Abstract:
Objective: Atrial fibrillation (AF) is the most widely recognized cardiovascular arrhythmia. Symptomatic AF can decrease quality of life, functional condition, and performance of the heart. In order to make a clinical decision strategy, the present study attempts to investigate the optimal number of cross validation (CV) [NOCV] on the prediction of postoperative AF based on the several data mining and voting ensemble approaches. Material and Methods: The retrospective dataset included complete medical records of 2888 individuals after coronary artery bypass grafting. The subjects were divided into two groups: AF group (n=360) and non-AF group (n=2528), respectively. Data mining approaches including artificial neural networks (ANN), Naïve Bayes (NB) and logistic regression (LR) were constructed for the prediction of the presence or absence of AF. Additionally, voting ensemble strategy was employed in order to improve predictive accuracy. NOCV was optimized by Grid search. For evaluating the predictive performance, accuracy and area under curve (AUC) of the receiver operating characteristics (ROC) graph were considered as evaluation index. Results: After removing the missing values and outliers, this research consisted of 2694 subjects; 327 (12.1%) in AF group and 2367 (87.9%) in non-AF group, respectively. The largest accuracy of 87.8% for LR was observed in the 8-fold CV. Similarly, voting ensemble yielded the highest AUC value of 0.896 in the 10-fold CV. The highest accuracy and AUC were 87.2% and 0.896 for voting, 87.8% and 0.729 for LR, 87.1% and 0.668 for ANN, and 79.4% and 0.675 for NB, consecutively. Conclusion: The results of the current research demonstrated that the constructed models yielded different results for predicting postoperative AF in the determination of the optimal NOCV. It is recommended to determine the optimal NOCV by optimization techniques. The proposed voting was an acceptable and promising approach for predicting AF, and thence, can support clinicians in the clinical decision making.
