Session S31.5

Automated Prediction of Spontaneous Termination of Atrial Fibrillation from ECG

D Hayn, K Edegger, D Scherr, P Lercher, B Rotman,
W Klein, G Schreier

ARC Seibersdorf research GmbH
Graz, Austria

An algorithm for automated analysis of ECG signals featuring atrial fibrillation (AF) has been developed. The algorithm was intended to differentiate ECGs with AF that will spontaneously terminate from signals, where it won’t. Development and validation of the algorithm has been done using the AF Termination Challenge Database from PhysioNet.
The algorithm comprised of four major steps. First, QRS complexes were detected using our existing biosignal processing system. In a second step all beats were divided into classes of heartbeats with similar QRS morphologies. For each class of heartbeat of each signal an averaged heartbeat was calculated. The third step was the subtraction of the corresponding averaged heart beats from the original sequence after alignment with each QRS complex in order to remove the ventricular signal parts – i.e. the QRS complex and the T wave. Due to high variations in QRS morphology caused by AF and due to the low sampling rate of the signals, the QRS complexes themselves were thereafter blanked completely by linear interpolation of the signal within a certain time region around each R wave. The forth and final step was filtering the signal and calculating the pseudo-periodogram using a short time Fourier transformation method. The frequency corresponding to the highest spectral peak (major frequency) within the pseudo-periodogram was compared to a predefined threshold value.
The algorithm has been developed using the whole learning-set as well as the test-set-b of the AF Termination Challenge Database. We found that for ECGs with a major frequency lower than 6.3 Hz AF was likely to terminate within up to one minute. Our algorithm was able to separate 100% of the extended test data set into the two groups “will terminate within up to one minute” and “will not terminate within one hour”. Validation was done using the test-set-a of the AF Termination Challenge Database. With our initial vote we reached a score of 90% (27/30) correctly classified signals for the test-set-a, a value which is like to increase by fine-tuning the algorithm.