Diabetes screening with hemoglobin A1c prior to a change in guideline recommendations: prevalence and patient characteristics

The Canadian Primary Care Sentinel Surveillance Network (CPCSSN) is Canada's first multi-disease EMR-based surveillance system [13]. The North Toronto Research Network (NorTReN) is one of nine networks currently participating in CPCSSN. NorTReN is a practice based primary care research network affiliated with the Department of Family and Community Medicine at the University of Toronto. Family physicians participating in NorTReN contribute their EMR data; an anonymized version of applicable data is sent to CPCSSN's central data repository from each site and is then de-identified and aggregated in a single national database. Posters informing patients about the study are present in the waiting rooms of participating practices, and patients can opt out. One patient has opted out of the NorTReN database, while 111 patients have opted out of the national CPCSSN database.

We obtained data from the Electronic Medical Records (EMRs) of community based family physicians in Toronto, Ontario. Thirteen family physicians participating in NorTReN who had used an EMR for three years or more and who currently contribute data to CPCSSN were included in this study. All physicians were community-based, were members of a Family Health Team [14], and had rostered practices in which patients formally enrolled with their physician, allowing the identification of practice panels [15, 16]. A Family Health Team consists of a group of physicians paid largely by capitation (a set fee based on rostered patients' age and gender); physicians work in multidisciplinary practices, along with nurses, social workers and other allied health professionals. The physicians in this study used a single EMR (Nightingale on Demand^®).

The eligible population included all non-diabetic rostered patients age 19 and over who had seen their family physician at least once between January 1^st 2007 and December 31^st 2009 and were active as of December 31^st 2009.

In order to capture tests that were ordered for screening purposes (and not for disease monitoring), we identified and excluded all known diabetic patients. To define diabetes, we followed a previously validated algorithm for administrative databases [17] and used two billing codes for diabetes in two years. This algorithm only had a sensitivity of 86% [17]. However, we are able to access data unavailable in administrative databases; our definition added the presence of diabetes in the problem list and we used free-text search terms that would likely maximize the number of true positives for the diagnosis of diabetes. If neither billing nor problem list data were present, we then searched for the presence of any current hypoglycemic medication (insulins, sulfonylureas, metformin, thiazolidinediones, incretins, or acarbose), or for an Hgb A1c of 7% or greater, or two fasting blood glucose tests of 7 mmol/l or greater [1]. We then manually excluded patients with polycystic ovarian syndrome (metformin can be used clinically for ovulation induction and management of metabolic syndrome in PCOS) [18‐20], gestational diabetes, secondary diabetes, and hyperglycemia not otherwise specified based on coding for these conditions. The approach is similar to that of a validated algorithm using EMR data [21]. Further development of electronic data queries for case definitions had been undertaken as part of CPCSSN [13].

Laboratory screening increases with age [22]. Screening in younger age ranges is recommended only for patients deemed to be at higher risk of chronic diseases [1, 4, 11, 23]. This may confound our comparisons of the screened and non-screened groups as screening in general may be more limited in younger groups. To limit this source of bias, we present descriptive data restricted to those age 45 and over by December 31^st 2009. By that age, both Canadian and American guidelines recommend diabetic screening for patients at low risk [1, 4]. Patients presenting for a periodic health exam may be more likely to have screening blood work [22, 24, 25], as are those with a chronic health condition. There may be systematic differences between those having screening bloodwork and those not having this done [22]. We used the presence of total cholesterol in the past three years as a marker of a blood test done for screening [22, 26] or chronic cardiovascular disease management. We then also calculated the proportion of screening Hgb A1c tests done for patients age 45 or over who had at least one total cholesterol done, that is those presumed to have presented for screening bloodwork.

We extracted values for Hgb A1c, as well as for data associated with cardiovascular risk factors [26‐29]. The following data elements, if present in the EMR in the past three years (January 1 2007 to December 31 2009), were extracted: age in years, gender, current smoking status (if available in the patient's health profile), Hgb A1c, LDL cholesterol, fasting glucose levels, weight, BMI, waist circumference and blood pressure. Age was calculated as 'test done date' minus July 1, birth-year for all patients with Hgb AIc test done. For those without an Hgb A1C test, age was calculated as 'study end date' of December 31^st 2009 minus July 1, birth-year for all patients. For laboratory measurements, we chose the value occurring immediately prior to the screening Hgb A1c test. If someone was screened on multiple occasions, we considered time of most recent screen and determined values of predictors from this time point. For patients that were never screened, we considered values of the predictors from the most recent time point. The Postal Code Conversion File was used to assign neighbourhood income quintile to the patients' residential postal code.[15]

Data completeness in EMRs used in routine care may be problematic, as clinicians may neglect to enter data or data may be entered in fields that are not captured during searches; some laboratory tests may not be received electronically [30]. To give an indication of data completeness we have included the percentage of missing values by study variable and patient group in the results section.

We first calculated the proportion of non-diabetic patients who had at least one Hbg A1c done in the past three years. We then calculated the proportion of patients with at least one Hgb A1c test in each year of interest. We used logistic generalized estimating equation with year treated as a continuous variable to test for a non-zero slope in Hbg A1c provision.

We compared characteristics of patients age 45 and over who did and did not have at least one Hgb A1c in the three years of interest. We used descriptive statistics such as means, medians, ranges, standard deviations and percentages to summarize baseline data. Logistic regression was used to identify clinically relevant factors associated with Hgb A1c testing [31]. First, we considered odds ratios and p values for each of the univariate predictors of screening. We then modelled screening (multivariable/adjusted model) as a function of various biochemical, demographic and behavioural covariates. All tests were two tailed and p value of < 0.05 was considered statistically significant. We used SAS version 9.2 for the analysis and data manipulation.

This study was approved by the North York General Hospital's Research Ethics Board. All physicians provided signed, informed consent.

This study is limited to community-based family physicians participating in a Practice Based Research Network in Toronto and using a single EMR; results may not be applicable to other settings or other EMRs. Nevertheless, there is no a priori reason to believe that the use of Hgb A1c for screening would be systematically different in this group of physicians; therefore we believe the results from this study provide valuable insight into the gradual increase in Hgb A1c utilization.

Our criteria for determining the presence of diabetes are similar to those others have employed, but have not been formally validated. Validation of chronic disease criteria in the CPCSSN database is ongoing. Furthermore, we could not reliably include patients with an Hgb A1c of 7% or more into the screened cohort because data in the EMR were not complete enough to allow us to determine whether the diagnosis of diabetes existed prior to the test. Thus, there was uncertainty as to whether they were screened or monitored. Including patients with a Hgb A1c of 7% or more in the screening group would have inflated our results.

EMR data abstracted for this study were fairly complete, but there were differences in missing data between the populations screened and not screened with Hgb A1c. Nearly all (95%) of patients age 45 or over with a Hgb A1c screen had a LDL cholesterol present in the EMR. However, only 72% of those without Hgb A1c had at least one LDL cholesterol level present in the three years studied. We have no reasons to assume that rates of missing EMR data (test done, but data not present in an electronically auditable form) are different between the two groups, so it is possible that some patients did not have a Hgb A1c done simply because they did not present for any screening blood work.

This study examined factors extensively documented in the literature as being associated with cardiovascular risk factors [26‐29] and clinically likely to influence a family physician's risk estimation. There are a large number of data elements in the EMR, and we wished to be conservative in order to exclude spurious associations that could be significant due to the large sample size. Therefore, we did not include LDL cholesterol in the model because it had only marginal statistical association and limited clinical plausibility. It is possible that consideration of other variables for the model may have altered the results, but in a sensitivity analysis, we included triglycerides and this did not change our findings.