Phase information of time-frequency transforms as a key feature for classification of atrial fibrillation episodes

Patients suffering from atrial fibrillation can be classified into different subtypes, according to the temporal pattern of the arrhythmia and its recurrence. Nowadays, clinicians cannot differentiate a priori between the different subtypes, and patient classification is done afterwards, when its clinical course is available. In this paper we present a comparison of classification performances when differentiating paroxysmal and persistent atrial fibrillation episodes by means of support vector machines. We analyze short surface electrocardiogram recordings by extracting modulus and phase features from several time-frequency transforms: short-time Fourier transform, Wigner-Ville, Choi-Williams, Stockwell transform, and general Fourier-family transform. Overall, accuracy higher than 81% is obtained when classifying phase information features of real test ECGs from a heterogeneous cohort of patients (in terms of progression of the arrhythmia and antiarrhythmic treatment) recorded in a tertiary center. Therefore, phase features can facilitate the clinicians' choice of the most appropriate treatment for each patient by means of a non-invasive technique (the surface ECG).