ECG based prediction of atrial fibrillation using support vector classifier

Abstract

In patients undergoing Coronary Artery Bypass Grafting (CABG) surgery postoperative atrial fibrillation (AF) occurs with prevalence of up to 40%. The highest incidence is between the second and third day after the operation. Following cardiac surgery AF causes various complications, hemodynamic instability, and can cause heart attack, cerebral and other thromboemolisms. AF increases morbidity, duration and expense of medical treatment. This study aims to identify patients at high risk of postoperative AF. An early prediction of AF would provide a timely prophylactic treatment and would reduce incidence of arrhythmia. Patients at low risk of postoperative AF could be excluded from the side effects of anti-arrhythmic drugs. The investigation included 50 patients in whom lead II electrocardiograms were continuously recorded for 48 hours following CABG. Univariate statistical analysis was used in the search of signal features that might predict AF. The most promising identified features were: P wave duration, RR interval duration and PQ segment level. On the basis of these a nonlinear multivariate prediction model was made deploying a Support Vector Machine (SVM) classifier. The prediction accuracy was found uprising over the time. At 48 hours following CABG; the measured best average sensitivity was 95.9% and specificity 93.4%. The positive and negative predictive accuracy were 88.9% and 98.8%, respectively and the overall accuracy was 94.6%. In regard to the prediction accuracy, the risk assessment and prediction of postoperative AF are optimal to be done in the period between 24 and 48 hours following CABG.