Clinical Decision Support Algorithm for Prediction of Postoperative Atrial Fibrillation Following Coronary Artery Bypass Grafting

Thumbnail Image
Seaborn, Geoffrey
Computer Science
Introduction: Postoperative atrial fibrillation (POAF) is exhibited by 20-40% of patients following coronary artery bypass grafting (CABG). POAF is associated with increased long-term morbidity and mortality, as well as additional healthcare costs. I aimed to find techniques for predicting which patients are likely to develop POAF, and therefore who may benefit from prophylaxis. Methods: Informed consent was obtained prospectively from patients attending for elective CABG. Patients were placed in the POAF group if atrial fibrillation (AF) was sustained for at least 30 seconds prior to discharge, and were placed in the ‘no AF’ (NOAF) group otherwise. I evaluated the performance of classifiers including binary logistic regression (BLR), k-nearest neighbors (k-NN), support vector machine (SVM), artificial neural network (ANN), decision tree, and a committee of classifiers in leave-one-out cross validation. Accuracy was calculated in terms of sensitivity (Se), specificity (Sp), positive predictive value (PPV), negative predictive value (NPV), and C-statistic. Results: Consent was obtained from 200 patients. I excluded 21 patients due to postoperative administration of amiodarone, 5 due to perioperative AF ablation, and 1 due to both. Exclusions were also made for 8 patients with a history of AF, 2 patients with cardiac implantable electronic devices (CIED), and 3 patients with no CABG (valve replacement only). POAF was exhibited by 54 (34%) of patients. Factors significantly associated (P<0.05) with POAF were longer postoperative hospital stay, advanced age, larger left atrial (LA) volume, presence of valvular disease, and lower white blood cell count (WCC). Using BLR for dimensionality reduction, I created a feature vector consisting of age, presence of congestive heart failure (CHF) (P=0.06), valvular disease, WCC, and aortic valve replacement (AVR). I performed leave-one-out cross validation. In unlabeled testing data, I obtained Se=70%, Sp=56%, PPV=89%, NPV=26%, and C=58% using a committee (BLR, k-NN, and ANN). Conclusion: My results suggest that prediction of patients likely to develop POAF is possible using established machine learning techniques, thus allowing targeting of appropriate contemporary preventative techniques in a population at risk for POAF. Studies appear warranted to discover new predictive indices that may be added to this algorithm during continued enrolment and validation.
External DOI