Materials and Methods
We analyzed data from a completed clinical trial that randomized 100 patients with stable heart failure and impaired systolic function to receive either 12 weeks of 1-hour group tai chi classes or time-matched heart failure education (the control) in addition to usual care. Methodological details and the primary trial results have been published elsewhere (Yeh et al. 2011). Patients who chose to participate in the study provided written informed consent; this study was approved by the institutional review boards of all participating institutions.
Participants were recruited from ambulatory clinics (primary care, general cardiology, and specialty heart failure practices) at three academic medical centers in and around Boston, Massachusetts. The inclusion criteria were physician diagnosis of chronic systolic heart failure, left ventricular ejection fraction ≤ 40% in the past 2 years; stable medical regimen that was defined as no major changes in medication in the past 3 months; and a designation of class I, II, or III for heart failure as defined by the New York Heart Association (Criteria Committee of the New York Heart Association 1994). The exclusion criteria were unstable angina or myocardial infarction in the past 3 months, major cardiac surgery within the past 3 months, history of cardiac arrest in the past 6 months, history of cardiac resynchronization therapy in the past 3 months, unstable serious ventricular arrhythmias, unstable structural valvular disease, current participation in a conventional cardiac rehabilitation program, diagnosis of peripartum cardiomyopathy within the preceding 6 months, inability to perform a bicycle stress test, lower extremity amputation or other inability to ambulate because of conditions other than heart failure, severe cognitive dysfunction (Mini-Mental State Examination score ≤ 24; Folstein et al. 1975), inability to speak English, or regular practice of tai chi.
Biomarkers
Blood samples were drawn from participants in a nonfasting state at the time of the study visit. BNP was analyzed in whole blood collected in EDTA (ethylenediamine-tetraacetic acid) using a commercially available point-of-service meter (fluorescence immunoassay; Biosite Triage BNP Test; Biosite Diagnostics, San Diego, CA). Serum samples were also analyzed for CRP using DPC (Diagnostic Products Corporation) Siemens Immulite high sensitivity hsCRP immunoassay (Siemens AG, Munich, Germany), for endothelin-1 using DPC Siemens chemiluminescent QuantiGLO ELISA (Siemens AG), and for TNF using Quantikine TNF-α immunoassay (R&D Systems, Inc., Minneapolis, MN). Blood samples were drawn at baseline, at 6 weeks, and at 12 weeks. The percentages of the coefficient of variation (CV%) of the intraassay and interassay for these kits were 8.8–11.6% and 9.9–12.2%, respectively, for BNP; 4.2–6.4% and 4.8–10.0%, respectively, for CRP; 3.1% and 6.7%, respectively, for endothelin-1; and 5.3% and 8.4%, respectively, for TNF.
Pollution and Weather Data
Ambient temperature, relative humidity, dew point temperature, and barometric pressure were obtained from the National Weather Service (National Climatic Data Center 2011) daily summaries of meteorological data measured at Logan International Airport (Boston, MA). Apparent temperature is a metric used to describe how people perceive the combination of temperature and humidity (Steadman 1984). The values for apparent temperature are based on the measures of ambient and dew point temperature and were calculated using the following formula:
We obtained hourly measures of ambient fine particulate matter (PM ≤ 2.5 μm in aerodynamic diameter; PM2.5) from the Harvard Countway Library SuperSite (Boston, MA) and ozone from the averages of five U.S. Environmental Protection Agency monitors located within the Boston Area.
Statistical Methods
The biomarker data (CRP, BNP, endothelin-1, and TNF) were examined using summary statistics and distributional plots to identify outliers and assess normality. Log-transformation of biomarkers was performed for all analyses. Outliers were defined as values falling outside 1.5 times the interquartile range of the data. There were two outliers identified for endothelin-1 and one outlier for TNF. Analyses were conducted including all observations and also excluding outlier observations.
Separate models were constructed for one to four day moving averages of ambient and apparent temperature for each biomarker using fixed-effects regression in R (version 2.9; R Foundation for Statistical Computing, Vienna, Austria) with the plm package for repeated-measures data. Unlike mixed-effects models commonly used in unbalanced repeated-measures analyses, which borrow information at the population level to account for within- and between-person variability, fixed-effects regression controls for all between-person differences by estimating subject-specific intercepts, thus isolating the within-person estimates of association. Although this approach is more commonly used in the field of econometrics (Hausman 1978; Hausman and Taylor 1981), it also has previously been used in air pollution studies (Gold et al. 2000; Rückerl et al. 2006). In this model, each participant forms their own stratum analogous to a conditional logistic regression framework for binary outcomes in a case-crossover study. The model provides estimates for the time-varying covariates only, because time invariant covariates do not change within person. The model takes the general form
where yit is the outcome y at time t for the ith subject, Xit is a matrix of time-varying predictors, αi is a random intercept for each subject, and eit is the error term. In the primary analyses, ambient and apparent temperature were modeled as continuous linear functions. All models were adjusted for time and seasonality using a harmonic function [sine (2π × day of year/365.25) and cosine (2π × day of year/365.25)], day of week, and body weight at each visit as a marker of hydration status. Models of ambient temperature were additionally adjusted for corresponding moving-average values of relative humidity and barometric pressure. Models of apparent temperature were adjusted for barometric pressure only because the formula for apparent temperature accounts for the dew point temperature. We also tested for effect modification by diabetes status, sex, and randomization to the tai chi treatment arm of the study using cross-product terms and examining statistical significance at p < 0.05.
In a sensitivity analysis, we adjusted for corresponding moving averages of PM2.5 and ozone in separate models. We also included both linear and quadratic terms for relative humidity and pressure. To further explore potential nonlinear relationships, we used the mgcv package for generalized additive models (version 1.6; R Foundation for Statistical Computing) to fit penalized splines. We hypothesized that changes in temperature could be associated with different physiologic responses in summer and winter. Therefore, we performed sensitivity analyses by subsetting to season when visits began, defined as warm (March through August) or cool (September through February) seasons and also tested for interactions using a cross-product term and examining statistical significance of this term at p < 0.05. In a sensitivity analysis, we also fit linear-mixed models with random effects for all four biomarkers using the plm package. No material difference was observed, so we present results for fixed effects only. All results are presented as a percent change for a 5°C increase in either ambient or apparent temperature measure with 95% confidence interval (CI). This increment was chosen to be representative of day-to-day variability in temperature in the Boston area.