Electronic Appendix A
Electronic-Figure 1: Comparison of 100 blinded manual vs. automated measurement of P-wave amplitude in lead V1 and aVR.
Electronic Appendix B
Patient, Procedure, and Outcome Data
Demography: Gender, age at operation (years), race, body surface area (m2), body mass index (kg•m2), preoperative systolic and diastolic blood pressure (mmHg), preoperative pulse pressure
Preoperative symptoms: New York Heart Association functional class (I-IV), Canadian Angina class (0-4), emergency operation
Cardiac comorbidity: History of atrial fibrillation or flutter, history of ventricular arrhythmia, history of heart failure, previous cardiac operation, history of endocarditis, history of hypertension
Non-cardiac comorbidity: History of smoking, history of peripheral arterial disease, previous stroke, history of chronic obstructive pulmonary disease, history of renal disease (creatinine >2 mg•dL-1), history of treated diabetes mellitus
Valve pathology: Aortic valve regurgitation (grade 0-4), aortic valve stenosis (grade 0-5), mitral valve regurgitation (grade 0-4), mitral valve stenosis (grade 0-5), tricuspid valve regurgitation (grade 0-4), pulmonary valve regurgitation(grade 0-4), pulmonary valve stenosis (grade 0-5)
Echocardiography: Left ventricular ejection fraction (%), left atrial volume (cc)
Angiography: Number of coronary systems >50% stenosis, left anterior descending coronary artery disease (>70% stenosis), left circumflex artery disease (>70% stenosis), left main trunk disease (>70% stenosis), right coronary artery disease (>70% stenosis)
Quantitative
electrocardiography: Atrial heart rate (bpm), ventricular heart rate (bpm), left ventricular (LV) hypertrophy by Cornell criteria (μV), Cornell voltage duration product (ms), LV hypertrophy by Romhilt-Estes score (μV), LV hypertrophy by Sokolow-Lyon criteria (μV), Sokolow-Lyon voltage duration product (μV•ms), P-wave dispersion (ms), P-wave amplitude I (μV), P-wave amplitude II (μV), P-wave amplitude III (μV), P-wave amplitude aVF (μV), P-wave amplitude aVL (μV), P-wave amplitude aVR (μV), P-wave amplitude V1 (μV), P-wave amplitude V2 (μV), P-wave amplitude V3 (μV), P-wave amplitude V4 (μV), P-wave amplitude V5 (μV), P-wave amplitude V6 (μV), P-prime amplitude I (μV), P-prime amplitude II (μV), P-prime amplitude III (μV), P-prime amplitude aVF (μV), P-prime amplitude aVL (μV), P-prime amplitude aVR (μV), P-prime amplitude V1 (μV), P-prime amplitude V2 (μV), P-prime amplitude V3 (μV), P-prime amplitude V4 (μV), P-prime amplitude V5 (μV), P-prime amplitude V6 (μV), P-wave area I (1/[4.88*4] μV•ms), P-wave area II (mV•ms), P-wave area III (mV•ms), P-wave area aVF (mV•ms), P-wave area aVL (mV•ms), P-wave area aVR (mV•ms), P- wave area V1 (mV•ms), P-wave area V2 (mV•ms), P-wave area V3 (mV•ms), P-wave area V4 (mV•ms), P-wave area V5 (mV•ms), P- wave area V6 (mV•ms), P-wave axis (degrees), P-wave duration I (ms), P-wave duration II (ms), P-wave duration III (ms), P-wave duration aVF (ms), P- wave duration aVL (ms), P-wave duration aVR (ms), P-wave duration V1 (ms), P-wave duration V2 (ms), P-wave duration V3 (ms), P-wave duration V4 (ms), P-wave duration V5 (ms), P-wave duration V6 (ms), P-wave duration (median, all leads; ms), P-prime duration I (ms), P-prime duration II (ms), P- prime duration III (ms), P-prime duration aVF (ms), P-prime duration aVL (ms), P-prime duration aVR (ms), P-prime duration V1 (ms), P-prime duration V2 (ms), P-prime duration V3 (ms), P-prime duration V4 (ms), P-prime duration V5 (ms), P-prime duration V6 (ms), PR interval (median, all leads; ms), QRS area I (mV•ms), QRS area II (mV•ms), QRS area III (mV•ms), QRS area aVF (mV•ms), QRS area aVL (mV•ms), QRS area aVR (mV•ms), QRS area V1 (mV•ms), QRS area V2 (mV•ms), QRS area V3 (mV•ms), QRS area V4 (mV•ms), QRS area V5 (mV•ms), QRS area V6 (mV•ms), QRS axis (median beat; degree), QRS duration (all leads, median; ms), QT interval (all leads, median; ms)
Electrocardiographic
diagnoses: Right, left and bi-atrial enlargement, myocardial infarction (anterior, septal, inferior), presence of LV hypertrophy, presence of bundle branch block
Medication use: Preoperative ACE-inhibitor/angiotensin II receptor blocker, preoperative calcium channel blocker, preoperative beta blocker, preoperative inotrope, preoperative nitrate (oral and intravenous), preoperative digoxin
Laboratory data: Preoperative creatinine (mg•dL-1), creatinine clearance (mL•min-1•1.73 m-2), preoperative blood urea nitrogen
(mg•dL-1), preoperative hematocrit (%)
Operative Data
Support: Cardiopulmonary bypass time (min), aortic clamp time (min), off-pump coronary artery bypass grafting
Surgery type: Coronary artery bypass grafting, number of coronary arteries bypassed, number of internal thoracic artery grafts, aortic valve repair/replacement, mitral valve repair/replacement, tricuspid valve repair/replacement, pulmonary valve repair/replacement, number of postoperative fresh frozen plasma/platelet/red blood cell transfusions
Experience: Date of operation as interval to a fixed date before cohort surgeries
Outcomes data
Cardiac outcomes: Postoperative atrial fibrillation
Electronic Appendix C
Detailed Description of Model Selection and Model Comparison Techniques
Missing Data
Missing values for covariables were imputed with informative 5-fold multiple imputation using the Markov chain Monte Carlo (MCMC) algorithm(1) of PROC MI in SAS version 9.2 (©SAS Institute Inc., Cary, NC). A separate bagging analysis of complete cases confirmed our results derived from multiple imputation datasets.
Data Preparation
To describe relationships between continuous variables and outcome accurately, we considered their inverse, logarithmic, squared, and exponential transformations for variable selection. Restricted cubic spline transformation of continuous final model predictors was considered to improve final model fit.
Model Selection
Automated data-driven variable selection is subject to model instability, especially when multiple colinear related variables (e.g., P-wave variables) are considered for selection. The sequence of variable-entry into the model decides which of correlated variables are retained. The result of this phenomenon is poor reproducibility of prediction models derived from single data sets. Therefore, reproducibility appears to be inversely related to the number of variables available for adjustment.
To increase model stability and prediction accuracy, we identified predictors of postoperative AF with 500-fold bootstrap aggregation (bagging)(2). Automated stepwise multivariable logistic regression selected predictors of postoperative AF for each bootstrap sample at P<.01. Bootstrapped predictors were then aggregated by counting frequency of their retention across all 500 prediction models (reliability). Predictors with a reliability of 50% or greater (they were selected in ³50% of the 500 prediction models) contributed to the final bagged prediction model. To further account for colinearity, closely related variables (e.g., continuous variables and their nonlinear transformations, all P-wave measurements, LV hypertrophy indices, etc.) were clustered. If pre-defined variable clusters had a reliability of 50% or greater, variable constellations within the clusters determined which cluster variables were selected: If a constellation of three cluster variables had the highest reliability, the three with highest individual reliabilities were selected for the final bagged prediction model, even if individual reliability was less than 50%.
Individual and cluster reliabilities were obtained by aggregating the 5 multiple imputation data sets. Beta estimates were obtained from the first imputed data set and expressed as odds ratios and their 95% confidence intervals.
Model Evaluation
Discrimination of the prediction model was derived from the averaged C-statistic across all 5 multiple imputation data sets. Model calibration was assessed with 40-fold bootstrapping of the final prediction model to obtain a calibration curve of predicted versus observed probabilities and mean prediction error. Model fit was assessed with the Hosmer-Lemeshow goodness-of-fit test.
Model Comparison
Nested prediction models with (1) clinical predictors alone were compared to models with (2) clinical predictors and semi-quantitative ECG diagnoses of left, right and bi-atrial enlargement, and (3) clinical predictors and quantitative ECG measurements with integrated discrimination improvement of reclassification analysis (IDI)(3) and C-statistic. Integrated discrimination improvement (IDI) has been recognized as a more sensitive measure of model improvement with additional risk markers compared to changes in C-statistic(3, 4). It assesses reclassification as a continuous outcome across the range of probability of developing postoperative AF, where zero indicates no improvement of the model with new risk markers. To quantify the improvement of model discrimination, we describe relative IDI, defined as
Relative IDI= (EY1−EY0 ¸ EX1−EX0) −1 (3),
Where EY1 and EY0 indicate probabilities of patients with and without postoperative AF, respectively with inclusion of quantitative ECG predictors and EX1 and EX0 indicate probabilities of patients with and without postoperative AF, respectively without inclusion of quantitative ECG predictors in the model.
Evaluation of Clinical Correlates of P-wave amplitude in Lead aVR
To investigate echocardiographic and clinical variables associated with P-wave amplitude in lead aVR we applied linear regression with stepwise variable selection. Model calibration was assessed with bootstrap sampling and comparison of predictied verses observed P wave amplitude in aVR. R2 evaluated model fit of the data. Co-plots of continuous variables evaluated linearity of variable correlations, which were relaxed with restricted cubic splines.
The R functions calibrate, validate and improveProb of Frank Harrel’s Design library were used for these analyses (ãThe R foundation for Statistical Computing, www.r-project.org).
References
1. Rubin DB. Multiple imputation after 18+ years (with discussion). Journal of the American Statistical Association. 1996;91, 473-489.
2. Breiman L. Bagging predictors. Machine Learning. 1996;26:123–140.
3. Pencina MJ, D'Agostino RB S, D'Agostino RB,Jr, Vasan RS. Evaluating the added predictive ability of a new marker: From area under the ROC curve to reclassification and beyond. Stat Med. 2008 Jan 30;27(2):157,72; discussion 207-12.
4. Hlatky MA, Greenland P, Arnett DK, Ballantyne CM, Criqui MH, Elkind MS, et al. Criteria for evaluation of novel markers of cardiovascular risk: A scientific statement from the american heart association. Circulation. 2009 May 5;119(17):2408-16.
Electronic Appendix D
Electronic-Figure 2: Multivariable-adjusted relationship of age, left ventricular ejection fraction, left atrial volume, pulse pressure and P-wave amplitude in lead V1 with P-wave amplitude in lead aVR. Dashed lines represent 95% confidence intervals.