Study population
The assessment tools were derived using baseline data from the REasons for Geographic And Racial Differences in Stroke (REGARDS) study population. REGARDS is a nation-wide prospective cohort study of 30,239 men and women aged 45 and older. Details of the study design and recruitment have been previously published [4]. REGARDS participants were recruited between January 2003 and October 2007. Participants were identified using commercially available lists with the goal of equal representation of male and female and African-American and white participants. Geographically, participants were recruited to obtain a population with 20% residing in coastal North Carolina, South Carolina, and Georgia (Stroke Buckle), 30% from the remaining areas of North Carolina, South Carolina, and Georgia plus Tennessee, Mississippi, Alabama, Louisiana, and Arkansas (Stroke Belt), and 50% from the other 40 contiguous US states and the District of Columbia [4].
Potential participants were contacted by telephone. After verbal consent was obtained, data including demographic and socioeconomic factors, medical history, psychosocial factors, lifestyle factors, and cognitive function were collected using computer-assisted telephone interviews (see Additional file 1 for the data collection forms). Following the interview, a technician conducted an in-home physical exam, collected blood and urine, and measured anthropometrics. Written informed consent was obtained during the home visit. A food-frequency questionnaire and a family history questionnaire were left with participants to be completed and returned by mail. This research was conducted in accordance with the Declaration of Helsinki. The REGARDS study was approved by the Institutional Review Boards of the University of Alabama at Birmingham, the University of Vermont, Wake Forest University, and the University of Cincinnati.
A modified 7-lead ECG was used until April 2004, and after this point a standard 12-lead ECG was obtained [4]. Because 7-lead ECGs miss anterior and some lateral MIs, we excluded 9,043 participants without a 12-lead ECG. We additionally excluded participants who reported a history of myocardial infarction or revascularization procedures (n = 2,919). Finally, we excluded individuals who had missing data on potential correlates of UMI (n = 1,624) in order to conduct a complete-case analysis. After these exclusions, there were 16,653 participants with data available for development of the assessment tools.
Measurement of covariates
The REGARDS data collection instruments contained several standard, previously validated questionnaires used in health research, and additional items selected from questionnaires used in prior studies such as the Behavioral Risk Factor Surveillance System, the Atherosclerosis Risk In Communities Study, the Multi-Ethnic Study of Atherosclerosis, and the Greater Cincinnati/Northern Kentucky Stroke Study [4]. Pilot testing was conducted to ascertain the feasibility of the study processes in the target population [4]. Participants self-reported information about demographics and health behaviors. Urbanization was determined by the census tract in which the participant resided and was categorized as < 25% urban, 25%-75% urban, and >75% urban. Educational attainment was categorized as less than high school, high school graduate, some college, or college graduate or above. Income was categorized < $20,000, $20,000-$34,999, $35,000-$74,999, ≥ $75,000 per year, or refused. Participants answered a single question about current health insurance (yes or no). Participants self-reported regular aspirin use (yes or no) and time spent watching television or videos (none, 1–6 hours/week, 1 hour/day, 2 hours/day, 3 hours/day, or 4 or more hours/day). They additionally reported information about alcohol consumption, cigarette smoking, and exercise. Alcohol consumption was classified as none, moderate (0–7 drinks per week for women and 0–14 drinks per week for men), or heavy (>7 drinks per week for women and >14 drinks per week for men). Cigarette smoking was categorized as current, past, and never. Exercise was categorized as none, 1–3 times per week, or 4 or more times per week. Medication adherence was assessed using the 4-item Morisky scale [5]. Each item in the Morisky scale was considered separately.
Clinical measurements were obtained during the in-home study visit by trained technicians. Body mass index was calculated using measured height and weight. Two blood pressure measurements were taken, and the values averaged. The technicians obtained blood and urine specimens which were processed and shipped to a central laboratory. HDL cholesterol, LDL cholesterol, triglycerides, C-reactive protein, white blood cell count, and serum creatinine were measured in the blood samples. Albumin and creatinine were measured in the urine. Estimated glomerular filtration rate (eGFR) was calculated using the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) equation [6]. Because of the nonlinear relationship of eGFR with adverse health effects, we categorized it as ≥90 mL/min/1.73 m2, 60–89 mL/min/1.73 m2, 45–59 mL/min/1.73 m2, 30–44 mL/min/1.73 m2, and <30 mL/min/1.73 m2. Similarly, the urinary albumin to creatinine ratio was categorized as < 30 mg/g, 30–300 mg/g, and >300 mg/g. Heart rate was measured by ECG. The 6-item screener for cognitive function was administered in the telephone interview [7]. A score of 4 or less was considered cognitive impairment.
Participants reported their history of stroke, transient ischemic attack, deep vein thrombosis, peripheral vascular disease (amputation or surgery), dialysis for kidney failure, falls, diabetes, hypertension, and dyslipidemia. They also reported whether they were currently taking antihypertensive medications. Family history of MI was defined as report of MI in the mother or father of the participant regardless of age at the time of the event. We created variables for unrecognized diabetes, hypertension, and hyperlidipidemia. A participant was considered to have unrecognized diabetes if they reported no prior diagnosis, but had fasting serum glucose ≥126 mg/dL or nonfasting serum glucose ≥200 mg/dL. Similarly, participants were considered to have unrecognized hypertension if they reported no prior diagnosis, but had systolic/diastolic blood pressure ≥140/90 mmHg. Unrecognized dyslipidemia was defined as no self-reported history and total cholesterol ≥240 mg/dL, LDL cholesterol ≥ 160 mg/dL, or HDL cholesterol ≤ 40 mg/dL.
Self-reported health scales included overall health status (as excellent, very good, good, fair, or poor), the 4-item Cohen’s Perceived Stress Scale [8], the 4-item Center for Epidemiologic Studies Depression Scale (CESD-4) [9], and the Medical Outcomes Study Short Form-12 (SF-12) from which physical and mental component scores were calculated [10]. Participants self-reported stroke symptoms, using the Questionnaire for Verifying Stroke-Free Status [11], and whether they needed two or more pillows to sleep at night or woke during the night because of breathlessness.
Electrocardiogram reading and coding
ECGs performed during the in-home visit were read and coded by a central reading center at Wake Forest University by research staff blinded to other study data. Research staff received initial training and certification followed by quarterly quality control examinations with retraining and recertification when necessary. ECGs were read by a certified coder, followed by a review by a senior coder, and a final review by the center PI (EZS) which focused on major abnormalities. Disagreements were resolved through discussion between the coder, senior coder, and PI. When necessary, a senior coder who had not previously read the ECG in question was asked to resolve the debate. Participants were considered to have UMI if they reported no history of MI, but had evidence of MI on ECG. This included major Q/QS wave abnormalities by the Minnesota code (MC) scheme (MC 1-1-X through 1-2-X) or smaller Q/QS wave abnormalities (MC 1-3-X) with major ST-segment or T-wave abnormalities (MC 4–1, 4–2, 5–1 or 5–2) [12].
Statistical analysis
We computed means and standard deviations or percentages for each of the potential correlates for participants with and without UMI. We additionally calculated each participant’s Framingham risk score for coronary heart disease [13]. The groups with and without UMI were compared using linear regression for continuous variables and χ2 tests for categorical variables. Using logistic regression, we calculated the crude odds ratio and the c-statistic for each potential correlate. We used a logistic regression approach to create the basic assessment tool. Basic demographics (age, sex, and race), smoking and alcohol consumption, self-report of medical history, family history of MI, and basic clinical measurements (body mass index and blood pressure) available during routine office visits were entered into a logistic regression model. For continuous variables, the linear term and the square of the centered value were entered into the model. We then used a backward selection procedure to create a more parsimonious model, requiring a p < 0.20 for a variable to remain. Variables which were removed from the model through backward selection were reconsidered, and if they improved the c-statistic by ≥ 0.005, they were retained. The model was used to calculate the predicted probability of UMI for each participant, and a receiver operating characteristic curve was constructed. We examined the performance of the UMI assessment tools in subgroups defined by sex, race, and age (< 65 years or ≥ 65 years). The procedure was repeated for the expanded set of potential correlates described above including demographics, health behaviors, clinical measurements, medical history, participant reported health scales, and participant reported symptoms (expanded assessment tool). The ability of the Framingham risk score for coronary heart disease to discriminate between those with and without UMI was evaluated using the c-statistic and receiver operating characteristic curve. Analysis was conducted using SAS version 9.2 (Cary, NC), and p values < 0.05 were considered statistically significant.