A Retrospective Analysis of the Predictive Potential of Pre-operative Data on Post-operative Atrial Fibrillation

1.0 Background Currently, roughly thirty percent of coronary artery bypass graft (CABG) patients develop atrial fibrillation (AF) in the five days following surgery, increasing the risk of stroke, prolonging hospital stay three to four days, and increasing the overall cost of the procedure [1, 2]. According to some sources, over $1 billion is spent annually on this problem in the US alone [2]. Current pharmacologic and nonpharmacologic means of AF prevention are suboptimal, and their side effects, expense, and inconvenience limit their widespread use in all patients [3].

Though many methods have been presented touting high predictive value in terms of sensitivity and specificity for post-operative AF, none are reliable enough for use in a clinical setting. This may be due to the lack of a standardized method for the measurement of certain morphological P wave features in ECG analyses [4]. Furthermore, the medical community has relied on limited variable combination methods for much too long, especially while there are advanced methods of data mining and decision-making to be harnessed. A Bayesian network (BN) is an excellent tool for making decisions based on collected information and is even able to handle missing data points well [5]. By combining more types of data and expert knowledge into a BN with a Bayesian statistical approach, better accuracy is the likely result.

2.0 Objectives

The main objective of this research is to develop a Bayesian network (BN) classifier which can model/predict/assign risk of the occurrence of atrial fibrillation in coronary artery bypass graft patients through the incorporation of different types of patient data. Expert knowledge coming from doctors in the field will be combined using Bayesian statistics with patient data and electrocardiogram (ECG) analysis, improving on the Frequentist methods currently used. We intend to investigate profit or loss due to the inclusion of the following data types:

• Collected Data- Risk factors and other medical indicators recorded in the hospital
• ECG Features- Time, frequency, wavelet, and nonlinear domain features derived from the ECG signal showing AF prediction potential
• Expert Knowledge- Cardiologist modified probability distribution and frequency beliefs of input data based on past experience This study will analyze data from patients who underwent cardiac surgery at Emory University Hospital, Crawford Long Hospital, or the Atlanta Veterans Affairs Medical Center. The collected data will include demographic, pre-operative, operative, and post-operative all taken from the patient's chart. ECG and available telemetry data will also be collected and analyzed for morphological features, which may yield AF predisposition clues.