Algorithm for selecting predictors and prognosis of atrial fibrillation in patients with coronary artery disease after coronary artery bypass grafting

Abstract

Aim. To develop an algorithm for selecting predictors and prognosis of atrial fibrillation (AF) in patients with coronary artery disease (CAD) after coronary artery bypass grafting (CABG). Material and methods. This retrospective study included 886 case histories of patients with CAD aged 35 to 81 years (median age, 63 years; 95% confidence interval [63; 64]), who underwent isolated CABG under cardiopulmonary bypass. Eighty-five patients with prior AF were excluded from the study. Two groups of persons were identified, the first of which consisted of 153 (19,1%) patients with newly recorded AF episodes, the second — 648 (80,9%) patients without cardiac arrhythmias. Preoperative clinical and functional status was assessed using 100 factors. Chi-squared, Fisher, and Mann-Whitney tests, as well as univariate logistic regression (LR) were used for data processing and analysis. Multivariate LR and artificial neural networks (ANN) were used to develop predictive models. The boundaries of significant ranges of potential predictors were determined by stepwise assessment of the odds ratio and p-value. The model accuracy was assessed using 4 metrics: area under the ROC-curve (AUC), sensitivity, specificity, and accuracy. Results. A comprehensive analysis of preoperative status of patients made it possible to identify 11 factors with the highest predictive potential, linearly and nonlinearly associated with postoperative AF (PAF). These included age (55-74 years for men and 60-78 years for women), anteroposterior and superior-inferior left atrial dimensions, transverse and longitudinal right atrial dimensions, tricuspid valve regurgitation, left ventricular end systolic dimension >49 mm, RR length of 1000-1100 ms, PQ length of 170-210 ms, QRS length of 50-80 ms, QT >420 ms for men and >440 ms for women, and heart failure with ejection fraction of 45-60%. The metrics of the best predictive ANN model were as follows: AUC — 0,75, specificity — 0,73, sensitivity — 0,74, and accuracy — 0,73. These values in best model based on multivariate LR were lower (0,75; 0,7; 0,68 and 0,7, respectively). Conclusion. The developed algorithm for selecting predictors made it possible to verify significant predictive ranges and weight coefficients characterizing their influence on PAF development. The predictive model based on ANN has a higher accuracy than multivariate HR.