Accurate population-level denominators are essential for providing reliable estimates of disease burden and evaluating public health interventions [1]. Many governmental and academic entities rely on census-derived population estimates to determine the rate of procedures and prevalence of chronic diseases [2,3]. These estimates of condition prevalence can guide intervention efforts to improve health at the neighborhood, state, and national levels [4-6]. However, inaccuracies in census data and discrepancies between census data and health data measures—particularly among underrepresented populations—have at times led to anomalies in reporting [7]. For example, during the COVID-19 pandemic, vaccination rates exceeding 100% were reported in some groups in both California and Massachusetts [8,9]. In Minnesota, we developed a statewide distributed data network that captures electronic health record (EHR) data for more than 90% of the population, offering an alternative source for deriving population-level denominators for public health data [10-13].

In addition to syndromic surveillance, population estimates play a critical role in allocating public health resources and designing interventions to address inequities [14]. The Centers for Disease Control and Prevention (CDC), through the Behavioral Risk Factor Surveillance System (BRFSS) and its Population Level Analysis and Community Estimates (PLACES) project, reports estimates of chronic disease prevalence for adults aged 18 years and older at the census tract level using small-area modeling that incorporates survey and census data [15]. While these modeled estimates guide federal and state funding to address health disparities, their accuracy may be limited by sparse sampling at the neighborhood level. On average, BRFSS includes only approximately 5 individuals per census tract per year [16]. By contrast, our EHR-based system enables direct observation of diagnosed chronic disease prevalence in nearly every census tract across Minnesota, offering an opportunity to compare EHR-based and model-based small-area estimates [10].

In this study, we compared population denominators derived from census data and Minnesota-based EHRs and examined concordance between CDC and EHR-based estimates of hypertension and diabetes prevalence at the census tract level in the adult population. We assessed concordance between CDC-modeled and EHR-derived prevalence estimates across Minnesota and explored the area-level sociodemographic factors associated with prevalence estimates from EHR data and CDC PLACES. These findings have implications for improving the accuracy of population health surveillance and ensuring equitable distribution of public health resources.

The 2020 census estimate for the population of Minnesota was 5,707,254. A total of 5,271,191 (92.4% of the census estimate) unique individuals visited 1 of the 11 MNEHRC health care systems in 3 years (2018-2020; Table 1). The ratio of MNEHRC patients to the Minnesota statewide census estimate was higher for female individuals (0.97) than for male individuals (0.88) and higher for the older age groups (65 years and older: 1.05) than for the younger age groups (0-17 years: 0.83). The MNEHRC patient-to-census ratio also differed by race—the ratio was highest for Black Minnesotans (1.17) and lowest for American Indian and Alaska Native Minnesotans (0.68). The MNEHRC data also had a greater number of individuals with missing, other, or unknown race (MNEHRC-to-census ratio 1.31). Coverage by county was generally consistent across the state, with a median MNEHRC patient-to-census estimate ratio of 0.89 (IQR 0.73-0.96; Figure 1).

According to the MNEHRC data, the percentage of adults in Minnesota with diabetes in 2022 was 9.5% (415,914/4,376,805), and the percentage with hypertension in 2021 was 32.2% (1,365,413/4,243,000). Crude prevalence estimates from PLACES data for diabetes were 9.9% (435,481/4,389,028) and for hypertension were 29.9% (1,311,459/4,389,028). There was a moderate level of concordance between the prevalence of diabetes and hypertension at the census tract level among adults comparing MNEHRC and PLACES estimates (Figure 2). Relative to the PLACES estimates, MNEHRC estimates tended to be higher in high-prevalence census tracts. This was particularly true for hypertension. The percentage of census tracts where the MNEHRC estimate was within 10% of the PLACES estimate was 40.3% (605/1500) for diabetes and 42.3% (635/1500) for hypertension. The percentage of census tracts where the MNEHRC estimate was within 25% of the PLACES estimate was 77.9% (1168/1500) for diabetes and 79.7% (1195/1500) for hypertension (Figure 3).

Area-level factors associated with diabetes and hypertension prevalence were generally similar in linear regression models for MNEHRC and PLACES (Table 2). Factors associated with higher prevalence in all models included a higher proportion of persons aged 65 years and older, a higher SVI, nonurban census tracts, and a lower proportion of individuals with at least a high school degree or General Educational Development. A higher proportion of Black individuals in a census tract was associated with lower hypertension prevalence in MNEHRC models but higher diabetes and hypertension prevalence in PLACES models. SVI was a stronger predictor of diabetes and hypertension in MNEHRC models compared with PLACES models. Model R² for diabetes was 0.46 in the MNEHRC model and 0.75 in the PLACES model; for hypertension, model R² was 0.52 in the MNEHRC model and 0.81 in the PLACES model.

In this analysis comparing Minnesota statewide EHR data with census and CDC PLACES data, we demonstrated that population estimates from EHR data and the census are generally similar, though there are important differences by age, sex, and race. This finding has important implications for public health data; when possible, it is preferable to obtain numerators and denominators from the same source to avoid overestimating or underestimating prevalence or uptake of services. We also demonstrated moderate concordance between small-area prevalence estimates from EHR data and CDC PLACES. However, we did observe a greater than 25% relative difference in estimates for 22% of census tracts for diabetes and 20% of census tracts for hypertension. These results indicate the need for multiple data sources to accurately estimate small-area prevalence for chronic conditions.

Our results add to the literature evaluating the accuracy of BRFSS-based estimates. In the Boston validation study, BRFSS estimates were compared with estimates from a survey of 7340 individuals in 15 neighborhoods [27]. The BRFSS model estimates of diabetes prevalence were within the 95% CIs of the Boston survey for 80% of the neighborhoods. In the Missouri County-Level Study, 50,690 adults were surveyed across 115 counties [28]. The correlation between the BRFSS estimate for diabetes prevalence at the county level and the Missouri County-Level Study estimate was only 0.51. These prior studies were limited in that they compared estimates between 2 surveys. The results of this study allow a comparison between BRFSS and an estimate from a different data source, namely, EHR data.

Differences in population estimates between MNEHRC and the census may be due to multiple factors. Census undercounting may lead to a lower estimate of the number of Black Minnesotans. American Indian and Alaska Native Minnesotans may receive care at facilities that are not included in the MNEHRC, such as Indian Health Service clinics. The causes of differences in disease prevalence estimates between MNEHRC and PLACES are less apparent. PLACES may be inaccurate in certain census tracts due to their low survey rate, which necessitates modeling-based prevalence estimates. This may have led to some inaccuracies in prevalence estimates, which may be more apparent for the high-prevalence census tracts, as we observed with hypertension. Notably, the relationship between census tract characteristics and disease prevalence differed between MNEHRC and PLACES. A higher proportion of Black and American Indian or Alaska Native residents was associated with lower hypertension prevalence in MNEHRC data but higher hypertension prevalence in PLACES. This could represent differential access to care, lower rates of diagnoses, or PLACES model assumptions—such as age distribution—that are inaccurate in Minnesota. SVI was more strongly associated with hypertension and diabetes prevalence in MNEHRC models than in PLACES models, which may be due to a stronger association between SVI and prevalence in Minnesota compared with national estimates. The higher overall R² in the PLACES model is expected as those estimates are inherently model based using census tract characteristics.

Estimating small-area prevalence using EHR data has several potential advantages. First, EHR data can capture nearly the entire population. In our MNEHRC data, more than 90% of the census-estimated population had an encounter at a participating health care system within a 3-year period. Comparatively, surveys collect data from a small number of individuals within each census tract, and most census tracts are not represented in the National Health and Nutrition Examination Survey at all. This study included the availability of statewide data from a large, multi–health care system distributed EHR network that is scalable in other states. The number of adults in the EHR data for most census tracts is >75% of the census estimate for that population, which makes the EHR-based estimates robust. This allowed for, what we believe is, the first comparison of CDC PLACES– and EHR-based small-area estimates of diabetes and hypertension prevalence. Second, established data collection methods reduce the need for expensive and time-consuming survey collection and field work. Third, EHR data can be produced in near real time. While our analysis cannot comment on which approach (EHR based vs survey) is more accurate, it appears that using EHR data to assess area-level trends would provide similar information to survey-based data and reduce the risk of errors related to discrepancies between numerator and denominator sources. Both EHR data and survey data are prone to some misclassification. While inaccuracies in EHR data may be a particular concern for people who have infrequent health care use, a common concern with survey data is recall bias. Factors that would favor use of EHR data include (1) high percentage of population included in EHR data sources, (2) ability to use objective data to identify conditions (eg, blood pressure for hypertension and hemoglobin A_1c values for diabetes), (3) conditions where EHR data have a high sensitivity, (4) conditions with low prevalence that would require surveying thousands of individuals, (5) conditions that patients may be unlikely to report on surveys either due to stigma or because they are unaware they have the disease, (6) when disease prevalence estimates for key subpopulations would be informative, and (7) when assessing local temporal trends in disease prevalence.

Our study has important limitations. First, the EHR patient count may overestimate the population due to incomplete deduplication, though this is likely to be small. Second, EHR data are unable to account for out-migration, which might represent 1% to 2% of the effect on the population in any given year [29]. By contrast, the EHR patient count may underestimate the population due to individuals being seen by health care providers not included in MNEHRC and individuals who have not been seen by a health care provider in the last 3 years. Third, patients who never seek care may vary in important ways from health care–seeking patients, resulting in EHR-based population and prevalence estimates that do not reflect the entire population. By contrast, the census and PLACES data represent individuals who respond to surveys. Relatedly, both EHR estimates and survey estimates likely underestimate disease prevalence due to undiagnosed disease and lack of disease awareness, but EHR estimates may underestimate disease prevalence less than self-report. Fourth, our study does not allow us to identify a threshold of EHR data coverage above which it becomes more appropriate to use EHR data rather than survey data for public health decision-making. The results of this study are relevant to distributed data networks with robust geographic coverage, particularly those that use OMOP. Finally, our models were estimated via ordinary least squares regression, which assumes independence across census tracts. Because tracts are spatially correlated, uncertainty surrounding coefficient estimates is likely understated, although the estimates themselves are unbiased.

In conclusion, we identified important differences between EHR and census data by age, sex, and race. These differences should be considered in public health reporting when 2 different data sources are used for the numerator and denominator, as was frequently done during the COVID-19 pandemic, where vaccination rates were estimated using census denominators. The prevalence of diabetes and hypertension at the census tract level was generally concordant between MNEHRC and PLACES, but differences suggest that EHR data may help inform targeting of public health interventions to reduce prevalence and adverse outcomes from chronic conditions. Our analysis suggests that there are both similarities and important differences between small-area estimates derived from EHR and survey data. Such differences suggest that further research with validation against an assessment of all individuals using clinical gold standards is needed to determine the optimal collection method for local estimates of health conditions.