Session S91.5

Computers in Cardiology/Physionet Challenge 2004: AF Classification Based on Clinical Features

M Lemay, Z Ihara, JM Vesin, L Kappenberger

Swiss Federal Institute of Technology
Lausanne, Switzerland

The challenge 2004 involved the classification of ECG signals of AF patients into three categories: types N, S, and T corresponding to AF episodes terminating never, soon or immediately, respectively. In our study, 7 different features were used, extracted by the experienced clinician among the authors (LK) from the supplied training set. The first feature is the main orientation of the F-waves in the atrial activity (AA). When AF terminates soon, P-waves will reappear and F-waves are more organized, translating into a stable peak direction of F-waves. The second feature relates to the RR intervals. Two interesting phenomena were observed. First, when ventricles beat at high rate, the AF seem to terminates soon, and second, when atria rhythm tries to return to normal, the ventricular rhythm also changes. The third feature relates to the intervals between F-waves. When AF front waves return to sinus rhythm, the AA intervals decrease. The fourth feature is the amplitude of AA. When the AF front waves are more organized, their amplitude perceived in the ECGs are higher. The fifth feature used is the periodicity in the amplitude of AA. When AF stops, slow variations appear in the AA amplitudes. The sixth feature exploits the similarities among AAs in the two leads provided. During complete AF, the wave fronts are not similar, but when the process slowly returns to sinus rhythm, their shapes become similar then viewed from different point of observation (leads). The final feature is the power in the high frequency range of the AA segments. The F-waves close to a P-wave are steeper than the F-waves in complete AF. Algorithms were developed to extract these features from the provided ECG data. We used a support vector machine (SVM) with radial basis and linear function kernel to classify these criteria. The SVM was trained by using the data in the supplied training set, using a boundary decision rule while keeping the number of subset as low as possible to avoid over fitting. When applied to the training set, the application of the features used and the derived decision rules resulted in 100% correct classification, demonstrating that the features as such have potential. When applied to the test set we obtained the following results: set A = [NNTNNTNNNNNNNTNTNTTNTNTTNNTNTN] and set B = [STTTSSSSTSSSTTSSSSTS].