Original Article
POPULATION HEALTH MANAGEMENT Volume 0, Number 0, 2014 ª Mary Ann Liebert, Inc. DOI: 10.1089/pop.2014.0058
Exploring Electronic Health Records as a Population Health Surveillance Tool of Cardiovascular Disease Risk Factors Abbey C. Sidebottom, MPH,1 Pamela Jo Johnson, MPH, PhD,2,3 Jeffrey J. VanWormer, PhD,4 Arthur Sillah, MPH,1 Tamara J. Winden, MBA,1 and Jackie L. Boucher, MS, RD 5
Abstract
The objective of this study was to examine the utility of using electronic health record (EHR) data for periodic community health surveillance of cardiovascular disease (CVD) risk factors through 2 research questions. First, how many years of EHR data are needed to produce reliable estimates of key population-level CVD health indicators for a community? Second, how comparable are the EHR estimates relative to those from community screenings? The study takes place in the context of the Heart of New Ulm Project, a 10-year population health initiative designed to reduce myocardial infarctions and CVD risk factor burden in a rural community. The community is served by 1 medical center that includes a clinic and hospital. The project screened adult residents of New Ulm for CVD risk factors in 2009. EHR data for 3 years prior to the heart health screenings were extracted for patients from the community. Single- and multiple-year EHR prevalence estimates were compared for individuals ages 40–79 years (N = 5918). EHR estimates also were compared to screening estimates (N = 3123). Single-year compared with multiyear EHR data prevalence estimates were sufficiently precise for this rural community. EHR and screening prevalence estimates differed significantly— systolic blood pressure (BP) (124.0 vs. 128.9), diastolic BP (73.3 vs. 79.2), total cholesterol (186.0 vs. 201.0), body mass index (30.2 vs. 29.5), and smoking (16.6% vs. 8.2%)—suggesting some selection bias depending on the method used. Despite differences between data sources, EHR data may be a useful source of population health surveillance to inform and evaluate local population health initiatives. (Population Health Management 2014;xx:xxx–xxx)
would address the challenges of having to extrapolate results from national surveys such as the National Health and Nutrition Examination Survey or the Behavioral Risk Factor Surveillance System to inform local initiatives.6,7 Given the public health impact of cardiovascular disease (CVD),8,9 surveillance for CVD risk factors using EHR data may have a high degree of application for population health improvement initiatives. Research exploring the use of EHR data for population health surveillance of CVD risk factors is currently limited. A conceptual model for the use of EHR data to conduct population health surveillance of CVD risk factors was previously presented as part of the Heart of New Ulm (HONU) Project.6 The objective of the present study was to examine the utility of using EHR data for periodic community health surveillance. To achieve this objective, the
Introduction
A
s the adoption of electronic health record (EHR) systems continues to increase in the United States,1 so have calls for secondary use of EHR data for outcomes, quality of care, and epidemiological research, as well as for public health functions such as identifying epidemic outbreaks,2 and disease or health behavior surveillance.3,4 The potential public health and population-based uses of EHR data are being recognized and fostered by federal priorities such as the Health Information Technology for Economic and Clinical Health Act, which includes Meaningful Use standards for EHRs.5 EHR data have the potential to address situations in which community health surveillance data are not available and it is cost prohibitive to conduct primary data collection. The availability of EHR data at local levels also 1
Division of Applied Research, Allina Health, Minneapolis, Minnesota. Center for Spirituality & Healing, University of Minnesota, Minneapolis, Minnesota. 3 Division of Epidemiology & Community Health, University of Minnesota, Minneapolis, Minnesota. 4 Epidemiology Research Center, Marshfield Clinic Research Foundation, Marshfield, Wisconsin. 5 Department of Education, Minneapolis Heart Institute Foundation, Minneapolis, Minnesota. 2
1
2
research team sought to answer 2 research questions. First, how many years of EHR data are needed to produce reliable estimates of key population-level CVD health indicators? Second, how comparable are the EHR estimates relative to those from community heart health screenings? Methods Setting
The HONU Project is a 10-year community health initiative designed to reduce the incidence of myocardial infarctions and CVD risk factors in New Ulm, Minnesota, with the specific target population of approximately 7855 residents of the 56073 zip code, ages 40–79 years (2010 Census).10,11 New Ulm is a rural community 100 miles southwest of the Minneapolis– St. Paul, Minnesota, metro area. It has 1 medical center that includes a clinic and hospital, providing medical care to nearly all community residents. At the time HONU began, it was estimated that 90% of the town’s adult residents were served by the medical center, making it possible to use the EHR to assess changes in the population. HONU is a collaborative partnership of Allina Health, the Minneapolis Heart Institute Foundation, the New Ulm Medical Center, and the community. HONU interventions are aligned with a social-ecological model of health determinants and health promotion addressing CVD risk factors at individual, social, institutional, community, and policy levels.12 Serial heart health screenings are a key feature of the HONU program and were first implemented in 2009.11,13 Screenings of employees were aggregated for worksites to identify opportunities for employee wellness programming, and aggregated data from screenings were used to create a community CVD risk profile to prioritize aspects of intervention development. Data sources
The data sources for this study include the 2009 community heart health screenings and the Allina EHR. This study was approved by the Institutional Review Board of Allina Health. Screening data. Heart health screenings were held from mid-April to mid-December 2009, with 109 screening events at locations throughout the community, including worksites, community centers, churches, and the clinic. Screenings were free and open to any adults in the general area. Attendance at the screening was entered as a visit into the EHR. Screenings included a health history and CVD risk factor survey, anthropometric measures (height, weight, blood pressure [BP], waist circumference), and a venipuncture (lipids and glucose). Three BP measures were taken with an automatic sphygmomanometer after the participant had been sitting for 3 minutes. The mean of the last 2 measures was used for analytical purposes. Participants were instructed to fast for 12 hours before screening. Further details about the 2009 screening have been published previously.11,13 A total of 4166 adults from the 56073 zip code participated in the 2009 screenings. The study sample is limited to adults ages 40–79 years (N = 3123). EHR data. Allina uses a single EHR system, developed by Epic Systems Corporation (Verona, WI), across the
SIDEBOTTOM ET AL.
entire health system, which includes 11 hospitals and more than 90 clinics. So, although data extracts selected only residents of the HONU target area, data may be included from visits to any Allina facility during the time period. The EHR extract selected patients ages 40–79 years residing in the 56073 zip code who had at least 1 ambulatory encounter (ie, office, nurse only, preoperative, or ob-gyn visits) between April 1, 2006, and March 31, 2009 (ie, within 3 years prior to the start of the heart health screenings). Selecting EHR data from before the screenings was important because of the enhancement of EHR content that occurred during the screening period, including registration of new patients and entry of screening results into the EHR. Patients were excluded if their only visit during the time period was any of the following visit types: emergency department, hospital inpatient, hospital observation, or hospital outpatient. Patients also were excluded if they did not have electronic documentation of consent to allow the use of their medical records for research purposes. The most recent values within the 3-year time period were extracted for fasting cholesterol, BP, and body weight. Height and smoking values were selected as the most recent available before the end of the extract period without a limit on how far back because of workflow related to these 2 measures. Height was combined with weight to calculate body mass index (BMI). Both fasting cholesterol and BP values were taken only from ambulatory visits while measures of height, weight, and smoking status were allowed from any type of visit. The full 3 years of EHR data included 6659 patients, but 741 (11%) were excluded because of no consent to use their records for research. The final sample of 5918 patients ages 40–79 years represents three fourths of all residents of this target area and age group per US Census data. The EHR data were further disaggregated into subsamples representing 3 time frames: 1 year of data (patients whose most recent values occurred within the 12 months prior to March 31, 2009), 2 years of data (patients whose most recent values occurred within 2 years prior to March 31, 2009), and 3 years of data (April 1, 2006, to March 31, 2009). Because of the way smoking status is updated in the EHR, the most recent smoking status was used from any date prior to March 31, 2009, to create a prevalence measure of smoking; 1-, 2-, and 3-year estimates could not be calculated. Measures
Four common modifiable CVD risk factors were examined: BMI, measured as a continuous measure and a dichotomous measure of obese (BMI ‡ 30 kg/m2) or not obese; BP as continuous measures of systolic blood pressure (SBP) and diastolic blood pressure (DBP) and as a dichotomous measure of uncontrolled BP (SBP mmHg ‡ 140 or DBP ‡ 90 mmHg); total cholesterol (fasting) as a continuous measure and a dichotomous measure of high cholesterol (‡ 200mg/dL); and smoking status as a dichotomous measure (current smoker or not). Measures of high cholesterol and uncontrolled BP are based only on test result values and do not incorporate use of medication. Analysis
Summary statistics of demographic characteristics were produced to describe the population in each data source
EHR AS A POPULATION HEALTH SURVEILLANCE TOOL FOR CVD
compared to Census information. Two-sample t tests were used to assess significant differences in mean values from the 1, 2, and 3 years of EHR data. Comparisons were done between the 1- and 2-year estimates and between the 2- and 3-year estimates. Two-sample t tests also were used to test for differences between the screening data and the EHR 1-year estimates. Z tests were used for categorical measures. All analyses were conducted using Stata version 12 statistical software (StataCorp LP, College Station, TX). Results
Not all CVD health indicators were documented at the same frequency in the EHR, with cholesterol having the lowest data availability within a 3-year time period. Specifically, 98% had data on BP, 81% had data on BMI, and 67% had data on total cholesterol available within those 3 years. Of this same set, 67% had data on all 3 biometric measures of interest within a 3-year time period. Smoking status was available for 88% of patients who had a visit within the 3-year time period. Availability of data was lower when looking only within the most recent 1-year period of EHR data. Out of the 5918 patients who had a visit in the past 3 years, 81% had data on BP, 69% had BMI data, and 43% had total cholesterol data within the previous 1 year. Only 41% (2447; approximately 31% of the population) had data available on all 3 biometric measures of interest within the previous year. Among subjects in the HONU screening data, 98% had complete data for all 4 health indicators. Table 1 shows a demographic comparison of age and sex between the EHR data set, the screening data set, and the year 2010 US Census for the target population. Relative to Census data, the EHR data set seemed to provide a closer reflection of sex, whereas the screening data set provided a slightly closer match to age. Racial and education level measures were not compared because of the different methods used to measure them across systems. To address the first research question regarding how many years of EHR data are needed to produce reliable estimates of key population health indicators for a community, CVD risk factor values were compared using 1, 2, and 3 years of EHR data. Table 2 shows estimates of continuous measures of each indicator by the number of EHR data years used to produce the estimate. The 2-year esti-
3
mates are not significantly different from the 1-year estimates, and the 3-year estimates are not significantly different from the 2-year estimates. Community assessments often focus on the percent of the population exceeding a threshold that indicates increased risk (as opposed to mean values). In alignment with that style of measurement, Figure 1 depicts the estimates of the percent of the population meeting or exceeding each of the threshold values (ie, obese, uncontrolled BP, high cholesterol) by years of EHR data aggregated. Significance testing found no differences when comparing the percent meeting or exceeding threshold levels for the 1-year compared to the 2-year estimates and the 2-year compared to the 3-year estimates. The overall finding from comparisons in Table 2 and Figure 1 indicate that the number of years of data aggregated, between 1 and 3, do not produce statistically nor clinically different point estimates or levels of precision. To examine the question of how EHR estimates of risk factor levels compare to those from community screening data, the 1-year EHR estimates were compared to the HONU screening estimates for 4 key CVD risk factors. All measures were significantly different in these comparisons (Table 3), but the direction of the difference varied by individual risk factor. Specifically, EHR estimates were higher than screening estimates for BMI and smoking, while they were lower than screening estimates for BP and cholesterol. Mean differences between the 2 data sources indicate BMI is 0.7 kg/m2 higher in the EHR data, while SBP is 4.9 mmHg lower, DBP is 5.9 mmHg lower, and total cholesterol is 15.0 mg/dL lower in the EHR data compared to the screening data. When examined using proportions meeting thresholds of risk, the estimate of smoking prevalence was 8.4 percentage points higher and the estimate of obesity was 4.1 percentage points higher in the EHR data while the estimate of uncontrolled BP was 10.7 percentage points lower and the estimate of high cholesterol was 15.0 percentage points lower in the EHR. Figure 2 graphically displays estimates of the percent of the population meeting or exceeding each of the risk threshold values by data source. Discussion
This study examined the use of the EHR as a population health surveillance tool for CVD risk factors in a rural
Table 1. Age and Sex of Adults, Ages 40–79 Years, by Data Source Census Dataa
Sex Female Male Age group 40–49 years 50–59 years 60–69 years 70–79 years a
EHR Datab
HONU 2009c
No.
%
No.
%
No.
%
4009 3846
51.0% 49.0%
2709 2145
55.8% 44.2%
1,809 1,314
57.9% 42.1%
2190 2695 1799 1171
27.9% 34.3% 22.9% 14.9%
1282 1620 1128 824
26.4% 33.4% 23.3% 17.0%
870 1,110 722 421
27.9% 35.5% 23.1% 13.5%
Census 2010 for zip code 56073, using SF1 data. EHR data include anyone who had at least 1 measure for cholesterol, blood pressure, or BMI between April 1, 2008, and March 31, 2009. HONU uses no weights or population control adjustments. BMI, body mass index; EHR, electronic health record; HONU, Heart of New Ulm b c
4
SIDEBOTTOM ET AL.
Table 2. Community-Level Continuous Measures for Adults Ages 40–79 Years by Number of EHR Data Years Selected.
Body Mass Index Systolic BP Diastolic BP Total Cholesterol
1-Year EHR Data
2-Year EHR Data
3-Year EHR Data
No.
Mean
SD
No.
Mean
SD
T
P Value
No.
Mean
SD
T
P Value
4093 4787 4783 2515
30.2 124.0 73.3 186.0
6.5 15.6 9.9 38.5
4646 5554 5549 3554
30.1 123.8 73.4 186.9
6.4 15.6 9.9 38.0
0.87 0.52 - 0.72 - 0.92
0.383 0.602 0.475 0.358
4780 5805 5800 3972
30.1 123.8 73.5 187.5
6.4 15.6 9.9 37.5
0.15 0.07 - 0.38 - 0.67
0.879 0.945 0.707 0.502
BP, blood pressure; EHR, electronic health record; SD, standard deviation
community. A comparison of estimates using 1, 2, or 3 years of data found no meaningful differences despite the fact that there was much less data available in the 1-year sample compared to the 3-year sample. This suggests that relatively brief, 1-year time periods of EHR data were sufficiently precise to produce reasonable risk factor prevalence estimates for this rural community, and that a 1-year sample of EHR data may be a relatively stable representation of other proximal time points in the EHR. How well a single or multiyear estimate measures prevalence is influenced by how many people visit a clinic on an annual basis and what measures are collected during those visits. In this rural community, approximately 75% of residents ages 40–79 years had an office visit within 3 years, and about 50% had data on 3 biometric measures within 3 years. Approximately 31% of the target population had data on all 3 biometric measures within the past year. National data indicate that approximately 80% of adults have an office visit with a medical provider annually, and that this is higher in older age groups (84% among those ages 45–64 years, 93%
among those ages 65–74 years).14 Although the research team is unable to calculate an annual attendance rate from the study data, the 3-year visit rate appears to be a slightly lower than the national level for having at least 1 visit in a year. In contrast, there were clear differences observed between EHR-derived versus community screening-derived prevalence estimates of CVD risk factors. This was somewhat expected, given the different mechanisms by which individuals present under these 2 mechanisms. The EHR data in this study were collected as part of routine care delivered during ambulatory medical visits, which is assumed to occur based on need and typically for reasons other than CVD prevention. The screening data in this study were collected as part of a specific community-wide CVD prevention initiative that individuals volunteered to participate in. Assuming these findings can be replicated in other care systems, there are numerous implications for population health planners. With a focus on the opportunity EHR data create for identifying those at high risk for CVD who could
FIG. 1 Comparison of risk factor estimates by number of years of EHR data. BP, blood pressure; CVD, cardiovascular disease.
EHR AS A POPULATION HEALTH SURVEILLANCE TOOL FOR CVD
5
Table 3. Community-Level Estimates for Adults Ages 40–79 years, HONU Community Screening Data Versus EHR Data. HONU Screening
1-Year EHR Data
Continuous Measures
No.
Mean
SD
No.
Mean
SD
T
P Value
Body Mass Index Systolic BP Diastolic BP Total Cholesterol
3096 3114 3114 3123
29.5 128.9 79.2 201.0
6.0 16.5 10.1 37.8
4093 4787 4783 2515
30.2 124.0 73.3 186.0
6.5 15.6 9.9 38.5
- 4.9 13.4 26.0 14.6
< 0.001 < 0.001 < 0.001 < 0.001
Proportion meeting threshold Current smoker Obese Uncontrolled BP High Cholesterol
No.
%
SE
No.
%
SE
Z
P Value
3108 3096 3114 3123
8.2% 40.2% 27.9% 49.6%
0.5% 0.9% 0.8% 0.9%
5203 4093 4783 2515
16.6% 44.3% 17.2% 34.6%
0.5% 0.8% 0.6% 1.0%
- 10.7 - 3.5 11.3 11.3
< 0.001 < 0.001 < 0.001 < 0.001
BP, blood pressure; EHR, electronic health record; HONU, Heart of New Ulm; SD, standard deviation; SE, standard error
benefit from preventive treatment, a recent study from 1 health system in Washington State used EHR data from the previous 5 years to calculate composite CVD risk scores. Findings indicated sufficient data for up to 84% of the population to calculate this metric, with cholesterol being the most common missing variable, as was observed in the present study.15 Another study examined the use of an algorithm applied to EHR data to accurately identify patients with type 2 diabetes using different time periods of data ranging from 1 to 11 years. That study found lower accuracy with only the most recent 1 or few years of data compared to longitudinal data for multiple years.16 Findings from these 2 studies15,16 and the present study indicate the number of years of data to accurately identify risk levels may vary by the specific risk factor, and methodological considerations
regarding number of years of data necessary may vary depending on whether the goal is to identify individuals with a diagnosis or estimate population levels of risk. No other studies were found comparing EHR data to community screening data for CVD risk factors, but 1 study compared colorectal cancer screening rates from EHR data to selfreported data in California. Those results found up to 9 percentage points higher self-reported cancer screening rates in the survey relative to actual screening rates in the EHR.4 One potential limitation of the 2 methods compared is related to how representative the sample is to the target population. Missing data are of particular concern if the group with available data differs substantially from the group with missing data (ie, not missing at random) because prevalence estimates may be biased. One year of EHR data
FIG. 2 Comparison of risk factor estimates by source of data. BP, blood pressure; CVD, cardiovascular disease; EHR, electronic health record; HONU, Heart of New Ulm.
6
seems to be an efficient time frame by which prevalence estimates of CVD risk factors can be derived, showing no divergence from more complete 3-year estimates. The 1year time frame avoids the potential for seasonality effects in smaller time windows and, relative to longer time frames, also potentially avoids ‘‘averaging out’’ secular trends or unnecessary delays in getting enough data to establish point estimates (for baseline or follow-up). The utility of community screenings to provide a complete picture of population health seems more nuanced though. When opened up to the whole community on a volunteer basis, as was done in HONU, there is potential for bias. Interestingly, however, such biases in the HONU project did not occur in one direction. Researchers are often critical that health promotion programs tend to attract those who need them least (ie, the ‘‘worried well’’).17–19 The number of individuals in the 1-year EHR data set (2447; 31% of the population) was roughly similar to the number in the screening data set, which also occurred within a single year (3123; 40% of the population), but the samples clearly differed by risk factor status. There was evidence of a wellness bias toward screenings, as the proportion of smokers and obese individuals was lower in the screening data set as compared to the EHR. Paradoxically, however, there also was evidence of a wellness bias in the EHR, as the proportion of individuals with uncontrolled BP and high total cholesterol was lower than in the screening data set, indicating improved management of these risk factors. Reasons for such discrepancies are speculative at this point, but could be a function of several factors. Smoking and obesity are associated with chronic diseases, which may drive individuals to seek medical attention and thus have higher representation in the EHR (and perhaps also gives them a sense of less need to be screened). The higher levels of uncontrolled BP and high cholesterol observed in the screening data set may be an indication of low awareness of such conditions because they are often asymptomatic and therefore individuals may not seek medical care for them. More research is currently planned in this area to determine if the community screenings indeed detected undiagnosed hypertension and dyslipidemia, which could eventually result in better downstream control. Despite the differences identified between the EHR and community screening data, and the potential bias in estimates from each source, actions to address CVD risks in a local community health initiative informed by either set of measures may not have differed substantially. Additionally, the relatively high proportion of the community represented within the record indicates that EHR data can provide reasonable estimates of CVD risk factors and changes in those risks over time in the context of population health planning. Although EHR systems lack systematic measures of lifestyle CVD risks often collected in the context of community screenings, it is possible that movements toward population health in health care could promote the incorporation of such measures into the EHR. The use of EHR data for community assessment offers a more affordable method than primary data collection. The utility of EHR data for surveillance purposes in the current study was particularly feasible given the single integrated system serving the area. The methods described here may be of particular importance for rural areas that are typically served by a single system. Approximately 23% of
SIDEBOTTOM ET AL.
the US population lives in a rural area and 2000 rural hospitals serve as the primary health care source for their geographic region.20 Rural communities may benefit from this method because they may not have access to public health estimates of health risks for their local geographic area or the resources to collect data. Other studies should be conducted in communities served by multiple providers using either single-provider EHR estimates or data from health information exchanges. Additional studies are needed to confirm the present study’s finding that estimates from EHR and community assessment data are similar enough that the policies and approaches applied to address risk would be the same. If so, EHR data is likely the quickest and most financially feasible source of community surveillance data. EHRs have enormous potential to be a rich source of data to support population health surveillance. In order to optimize the EHR for population health surveillance, increased use of EHRs should be coupled with the use of health surveillance tools and data collection methods that enable both clinical and population health management goals. EHR surveillance methods are becoming more routine as part of health care systems’ information management practices and may be able to provide an efficient and affordable picture of a given population’s CVD health. Author Disclosure Statement
Ms. Sidebottom, Dr. Johnson, Dr. VanWormer, Mr. Sillah, Ms. Winden, and Ms. Boucher declared no conflicts of interest with respect to the research, authorship, and/or publication of this article. The authors received no financial support for this article. References
1. Jha AK, Burke MF, DesRoches C, et al. Progress toward meaningful use: hospitals’ adoption of electronic health records. Am J Manag Care. Dec 2011;17(12 Spec No.): SP117–SP124. 2. Arsand E, Walseth OA, Andersson N, et al. Using blood glucose data as an indicator for epidemic disease outbreaks. Stud Health Technol Inform. 2005;116:217–222. 3. Allen-Dicker J, Klompas M. Comparison of electronic laboratory reports, administrative claims, and electronic health record data for acute viral hepatitis surveillance. J Public Health Manag Pract. 2012;18:209–214. 4. Palaniappan LP, Maxwell AE, Crespi CM, Wong EC, Shin J, Wang EJ. Population colorectal cancer screening estimates: comparing self-report to electronic health record data in California. Int J Canc Prev. 2011;4(1). pii 28540. 5. Blumenthal D, Tavenner M. The ‘‘meaningful use’’ regulation for electronic health records. N Engl J Med. 2010; 363:501–504. 6. VanWormer JJ. Methods of using electronic health records for population-level surveillance of coronary heart disease risk in the Heart of New Ulm Project. Diabetes Spectrum. 2010;23:161–165. 7. Johnson PJ, Blewett LA, Davern M. Disparities in public use data availability for race, ethnic, and immigrant groups: national surveys for healthcare disparities research. Med Care. 2010;48:1122–1127. 8. Kochanek KD, Jiaquan X, Murphy SL, Minino AM, Kung HC. Deaths: Final Data for 2009. National Vital Statistics Reports. 2012;60(3):1–117.
EHR AS A POPULATION HEALTH SURVEILLANCE TOOL FOR CVD
9. Roger VL, Go AS, Lloyd-Jones DM, et al. Heart disease and stroke statistics—2012 update: a report from the American Heart Association. Circulation. 2012;125(1):e2–e220. 10. Boucher JL, Pereira RF, Graham KJ, Pettingill RR, Toscano JV, Henry TD. The Heart of New Ulm: a vision for the future. J Cardiovasc Transl Res. 2008;1:310–316. 11. VanWormer JJ, Johnson PJ, Pereira RF, et al. The Heart of New Ulm Project: using community-based cardiometabolic risk factor screenings in a rural population health improvement initiative. Popul Health Manag. 2012;15:135–143. 12. McLeroy KR, Bibeau D, Steckler A, Glanz K. An ecological perspective on health promotion programs. Health Educ Q. 1988;15:351–377. 13. VanWormer JJ, Fyfe-Johnson A, Boucher JL, et al. Stress and workplace productivity loss in the Heart of New Ulm project. J Occup Environ Med. 2011;53:1106–1109. 14. Blackwell DL, Lucas JW, Clarke TC. Summary health statistics for U.S. adults: national health interview survey, 2012. Vital and Health Statistics. 2014;10(260):95. 15. Green BB, Anderson ML, Cook AJ, et al. Using body mass index data in the electronic health record to calculate cardiovascular risk. Am J Prev Med. 2012;42:342–347. 16. Wei WQ, Leibson CL, Ransom JE, Kho AN, Chute CG. The absence of longitudinal data limits the accuracy of high-throughput clinical phenotyping for identifying type
17.
18. 19.
20.
7
2 diabetes mellitus subjects. Int J Med Inform. 2013;82: 239–247. Martinson BC, Crain AL, Sherwood NE, Hayes MG, Pronk NP, O’Connor PJ. Population reach and recruitment bias in a maintenance RCT in physically active older adults. J Phys Act Health. 2010;7(1):127–135. Mechanic D. Population health: challenges for science and society. Milbank Q. 2007;85:533–559. Pirie PL, Elias WS, Wackman DB, et al. Characteristics of participants and nonparticipants in a community cardiovascular disease risk factor screening: the Minnesota Heart Health Program. Am J Prev Med. 1986;2(1):20–25. American Hospital Association. The Opportunities and Challenges for Rural Hospitals in the Era of Health Reform. Washington, DC: American Hospital Association; 2011:1–2.
Address correspondence to: Abbey Sidebottom Allina Health 2925 Chicago Ave, MS 10039 Minneapolis, MN 55407-1321 E-mail:
[email protected] 