George B. Moody PhysioNet Challenge


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 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 S31.2

Fibrillatory Wave Analysis of the Surface ECG to Predict Termination of Atrial Fibrillation: The 2004 Computers in Cardiology/PhysioNet Challenge

S Petrutiu, AV Sahakian, J Ng, S Swiryn

Northwestern University

Evanston, IL, USA

Paroxysmal atrial fibrillation (AF) is self-terminating by definition, but the mechanism by which this occurs is not well understood. It has been reported that fibrillatory activity slows just prior to termination, and we have recently demonstrated that patients with paroxysmal AF have lower frequency fibrillatory waves than those with persistent or permanent AF.

Holter recordings from the Challenge database were used to develop and test algorithms for distinguishing between AF segments that are non-terminating (N), terminating within a second (T), and terminating within a minute (S). The database consisted of a training set containing 30 recordings (10 of each), and two test sets, one with 30 N or T and one with 20 S or T recordings. Following QRST cancellation, the power spectrum of each remainder ECG was calculated. We used the overall peak frequency in the 4 to 9 Hz band and the short-term peak frequency and peak power of the last 2 seconds to distinguish between N and T and between S and T recordings.

From the training set, the peak frequency ranges (mean ± SD) were 4.8-6.0 (5.3±0.4) Hz for T, 4.7-6.4 (5.2±0.6) Hz for S, 4.8-7.3 (6.5±0.8) Hz for N. Terminating recordings had a lower mean peak frequency than non-terminating recordings (p = 0.0002 for N and T, p = 0.0004 for N and S). Comparison of the penultimate and ultimate second showed a decrease in peak frequency in 8 of 10 T recordings with mean decrease of 0.7±0.6 Hz and an increase in 2 (0.3 Hz, 1.1 Hz). There was also a decrease in peak power in 8 of 10 T recordings with mean decrease of 1.96±2.26 mV2, and an increase in 2 (2.32 mV2, 3.96 mV2). The last second of the recording had a lower peak frequency for T when compared to S of the same patient in 9 of 10 patients with a mean difference of 0.9±0.6 Hz. Based on these criteria, between N and T we correctly classified 20 of 20 from the training set and 29 of 30 from the test set. Between S and T we correctly classified 20 of 20 from the training set and 20 of 20 from the test set.

There are frequency characteristics that can be used to discriminate terminating from non-terminating AF. Whether this reflects fewer circulating wavelets, slower firing of pulmonary vein foci, or other factors is unknown.

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