Background: New-onset postoperative atrial fibrillation (POAF) is the most common complication after valvular surgery, but its etiology and risk factors are incompletely understood. This study investigates the benefits of machine learning methods in risk prediction and in identifying relative perioperative variables for POAF after valve surgery.
Methods: This retrospective study involved 847 patients, who underwent isolated valve surgery from January 2018 to September 2021 in our institution. We used machine learning algorithms to predict new-onset postoperative atrial fibrillation and to select relatively important variables from a set of 123 preoperative characteristics and intraoperative information.
Results: The support vector machine (SVM) model demonstrated the best area under the receiver operating characteristic (AUC) value of 0.786, followed by logistic regression (AUC = 0.745) and the Complement Naive Bayes (CNB) model (AUC = 0.672). Left atrium diameter, age, estimated glomerular filtration rate (eGFR), duration of cardiopulmonary bypass, New York Heart Association (NYHA) class III-IV, and preoperative hemoglobin were high-ranked variables.
Conclusions: Risk models based on machine learning algorithms may be superior to traditional models, which were primarily based on logistic algorithms to predict the occurrence of POAF after valve surgery. Further prospective multicenter studies are needed to confirm the performance of SVM in predicting POAF.