The PaTH Clinical Data Research Network (CDRN) initiative aims to use clinical data from electronic medical records and patient reported outcomes to answer questions of clinical importance to patients, providers, and other stakeholders [40]. The research infrastructure provided through the PaTH network provides opportunities to researchers seeking to conduct studies examining variables collected by the PaTH network. PaTH AF is a longitudinal cohort study at four academic medical centers [40]. PaTH AF has two components: 1) a cohort of over 95,302 individuals identified through electronic medical records, this dataset is thus limited to de-identified information from electronic medical records, and 2) a “Consented Cohort” of 953 consented individuals who were recruited from the larger cohort. The PaTH CDRN utilizes a central Institutional Review Board approach which governs the entire process. The institutional review board approved this study (JHU IRB00064600).

Inclusion criteria for the PaTH AF Cohort included: diagnosis of AF by ICD 9–10 codes or electrocardiographic reading, 3 non-emergency department visits since 1/1/2011, age ≥ 18, and ability to speak, read, and understand English. Individuals were excluded if they received the diagnosis of AF within a month of cardiac or abdominal surgery, or a thyroid-related diagnosis, and 12 months before or after prescription for methimazole or propylthioracil. Individuals in the PaTH AF Cohort were identified through electronic medical records using a computable phenotype [41] implementing the inclusion and exclusion criteria.

Participants were recruited for the Consented Cohort to investigate patient-reported outcomes beginning in August 2015. Individuals identified in the larger PaTH AF clinical database were contacted with study invitations through in-person, email, phone, patient portal messaging and post mail techniques. Participants were screened to ensure they were correctly identified as meeting the inclusion and exclusion criteria through an online survey.

At the time of study enrollment, participants completed baseline questionnaires of demographic and health status information, and participants completed surveys on patient-reported outcomes every 6 months following completion of the baseline questionnaire. Data collection was done online. Participants recruited through email and post mail completed surveys through REDCap (Vanderbilt University, TN), and participants who were recruited through the patient portal completed the surveys in MyChart (EPIC software, WI). Either MyChart or REDCap was used to enroll and collect data on patients recruited in-person and by phone based on patient preference.

The cohort completes surveys on patient-reported outcomes at baseline and every 6 months. Healthcare utilization, procedures, medications, diagnoses, body mass index, laboratory data, and demographics are collected through electronic medical records. This analysis consists of participants’ first 3 patient-reported outcome surveys (baseline, 6 months, and 1 year), and electronic medical record data collected through the time of the 1-year post-enrollment survey.

The PaTH network is committed to high quality data and uses the PCORnet Common Data Model, a defined set of health data in a structured, consistent format [40]. The Common Data Model design prioritizes analytic functionality and a parsimonious approach to ensure the electronic medical record data is accurate and consistent across sites. Additionally, each site had a trained research assistant and coordinator to oversee the quality of the data collected via surveys.

Age, sex, education, and marital status were determined by self-report. Comorbidities, body mass index (BMI), and medication prescriptions were determined through electronic medical record data. Comorbidities were measured using the grouped Charlson comorbidities index score, which is calculated using the Deyo version of the Charlson comorbidities index [42, 43]. The grouped Charlson comorbidities index score rates comorbidities on a 0 to 2 scale.

Symptoms and quality of life were measured using the Atrial Fibrillation Effect on QualiTy of Life (AFEQT) [44]. The AFEQT is a disease-specific quality of life instrument consisting of four subscales: symptoms, treatment concerns, daily activities, and treatment satisfaction. The subscales for symptoms, treatment concerns, and daily activities are added together to create an AF-related quality of life score. The AFEQT has high internal consistency (Cronbach’s α: > 0.8) and has been validated in a large, community-based cohort [7, 9, 44, 45]. In this sample, the AFEQT had a Cronbach’s alpha of 0.93. The symptoms subscale consists of 4 items rated on a 7-point Likert scale and contains questions asking patients to report in the past 4 weeks, how much were they bothered by palpitations, an irregular heart beat, a pause in heart activity, and lightheadedness or dizziness as a result of their AF. The symptoms subscale was analyzed separately from the AF-related quality of life total score since patients’ reported symptoms are the primary focus of AF therapies. The symptoms subscale score is formed by calculating the mean score of the questions answered. The AFEQT total scale scores range from 0 to 100, with 0 representing the worst possible quality of life, and 100 representing the best (no impairment due to AF).

Emotional and functional status were measured using the Patient-Reported Outcomes Measurement Information System (PROMIS)-29 emotional distress (anxiety and depression) and physical function domains [46]. The PROMIS initiative was established to develop standardized, reliable, and valid item banks for measuring patient-reported outcomes that are highly accessible to researchers, patients and clinicians in multiple languages and free-of-charge [46‐48]. The emotional distress (4 items on anxiety and 4 items on depression) and physical function domains (4 items) are rated on a 5-point Likert-type scale. High internal consistency have been reported for both emotional (Cronbach’s α: > 0.8) [46, 48] and physical function domains (Cronbach’s α: > 0.8) [46, 48]. The raw scores of the anxiety, depression, and functional status domains were converted to a scaled T-score that range from 41 to 79.4 for depression, 40.3 to 81.6 for anxiety, and 22.9 to 56.9 for function using PROMIS scoring guidelines [49]. The scaled T-score is created based on PROMIS scores mean of 50 and standard deviation of 10 in a generalized, referent population: scores 0.5–1.0 standard deviation worse than the mean are interpreted as mild symptoms or impairment, scores 1–2 standard deviations are interpreted as moderate symptoms or impairment, and scores 2 standard deviations or more worse than the mean is interpreted as severe symptoms or impairment.

Rate and rhythm control medication were included as covariates in all final models. Patients were classified as receiving rate control medication or rhythm control medication if they had been prescribed a medication in an outpatient setting in the respective category within 1 year prior to their baseline questionnaire. Medications were categorized based on the American College of Cardiology’s recommendations [50]. Rate control medications included amiodarone, atenolol, carvedilol, diltiazem, metoprolol, and verapamil. Rhythm control medications included amiodarone, dofetilide, dronedarone, flecainide, ibutilide, propafenone, and sotatol. Ablation was classified based on provider billing code CPT 93656 within the past 3 years. Cardioversion was classified based on provider billing codes CPT 92960, 92,961, or ICD-9 99.61 within the past 3 years.

We used the statistical software STATA, version 15.0 (StataCorp, College Station, TX), for all analyses. We checked model assumptions of normality before statistical analyses. We calculated descriptive statistics for all variables included in the analysis to examine means, standard deviations, shapes of distributions for continuous variables, and frequencies for categorical variables. Patterns of missing data were assessed for each variable included in the models, and, for variables with greater than 5% missing, missing data was imputed using multiple imputation [51, 52]. We created 100 imputations for each patient-reported outcome (symptoms, AF-related quality of life, anxiety, and depression) since these 13% of these measures were missing at baseline, and 33% of these measures were missing across all 3 survey times (baseline, 6 months, and 1 year). Missing data across patient-reported outcomes did not significantly differ by sex, comorbidities, age, education, or BMI, however, missing data did significantly differ by site. When we compared results using completed-cases only against results using multiple imputation, we found that inferences did not differ across the models (see Appendix). For all of regression analyses described below, we examined multi-collinearity within each set and among all variables by examining the variance inflation factor (VIF). Variables with a high VIF (0.8 or greater), which is indicative of multi-collinearity, were considered for removal from the model. Patient reported health status was collinear with functional status and emotional distress, and thus was removed from the model.

Longitudinal multiple regressions with random effects tested the associations of individual characteristics to patient-reported outcomes over time. The Lagrangian multiplier test determined that random effects was appropriate (p < 0.05). Quality of life, symptom severity, emotional status and functional status were each treated as separate dependent variables. Variables were entered as sets into the model with individual characteristics entered in the first step, followed by comorbidities, BMI, and rate and rhythm control medications in the second step. Age in years was divided into 4 categories as follows: 1) > 18 and < 60, 2) > 60 and < 70, 3) > 70 and < 80, and 4) > 80. Site was dummy-coded and included in all models. The Wald test was used to test categorical variables. In the third and final set of models, change in patient-report outcomes over the 6-month intervals by individual characteristics was examined. The interaction of female sex and time, age group and time, and education level and time were each examined in separate models that accounted for comorbidities, BMI, rate control medications, and rhythm control medications.

Table 1 summarizes demographic data specific to all participants stratified by sex. Our study population (n = 953) was 65% male (n = 616), 93% white (n = 890), and 72 (±10) years old on average. The majority (63%) had at least a college degree. A diagnosis of heart failure was present in 28% of participants (n = 364), and the average Charlson Comorbidity Index among participants was 1.61 (±2.08). The majority of the population was overweight (35%) or obese (44%). The patient-reported outcomes of this sample were each lower than the general population mean of 50 [46]: the mean functional status of our population was 47.3 (±8.8), the mean anxiety score of our population was 48.9 (±8.8), and the mean depression score of our sample was 46.9 (±8.1). Participants reported a mean score of 2.21 (±1.40) on the AF symptoms subscale [1‐7], which can be interpreted as the participants were on average “a little bothered” by their AF symptoms. On average, the AFEQT total score was 62.72 (25.80) on a scale of 0 to 100, with 0 being the lowest AF-related quality of life.

Longitudinal multiple regression results testing individual characteristics associated with patient-reported outcomes are summarized in Table 2. In models including only individual characteristics (sex, education, and age) and patient-reported outcomes, female sex was associated with higher AF symptoms (β 0.19, 95% CI: 0.13, 0.46), anxiety (β 1.79, 95% CI: 0.65, 2.93), and depression (β 1.14, 95% CI: 0.10, 2.18), and lower function status (β − 2.03, 95% CI: -3.22, − 0.85). Female sex was not associated with AF-related quality of life in the individual characteristics model. In longitudinal models adjusting for illness characteristics (BMI, comorbidities, and heart failure) and rate and rhythm control medication prescription, female sex was significantly associated with higher symptoms of AF (β 0.14, 95% CI: 0.02, 0.25), anxiety (β 1.90, 95% CI: 0.7, 3.10) and depression (β 1.39, 95% CI: 0.27, 2.50), and poorer AF-related quality of life (β − 2.89, 95% CI: -5.74, − 0.03) and functional status (β − 1.99, 95% CI: -3.19, − 0.80). While women tended to have poorer patient-reported outcomes, female sex was not associated with illness characteristics on patient-reported outcomes (Fig. 1).

In models accounting for education, age, comorbidities, BMI, and rate and rhythm control medication, the association of female sex and patient-reported outcomes varied over time. Women reported significantly higher AF symptoms at baseline (β 0.29, 95%CI: 0.08, 0.50). The change in AF symptoms over time did not differ across sexes. Women reported higher symptoms of anxiety at baseline (β 2.08, 95% CI: 0.76, 3.40), 6 months (β 2.28, 95% CI: 0.77, 3.80) and 1 year (β 1.82, 95% CI: 0.32, 3.32). At baseline, women reported significantly higher symptoms of anxiety (β 1.44, 95% CI: 0.25, 2.63) but the change over time was non-significant between sexes. Women reported poor functional status at baseline (β − 2.23, 95% CI: -3.52, − 0.94), 6 months (β − 2.94, 95% CI: -4.39, − 1.48) and 1 year (β − 3.18, 95% CI: -4.62, − 1.75). While women reported poorer AF-related quality of life at baseline (β − 4.12, 95% CI: -8.10, − 0.14), the change over time in AF-related quality of life was not significantly different across sexes.

In longitudinal multiple regression models including only individual characteristics (age, sex, and education) with patient-reported outcomes, there was a significant decrease in symptoms of anxiety, depression, and AF with each increase in age category, from age > =60 and < 70, age > =70 and < 80, and age > =80 (Fig. 2). There was no association between AF-related quality of life and age. Compared to individuals under age 60, individuals between the ages of 70 and 80 (β − 2.68, 95% CI: -4.50, − 0.87) and older than 80 (β − 5.98, 95% CI: -7.94, − 4.02) reported poorer functional status.

In longitudinal models adjusting for illness characteristics (BMI, comorbidities, and heart failure) and rate and rhythm control medication prescription, there remained a decrease in symptoms of anxiety, depression, and AF with each increase in age category, from age > =60 and < 70, age > =70 and < 80, and age > =80. There remained no association between AF-related quality of life and age. Compared to individuals under age 60, individuals between the ages of 70 and 80 (β − 2.30, 95% CI: -4.11, − 0.50) and older than 80 (β − 6.76, 95% CI: -8.76, − 4.77) reported poorer functional status.

In models accounting for education, age, comorbidities, BMI, and rate and rhythm control medication, older age groups reported significantly lower symptoms of AF, anxiety, and depression compared to individuals under age 60 at baseline, 6 months, and 12 months (Table 3). With the exception of adults between the ages of 60 to 70 at baseline, older age groups reported poorer functional status at baseline, 6 months, and 12 months. Individuals between the ages of 60 to 70 reported better AF-related quality of life at 12 months (β 9.54, 95% CI: 2.57, 16.52) compared to individuals under the age of 60. AF-related quality of life at baseline and the change in AF-related quality of life at 6 months and 12 months was not significantly different in individuals over the age of 70 compared to individuals under the age of 60.

In longitudinal multiple regression models including only individual characteristics (age, sex, and education) with function, anxiety, and depression, functional status decreased with education level. Compared to individuals with greater than a college education, Individuals with a college education (β − 2, 95% CI: -3.51, − 0.49), individuals with some college (β − 2.5, 95% CI: -3.99, − 1.01), and individuals with high school or less (β − 2.47, 95% CI: -4.17, − 0.76) reported poorer functional status. Individuals with some college education reported higher symptoms of AF (β 0.14, 95% CI 0.01, 0.27), anxiety (β 1.73, 95% CI: 0.33, 3.13), and depression (β 1.11, 95% CI: 0.15, 2.38) compared to individuals with greater than a college education. Individuals with some college education reported poorer AF-related quality of life (β − 5.34, 95% CI: -9.19, − 1.49) than individuals with greater than a college education.

In longitudinal multiple regression models adjusting for illness characteristics and rate and rhythm control medication prescription, there was no association between education level and symptoms of depression and AF-related quality of life. Compared to individuals with greater than a college education, Individuals with a college education (β − 1.56, 95% CI: -3.04, − 0.09), individuals with some college (β − 1.77, 95% CI: -3.23, − 0.31), and individuals with high school or less (β − 2.66, 95% CI: -4.35, − 0.98) reported poorer functional status. Individuals with some college education reported higher symptoms of AF (β 0.32, 95% CI 0.12, 0.52) and anxiety (β 1.64, 95% CI: 0.19, 3.09) compared to individuals with greater than a college education.

In models accounting for education, age, comorbidities, BMI, and rate and rhythm control medication, individuals with lower levels of education reported significantly poorer functional status baseline, 6 months, and 12 months compared to individuals with greater than a college education. Individuals with some college reported higher symptoms of AF at baseline (β 0.42, 95% CI: 0.17, 0.68) and anxiety at baseline (β 1.86, 95% CI: 0.26, 3.45), at 6 months (β 2.71, 95% CI: 0.82, 4.61), and 12 months (β 2.13, 95% CI: 0.33, 3.94), and poor AF-related quality of life at baseline (β − 4.41, 95% CI: -8.25, − 0.57) compared to individuals with greater than a college education.

Site, BMI, comorbidities, and rate and rhythm control medication prescription in the past year were included as covariates across all final models. Site was associated with anxiety (β − 1.38, 95% CI: -1.94, − 0.82), functional status (β 0.96, 95% CI: 0.39, 1.53), and (β 3.80, 95% CI: 1.79, 5.82). BMI was associated with poorer AF-related quality of life (β 0.58, 95% CI: -0.83, − 0.33) and higher symptoms of depression (β 0.09, 95% CI: 0.01, 0.18). Comorbidities were associated with poorer functional status (β − 1.04, 95% CI: -1.63, − 0.44). Rate medication was associated with poorer functional status (β − 2.16, 95% CI: -3.44, − 0.87) and higher symptoms of depression (β − 1.41, 95% CI: 0.11, 2.70). Rhythm control medication was association with higher symptoms of AF (β 0.35, 95% CI: 0.19, 0.52).

These data represent observations from a multi-institutional cohort study. Participation in this study is voluntary and, thus, biases exist in patient selection, participating sites, and reporting. We performed multiple imputation to account for reporting bias due to missing data, and found that there was no difference in our conclusions when we examined completed cases only. Despite the use of multiple recruitment methods, the study sample was overwhelmingly white. The larger PaTH clinical database of individuals with AF was majority white (90%), which may be reflective of the population in treatment for AF [37, 39]. This pattern is consistent throughout AF clinical trials [37], there is a lower prevalence of this disease among black and Hispanic populations which adds to the challenge of enrolling a racially diverse group of patients. A further limitation is that we did not measure prior physical activity. There is evidence that endurance athletes are at a heightened risk for atrial fibrillation [63, 64], and accounting for prior athletic activity may have added to our understanding of the differences in atrial fibrillation experience by individual characteristics.

Importantly, this study relies on electronic medical record data for information on rate and rhythm control medication, BMI, and comorbidities. At least 3 non-inpatient visits in the 4 years prior to study enrollment was inclusion criteria to ensure enrolled participants had adequate and accurate electronic medical record data available. However, changes in medications, new diagnoses, and weight fluctuations may have occurred at visits to health systems not included in the PaTH network in parallel to the time of this study, and thus not been captured in this dataset. Site was included as a variable in all models to adjust for differences in both EMR data collection across institutions, and varying patient populations. Site was statistically significant across all models, which highlights the importance of taking site into consideration in multi-institutional research.