analyses were performed to identify clinical factors significantly associated with the two discordant phenotypes using the concordant phenotype as the reference group. Covariates identified from the univariate analyses to either have a P value <0.05 or to be established clinical factors known to influence the relationship between A1C and glucose were then included in multivariate logistic regression analyses to identify factors that are independently associated with discordance. All analyses were conducted at the patient level based on the last admission during the study period. Statistical analyses were performed using Stata statistical software, v. 15 (StataCorp., College Station, TX). P <0.05 was considered statistically significant.

Table 1 shows the baseline characteristics of the overall study population and by phenotype of concordance between mAG and eAG. The full cohort consisted of older, overweight, predominantly White patients with a slight male majority. The majority of patients (53.3%) were diagnosed with type 2 diabetes. The median LOS was 5.8 days, and most patients (60.6%) were treated at an academic hospital. There was a high prevalence of hypertension (31.8%). Acute illness or infection were suggested by a high prevalence of tachycardia (55.5%), fever (40.1%), and leukocytosis (median white blood cell [WBC] count 11,600). Reduced renal function (estimated glomerular filtration rate [eGFR] <60 mL/min/1.73 m²) was observed in 18.7% of patients. Patients received a median total daily dose (TDD) of insulin of 5.5 units, and a large percentage of patients (43.3%) were on a carbohydrate-controlled diet.

There were several differences in patient characteristics between the concordant group and each of the two discordant phenotypes. Of note, the discordant-low group (mAG < eAG) consisted of younger, predominantly Black (55.9%) patients with a higher prevalence of type 1 diabetes (24.8 vs. 4.6%) and lower prevalence of hypertension (65.5 vs. 75.5%) and differences in CKD stages. This group was more aggressively treated for diabetes with inpatient insulin (median of mean TDD 32.2 vs. 5.0 units), inpatient metformin (14.5 vs. 8.7%), a carbohydrate-controlled diet (73.2 vs. 42.4%), and insulin use at home (35.3 vs. 26.4%). The discordant-high group (mAG > eAG) had similar demographic characteristics to the concordant group but consisted predominantly of patients without coded diabetes (54.4 vs. 39.8%). LOS was shorter (4.4 vs. 5.9 days), and a smaller proportion of patients (14.1 vs. 25.5%) had undergone surgery. This group was treated more aggressively with inpatient insulin (median of mean TDD 17.3 vs. 5.0 units), home insulin (30.9 vs. 26.4%), inpatient high-dose glucocorticoids (30.0 vs. 8.1%), and home glucocorticoids (9.6 vs. 5.9%).

Figure 2 shows the relationship between mAG and eAG in a fitted line plot. There was a large amount of variation between mAG and eAG, with a coefficient of determination (R²) of 37.3%. Figure 3 shows the discordance between mAG and eAG in a Bland-Altman plot, with $Δ$ plotted against the average of mAG and eAG. There was a negative correlation between the difference and average, suggesting that mAG is higher than eAG when the average inpatient glucose value is lower, whereas mAG is lower than eAG when the average inpatient glucose value is higher. This model suggests that blood glucose values of ∼200 mg/dL have the smallest difference and thus the highest level of concordance.

Figure 4 demonstrates the fully adjusted odds ratios (ORs) for patient characteristics associated with both discordant phenotypes, with results for univariate and multivariable regression analysis provided in Supplementary Tables S2 and S3, respectively. Qualitatively, higher odds of mAG < eAG would be expected to correspond to lower odds of mAG > eAG and vice versa.

The results of the multivariable logistic regression analyses identified several patterns of association between patient characteristics and the discordant phenotypes. These patterns include no association (OR 1 for both discordant-low and discordant-high), single positive or negative association (OR >1 or OR <1 for either phenotype), dual consistent associations (OR >1 for one phenotype and OR <1 for opposite phenotype), or conflicting associations (OR <1 or OR >1 for both phenotypes).

Figure 5 summarizes these qualitative inferences. Lower-than-expected glucose was observed with both type 1 and type 2 diabetes, which were found to have a dual association favoring mAG < eAG. Factors that had a single measure of association observed for this discordant phenotype were Black or other race, increasing LOS, surgery, fever, community hospital, inpatient metformin use, outpatient use of other antihyperglycemic agents, and certain diets (i.e., carbohydrate-controlled, NPO, clear liquid, and unknown diet).

Higher-than-expected glucose was observed with intensive care unit (ICU) service, stage I hypertension, anemia, and high-dose glucocorticoid use in the hospital, which had dual associations favoring mAG > eAG. Factors that had a single measure of association observed for this discordant phenotype were care in an intermediate medical care (IMC) unit, psychiatry or other service, stage II hypertension, BMI, tachycardia, mean albumin, outpatient glucocorticoid use, maximum potassium level, medium-dose glucocorticoid use in the hospital, and use of insulin or insulin secretagogue before admission.

For some patient factors, a conflicting pattern of association was observed with respect to the discordant phenotypes. Age, CKD stage 5, and increasing sodium levels were associated with lower odds of both discordant phenotypes. Increasing insulin TDD was associated with higher odds for both discordant phenotypes.

In this retrospective cohort study using a large EHR database, we identified several patient characteristics associated with discordance between mAG and eAG in hospitalized patients. We found a weak correlation between mAG and eAG and an overall negative glycemic gap (mAG < eAG), which suggests that patients overall have improved glycemic control in the hospital. This observation may be explained by changes in glycemic management at the transition from home to hospital, the impact of glucose sampling frequency on average glucose calculations, conditions that affect A1C accuracy, and other clinical or patient factors that may predispose patients to hypoglycemia. Understanding when an A1C measurement may not accurately predict inpatient glycemic control is important for clinicians because it could affect selection of an initial antihyperglycemic regimen.

Current practice guidelines recommend obtaining an A1C during admission for hospitalized patients if this test has not been performed in the 2–3 months before admission (29). This recommendation is based on previous research showing that A1C levels strongly predict inpatient glycemic control in most hospitalized patients (30). Although A1C is a generally useful indicator of glycemic control, various factors can influence the accuracy of this test. Our study found that, in addition to factors known to influence A1C accuracy (e.g., anemia), several clinical factors were also associated with discordance between A1C-estimated and observed glucose levels.

Lower-than-expected glucose levels were most strongly associated with type 1 and type 2 diabetes compared with no coded diabetes or other diabetes type on the problem list. This finding may be explained by more intensive glucose management when a patient’s care is shifted to the hospital team, as well as relatively lower carbohydrate intake because of a prescribed inpatient diet. It is also possible that under-coding of diabetes could account for delayed recognition in a subset of patients who have diabetes but do not have the problem listed on their problem list, which could result in higher glucose values in those admissions (31).

The association of higher A1C in Black patients is consistent with published literature, but the clinical relevance of this difference is unclear (22). Increasing LOS and fever may reflect sicker patients, in whom hypoglycemia may simply be a marker of severe illness, and surgical patients are often NPO status, which may predispose to relative hypoglycemia.

We found that hospitalization in a community hospital was associated with lower-than-expected glucose values, which could reflect differences in the patient populations served or differences in practice patterns compared with academic hospitals. Because many patients on noninsulin medications at home are transitioned to insulin in the hospital, this practice may explain why patients receiving metformin and other home antihyperglycemic medications had lower-than-expected glucose levels. Furthermore, use of some antihyperglycemic medications (e.g., weekly glucagon-like peptide 1 receptor agonists, ultra-long-acting insulins, and sodium–glucose cotransporter 2 inhibitors) before admission may influence glycemic control because of the long duration of action of these medications. Thus, use of these medications could have contributed to lower inpatient glucose levels as the effects of the medications persisted.

Higher-than-expected glucose levels were most strongly associated with anemia, high-dose glucocorticoid treatment, hypertension, and ICU status. Anemia (i.e., hemolysis; treatment for anemia from iron, vitamin B12, and folate deficiency; or treatment with erythropoietin) is known to result in underestimation of A1C (8–12,16). Glucocorticoid treatment would be expected to result in acute elevations in glucose. ICU status likely reflects stress hyperglycemia in the setting of critical illness. Other factors less strongly associated with this phenotype were admission to IMC (severity of illness), psychiatry, or other services (potentially related to less aggressive inpatient glucose management on these services or a direct effect of mental illness on glucose control [32]); higher BMI (increased insulin resistance [33]); higher albumin levels (consistent with a previous study [19]); and higher potassium levels (potentially because of underlying hyporeninemic hypoaldosteronism [34]).

We found conflicting patterns of association between certain factors (i.e., age, CKD stage 5, higher sodium levels, and increasing insulin TDD) and the discordant phenotypes. Advancing age is thought to be associated with increasing A1C in patients with normal glucose tolerance, but we found age to be associated with lower odds of both discordant phenotypes in our cohort (35). We also found that CKD stage 5 (eGFR <15 mL/min/1.73 m²) was associated with lower odds of both discordant phenotypes, but other stages of CKD were not. This finding is not unexpected because advanced CKD, dialysis, and erythropoietin treatment have all been associated with spurious A1C values, often in an unpredictable direction (13–18). Finally, we found that higher insulin TDD was associated with higher odds of both discordant phenotypes, which may reflect more advanced diabetes. Because insulin is a relatively narrow therapeutic index medication, it is not surprising that there is greater glycemic variability in patients receiving higher insulin doses (36).

To our knowledge, there have been no previous studies evaluating the relationship between mAG and eAG in both noncritically and critically ill hospitalized patients with diabetes. Several studies have explored this relationship in critically ill patients as a measure of stress hyperglycemia to predict adverse outcomes and mortality (37–45). Overall, the literature suggests that greater discordance between observed and expected glucose levels is associated with worse prognosis and potentially increased mortality in critically ill patients with diabetes. However, these studies used initial glucose values at ICU admission or a mean value for the first week, whereas we calculated a mean value throughout patients’ hospital stays. Thus, our study presents a more complete picture of this relationship and is generalizable to a broader group of patients.

These findings may have implications for the use of A1C as a predictor of inpatient glycemic control in some patients, which could potentially influence selection of an initial antihyperglycemic regimen in ways that could result in ineffective or unsafe outcomes. For example, patients who are expected to have higher glucose values may be started on more aggressive initial insulin doses, which could predispose to hypoglycemia. Indeed, we observed a nearly double mean insulin TDD in the discordant-low compared with discordant-high phenotype (32 vs. 17 units, P <0.001); however, further studies are needed to determine whether discordant mAG and eAG levels lead directly to differences in management that affect glycemic outcomes. Although no single clinical factor identified in this study would be sufficient for clinicians to dismiss patients’ A1C in clinical decision-making, the presence of multiple factors (e.g., glucocorticoid use, anemia, critical illness, known type 1 or type 2 diabetes, race, and CKD) should alert clinicians to the possibility that A1C may not be indicative of the glycemic control that patients’ will experience while hospitalized.

The major strength of this study is its large sample size and inclusion of a broad number of clinical predictor variables. Because we evaluated all hospitalized adult patients with diabetes in a large health system in both critical and noncritical care settings, our findings are widely generalizable. The main limitation of this retrospective study was the influence of sampling frequency on the correlation between blood glucose and A1C. The accuracy of eAG is dependent on the sampling of SMBG, with increased accuracy expected as the number of SMBG measurements increases (46–48). Although eAG based on CGM data has traditionally been thought to have a good correlation with A1C (46,49,50), a recent study reported a high frequency of discordance between A1C and the CGM-derived glucose management indicator—an estimated A1C metric— in 641 patients with diabetes who use CGM (51). Additionally, the equation used to calculate eAG from A1C was derived from both CGM and seven-point SMBG, but blood glucose was measured approximately four times per day for the majority of patients in this study (26). Thus, we expected to have some inherent discordance between mAG and eAG. Regardless, we suspect that most providers are not aware of these issues when using eAG/A1C conversion calculators, which are readily available online and are likely being applied in clinical practice to diverse groups of patients in various settings, including hospitals. Furthermore, we believe that selection of the most extreme discordant phenotypes minimizes the contribution of sampling frequency on the observed discordance.

We identified several factors associated with discordance between mAG and eAG in the hospital, including Black race, strict inpatient antihyperglycemic control through medications and diets, signs of critical illness such as ICU service and abnormal vital signs, anemia, CKD, electrolyte abnormalities, and inpatient glucocorticoid doses. These factors should be considered when appraising A1C to determine an initial antihyperglycemic regimen at the transition from home to hospital.

The work of M.S.A. and N.M. was supported by a K23 grant (K23DK111986) from the National Institute of Diabetes and Digestive and Kidney Diseases.

No potential conflicts of interest relevant to this article were reported.

S.W. contributed to statistical analysis and interpretation of data and drafted the manuscript. M.S.A. contributed to study concept and design, data assessment, data synthesis, and statistical analysis. W.C. contributed to data analysis and reviewed the manuscript. N.M. contributed to study concept and design, data synthesis, statistical analysis, and interpretation of data and drafted the manuscript. All authors reviewed and edited the manuscript. N.M. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.