Session S42.2

Automated Prediction Of The Onset Of Paroxysmal Atrial Fibrillation From Surface Electrocardiogram Recordings

P. de Chazal, C. Heneghan

University of New South Wales
Sydney, Australia

A technique for predicting the onset of paroxysmal atrial fibrillation/flutter (PAF) through automated assessment of a single channel electrocardiogram (ECG) is presented. Algorithmic development has been carried out using the Physionet PAF database. This database consists of 100 pairs of half-hour two-channel ECG recordings (for a total of 200 half-hour ECG segments). Each pair of recordings is extracted from a two channel (Leads I and II) 24-hour ECG. These 100 pairs are drawn from two subject groups. Subjects in the first group experienced PAF at some point during the 24-hour ECG. For these subjects, one recording ends just before the onset of PAF, and the other recording is distant in time from any PAF. Subjects in the second group do not have PAF. In these subjects, the times of the recordings have been chosen at random. The database is divided into a learning set and a test set of equal size, each containing approximately equal numbers of subjects from the two groups. The classifications of the recordings in the learning set are provided, while those for the test set are withheld for independent validation of classifiers. Unvalidated QRS detection points are provided for all ECG files. A linear discriminant classifier was developed to discriminate half-hour segments prior to PAF from segments not associated with PAF. The classifier uses a ten-element feature vector based solely on R-R intervals. Features include mean R-R interval, R-R interval variance, serial correlation coefficients, and various spectral measures, including a spectral power-law exponent. After training, the classifier was validated on the test set of subjects suffering from PAF. The classifier was able to correctly identify the ECG segment immediately prior to PAF in 41 out of 50 cases in this test set.