George B. Moody PhysioNet Challenge

Logo

Quick links for this year's Challenge:

Please post questions and comments in the forum. However, if your question reveals information about your entry, then please email info at physionetchallenge.org. We may post parts of our reply publicly if we feel that all Challengers should benefit from it. We will not answer emails about the Challenge to any other address.

Session S42.4

A Methodology For Predicting Paroxysmal Atrial Fibrillation Based On ECG Arrhythmia Feature Analysis

W. Zong, R.G. Mark

Harvard-MIT

Cambridge, MA, USA

This work addresses the 2001 CinC Challenge for predicting the onset of paroxysmal atrial fibrillation/flutter (PAF) from the surface electrocardiogram (ECG). We developed a methodology based on ECG arrhythmia feature analysis for this purpose. We found the frequency of atrial premature contractions (APCs)to be a useful feature for predicting PAF. Our initial step was to visually examine all the ECG data in the training set of the PAF ECG database to get a better understanding of the ECG characteristics in the subjects during PAF, prior to PAF, and without PAF. Then, we employed a previously developed automated arrhythmia detection algorithm which identifies beat types (Normal, APC, VPC, Unknown, etc.), as well as rhythm types (SVT, VT, Sinus Tach, etc.). After investigating the arrhythmia patterns associated with and without PAF events, we found that APC frequency seemed be a convenient and useful predictive feature. Our algorithm detects and counts the number of APCs that occurred in each 30 minute ECG recording. For Event 1 (Screening), we utilized the total number of APCs from both recordings as the statistic for prediction. In order to establish an optimal threshold for this statistic that distinguishes between Normal and PAF subjects, we analyzed the receiver operating curve (ROC) characteristics based on the training data. The resulting threshold was used to screen the test data. Subjects in the test set were determined to have PAF when their total number of APCs was larger than the threshold. For Event 2 (Prediction), we again utilized the number of APCs detected in each 30 minute recording. We predicted that for each subject, the recording with a larger number of APCs would be followed by PAF. With the method outlined above, we achieved a score of 35 correct out of 50 for Event 1 and 44 correct out of 50 for Event 2. These encouraging results indicate that APCs are indeed of great value for predicting imminent PAF for at least the test set provided by the CinC Challenge. We believe that we can improve upon these results by incorporating into our method measures of other beat and rhythm features, RR interval variability, and P wave characteristics.


Supported by the National Institute of Biomedical Imaging and Bioengineering (NIBIB) under NIH grant number R01EB030362.

Back