HbA1c Evidence for a Prediabetes Diagnosis Delays Onset of Type 2 Diabetes

Page 1 of 15 HbA1c Evidence for a Prediabetes Diagnosis Delays Onset of Type 2 Diabetes Maurice C. Johnson, Jr.1, Howard A. Fishbein1*, Rebecca Jeffries Birch1, Qilu Yu2, Russ Mardon1, Wilson Pace3, Natalie Ritchie4, Jennifer K. Carroll5, Daniella Meeker6 1Westat, Rockville, MD, USA 2National Center for Complementary and Integrative Health (NCCIH), National Institutes of Health (NIH), Bethesda, MD, USA 3DARTNet, Aurora, CO, USA 4University of Colorado/Denver Health and Hospitals Authority, USA 5University of Colorado, USA 6University of Southern California, USA


Introduction
Diabetes mellitus affects 30.3 million individuals in the US 1 and is the seventh leading cause of death 2 , accounting for an estimated economic cost of $327 billion in 2017 3 . In adults, the majority of diabetes cases are type 2 diabetes. Prediabetes, also referred to as intermediate hyperglycemia, signifies an increased risk of developing type 2 diabetes 4,5,6 . Nearly 84.1 million US adults, or roughly one third of US adults, are estimated to have prediabetes 1 .
Diagnosing prediabetes may be an important initial step toward actions to slow or prevent diabetes onset. Research shows that lifestyle interventions, including physical activity, healthy diet, and treatment with drugs (e.g., metformin), can delay or prevent progression from prediabetes to diabetes 7,8 . The American Diabetes Association encourages health care providers to screen for prediabetes, followed by referring to diabetes prevention programs and prescribing metformin therapy 9 . Nonetheless, adoption of prediabetes screening tools and referrals to lifestyle interventions are relatively low 11 , as is use of metformin therapy 10 . Studies of electronic health records (EHRs) further show relatively few recorded prediabetes diagnoses in patients with glycemic evidence of the condition, and highlight knowledge gaps regarding the impact of diagnosis on progression to diabetes 11,12,13 . The debated utility of diagnostic protocols for prediabetes [14][15][16][17][18][19][20] may underlie low rates of screening, diagnosis, and treatment, and therefore further study is needed to understand patterns in prediabetes diagnosis and what association exists between receiving a diagnosis and progression to diabetes.
To understand factors that influence diagnosis of prediabetes, this study uses the published large-scale database of EHR records 21 to assess the demographic and clinical differences between diagnosed and undiagnosed patients with evidence of prediabetes based on glycosylated hemoglobin (HbA1c) results. The study further investigates differences in risk of progression to diabetes between diagnosed and undiagnosed prediabetes patients.

Study Design and Population
This retrospective observational study uses the Longitudinal Epidemiologic Assessment of Diabetes Risk (LEADR) database 21 , which is an aggregated EHRderived database for examining patient-level risk factors for diabetes among adults. The EHR data comes from participating healthcare organizations including primary care and multi-specialty integrated delivery systems with ambulatory and inpatient components located across the U.S and aggregated into four regions: Northeast (Connecticut, Vermont, and Ohio), South (North Carolina and Tennessee), Rocky (Wyoming, Nebraska, and Colorado), and West (California, Idaho, and Washington).
Patients in the LEADR database had at least four encounters within a participating healthcare network between January 1, 2010, and December 31, 2018, each at least 14 days apart and with at least 24 months between the first and last encounters. An encounter is defined as the unique date on which one or more conditions, observations, procedures, drugs, or visits are recorded within a healthcare network's EHR system. Patients with type 1 diabetes, type 2 diabetes, gestational diabetes, or unspecified diabetes prior to or during their first 12 months after initial identification were excluded from LEADR such that those remaining with subsequent evidence of diabetes were considered incident diabetes cases. The resulting cohort had over 2 million patients with geographic region serving as the only attributable provider characteristic. Standardization of EHR records across the participating LEADR healthcare networks was accomplished with the Observational Medical Outcomes Partnership Common Data Model (OMOP) 22 .
A series of steps were conducted to derive an analytic population that 1) limited analysis to incident cases of prediabetes, 2) ensured all patients had glycemic evidence of prediabetes, 3) ensured all patients had shared clinical and demographic data available (complete case analysis), and 4) contained the same population for both the logistic and cox regression analyses. To accomplish these goals, the study focused on patients newly found to have prediabetes, defined as a HbA1c value in the prediabetes range (5.7-6.4%) following a 12 month washout period with no evidence of the condition. We selected HbA1c as the only glucose-related value to identify patients with prediabetes as only a small percentage (< 5%) of the glucose measures in LEADR had a corresponding fasting indicator. Hence, most blood sugar values were considered random glucoses, which are not standard observations used for diagnosing diabetes 9 . Each patient's first HbA1c record in prediabetes range served as the baseline date.
Patients with at least one prediabetes diagnosis code were considered cases of diagnosed prediabetes (DxP), and those without a diagnosis code were defined as cases of undiagnosed prediabetes (UDxP). Determination of a prediabetes diagnosis used OMOP codes related to the International Classification of Diseases, 9 th (ICD-9) and 10 th (ICD-10) edition codes 790.2X and R73.03.
The study population was further limited to patients with at least one of each of the following: 1) HbA1c result, 2) BMI, 3) blood pressure results (i.e., systolic and diastolic values, or diagnosis or medication related to treatment of hypertension), and 4) lipid results (i.e., triglyceride, HDL, or non-HDL values, or diagnosis or medication related to treatment of elevated lipids). This restriction helped ensure patient-level information was similar across health care systems and facilitated the assumption that clinicians had comparable information when assessing patients' risk of prediabetes and diabetes.
Additional patients were excluded if missing demographic information (i.e., age, sex, or race/ethnicity), or they were pregnant or had gestational diabetes during the study period. Figure 1 outlines how the analytic population was derived from the LEADR cohort. Supplemental Table 1 provides a summary of the reasoning for each of the steps in developing the analytic population for this study.
New cases of type 2 diabetes were identified from diagnosis records, anti-diabetic drug prescriptions, HbA1c, fasting plasma glucose, and random glucose values. diagnosis as it was the most significant interaction identified. An interaction term between prediabetes diagnosis and follow-up time was further added to the final model after examination of Kaplan-Meier curves and Schoenfeld plots indicated non-proportionality; as a result, hazard ratios (HR) and 99% confidence interval for prediabetes diagnosis are estimated assuming the median follow-up time of 634 days (1.7 years).
Analyses were performed using SAS 9.4, with a p-value of < 0.01 to determine statistical significance.

Descriptive statistics of study population
The study population included 40,970 patients with a new case of prediabetes. Table 1 summarizes the demographic and clinical characteristics of the study population as well as the differences between the Dxp and UDxP groups. The UDxP group represented 76.8% of the study population. While 79.8% of patients with a low HbA1c level were undiagnosed, slightly over 70% of patients with medium and high HbA1c levels remained undiagnosed as well. With the exception of metformin prescriptions, the demographic and clinical characteristics between the DxP and UDxP groups differed significantly.
Average HbA1c was 0.05% higher in the DxP versus the UDxP populations (5.93% versus 5.88%). There was a higher proportion of medium and high HbA1c levels in DxP versus UDxP populations (40.4% versus 29.0%). The average follow-up time was nearly one year longer in the DxP versus UDxP groups (1003 Supplemental Table 2 provides detailed definitions of the study's demographic and clinical variables of interest.

Statistical Methods
Summary statistics among patient-level characteristics were examined by DxP and UDxP status. Differences between these groups were tested using a t-test for continuous variables and x 2 -test for categorical variables.
The association between a prediabetes diagnosis and progression to type 2 diabetes was explored using a rightcensored Cox regression model adjusting for age, sex, race/ethnicity, region, weight classification, HbA1c level, elevated lipid status, hypertension status, history of tobacco use, record of a metformin prescription, and number of encounters after baseline. Follow-up time for each patient was calculated from baseline to last recorded encounter date or earliest recorded date of diabetes, whichever came first. All possible interactions with prediabetes diagnosis status were investigated. The final adjusted model included an interaction term between HbA1c level and prediabetes The DxP population had a lower proportion of non-Hispanic white and non-Hispanic black patients than the UDxP population (50.9% versus 60.8%, and 7.8% versus 12.1%, respectively); however, the DxP population had a larger Hispanic population (35.6% versus 22.3%). Patients in the 40-64 age group made up a majority of the study population, with that age group having slightly more representation in the DxP population (64.0% and 60.8%). The DxP population also had slightly more females compared to the UDxP group (61.6% versus 58.1%).   Figure 2 presents the odds ratios (ORs) from the adjusted logistic regression. Supplemental Table 3 presents the adjusted parameter estimates, standard errors, and p-values from the adjusted logistic regression.

Multivariate analyses comparing diagnosed and undiagnosed prediabetes
HbA1c level was the most significant clinical factor that increased the likelihood of a prediabetes diagnosis. Compared to patients with a low HbA1c level, those with medium and high levels had ORs of 1. Compared to non-Hispanic white patients, Hispanic and Asian/American Indian/Alaskan Native/Other/ Multiracial patients were more likely to be diagnosed with prediabetes (ORs of 1.29, CI 1.18-1.42 and 1.27 CI 1.10-1.48, respectively). Patients in the 40-64 age group were more likely to be diagnosed than the 18-39 age group (OR 1.14, CI 1.04-1.27). Males were less likely than females to have a prediabetes diagnosis (OR 0.85, CI 0.80-0.91). Region and number of encounters were also significant factors in estimating odds of being diagnosed.  Figure 3 presents the hazard ratios from the adjusted Cox regression model at median time for a prediabetes diagnosis. Supplemental Table 4 presents the adjusted parameter estimates, standard errors, and p-values from the adjusted Cox regression model. The patient's HbA1c level was the most significant predictor of type 2 diabetes with chi-square values 5 to 10 times higher than for any other predictor (Supplemental Table 4). The effect of receiving a prediabetes diagnosis varied by HbA1c level (p<.0001 for interaction term). For each HbA1c level, UDxP patients progress to type 2 diabetes faster than DxP patients, and the difference in progression increases with higher baseline HbA1c levels. Specifically, as  In all three classifications of HbA1c level, patients in the UDxP population progress to type 2 diabetes at a faster rate. The survival rate at the end of follow-up for those at the low HbA1c level were ~95% for DxP and ~94% for UDxP (~1% difference). Among patients with a medium HbA1c level, the survival rate was approximately 87% for those diagnosed and 82% for those undiagnosed (~5% difference). Likewise, the survival rates for the high HbA1c level were 75% for those diagnosed and 55% for those undiagnosed (~20% difference). This pattern was   Furthermore, the higher the HbA1c level, the flatter the HR slope: the highest HR reached among the high HbA1c level group was nearly 1.5 by the end of the study period that was plotted (~7 years), while the low HbA1c level group reached an HR above 2.5 over the same time interval.

Longitudinal analyses for risk of type 2 diabetes
Other characteristics had significant relationships with development of type 2 diabetes. Among clinical factors, patients in obesity I, II, and III categories were more likely to develop type 2 diabetes than those in the normal/underweight classification (HRs ranged from 1.3-1.7). Other positive associations with development of type 2 diabetes were found for those with elevated lipids (HR 1.89

Conclusions
This study aimed to determine demographic and clinical differences between patients with and without a prediabetes diagnosis among incident prediabetes cases detected by HbA1c, as well as differences in rates of development of type 2 diabetes. Results show that greater BMI, higher HbA1c, elevated lipids, hypertension, and frequency of encounters are clinical factors associated with assigning a prediabetes diagnosis. Demographic factors associated with a diagnosis included being female, Hispanic or Asian/American Indian/Alaska Native/Other/ Multiracial, and being in the 40-64 age group. Results further indicate that assigning a prediabetes diagnosis is associated with slower development of diabetes, and the protective benefit of a diagnosis increases the higher a patient's baseline HbA1c level. Overall, higher HbA1c levels were associated with greater rates of diabetes onset, as was being younger, male, obese, and Asian/ American Indian/Alaska Native/Other/Multiracial, and having elevated lipids, hypertension, and more encounters within the health care system. This study leverages EHR data to investigate the demographic and clinical differences between undiagnosed and diagnosed prediabetes patients. Most prediabetes patients in this analysis were classified as undiagnosed, which is consistent with previous research 23,24,25 . Previous studies have found that EHRs are useful tools for identifying at-risk patients for prediabetes screening 14,15,26,27 , and this study adds insight into patient-level characteristics that may influence diagnosis. Findings from this study also highlight the potential value of assigning a prediabetes diagnosis to patients with evidence of the condition through an HbA1c test, and adds to the research for using HbA1c as a metric for assessing and treating prediabetes 19,28,29,30 . Having a prediabetes diagnosis may lead to patient-provider conversations about diabetes risks, resulting in patient changes in risk reduction behaviors 31,32,33 . The finding that the diagnosed group had nearly 8 more encounters than the undiagnosed group following the baseline prediabetes date may be an indication of patient activation following the prediabetes diagnosis.
Demographic differences in the assignment of a prediabetes diagnosis may reflect broad health equity and contextual factors 34,35,36 . There are many provider specific and system-level contextual factors that may influence clinical practice, including support and uptake of the recommended practice guidelines within each health system 37 . Decisions to assign a prediabetes diagnosis can also be affected by varying organizational protocols, availability of resources to address prediabetes, programming targeted to at-risk populations, and differing levels of provider awareness of diabetes prevention programs, resulting in differences in how clinicians assign diagnosis codes [38][39][40][41][42] . Variability in the strength of provider beliefs in prediabetes as a disease entity may also influence a decision to assign or not assign this diagnosis 19,28,43,44 . The high number of undiagnosed patients in our study could also be reflective of a lack of action to assign a medical condition based solely on a lab result in the prediabetes range. This could particularly be true among patients with a low HbA1c level as they made up 71% of the undiagnosed prediabetes group compared to 59.6% of the diagnosed group.
There are inherent limitations in observational studies and those utilizing EHR data, including that EHR data provide more information on patients engaged more frequently with health systems. Limitations in our analysis were minimized to the best of our ability. For example, large administrative datasets may lack precision and detail, and require a validated methodology for identifying patients with chronic diseases like diabetes 45 . We addressed this concern by applying rigorous standards to separate those with and without a prediabetes diagnosis code. While there is currently widespread interest in promoting and adopting EHRs as a viable research tool for prevention, low data density for various observations and discrepancies in regional data availability is challenging. We conducted a complete-case analysis under the assumption that each of the covariates in our analysis were observations available to providers when assessing patients' risk for prediabetes. However, we recognize that our analysis was biased towards patients with all these variables available in the EHR and acknowledge this is not fully representative of real-world scenarios.
Solely defining patients based on HbA1c may have overlooked a number of patients who had normal HbA1c observations with corresponding blood glucoses in the prediabetes range. However, given that most of the patients within LEADR did not have a fasting indicator (over 95%), we focused on HbA1c to serve as a reliable criterion for assessing prediabetes. Furthermore, while applying a 365day "washout" period has been used in previous studies to account for the prevalence of a condition, we recognize that this process does not fully guarantee all patients in our analytic population had incident prediabetes 46,47,48 .
Other limitations include lack of information on diet, physical activity, or socioeconomic status. To preserve confidentiality of participating healthcare organizations, provider-and clinic-level characteristics were not available to the LEADR research team, including information on clinic type or physician specialty -all potentially important factors in screening and treating prediabetes and diabetes.
Future research should continue examining prediabetes diagnosis patterns among providers and the potential shortand long-term health impacts of a prediabetes diagnosis on patients. Based on the demographic differences in diagnosis rates found in this study, we recommend further exploration of health equity factors of a patient, including social determinants, which may influence assignment of a prediabetes diagnosis. A focus on provider characteristics would also further help researchers gain a more robust understanding of the nonclinical characteristics influencing the assigning of a prediabetes diagnosis. This study also recommends further research on the mitigating factors on the progression to diabetes among patients with prediabetes, including referrals to and uptake of lifestyle intervention counseling.

4) A prescription for nicotine or varenicline
Patients that did not meet any of these criteria noted above were classified as no record of tobacco use

Metformin prescription
A prescription for metformin on the baseline date or following date or following AND prior to the first indication of diabetes, if any. Patients that did not meet any of these criteria noted above were classified as no record of metformin.
Case definitions for dependent variable (prediabetes diagnosis), outcome variable (onset type 2 diabetes), and covariates. Definitions are applied the LEADR dataset using the Observational Medical Outcomes Partnership Common Data Model. For both the logistic regression and Cox non-proportional hazards, the baseline date is the first encounter when HbA1c level is in prediabetes range (≥ 5.7% and <6.5%)  Right-censored Cox non-proportional hazards model with onset of diabetes as outcome. Estimates, standard errors (SE), chi-square and p-values presented. + every 10 encounters, centered around mean.