Feature subset selection and classification of intracardiac electrograms during atrial fibrillation

Several approaches have been adopted for the identification of arrhythmogenic sources maintaining atrial fibrillation (AF). In this paper, we propose a classifier that discriminates between four classes of atrial electrogram (EGM). We delved into the relation between levels of fractionation in EGM signals and the fibrillation substrates in simulated episodes of chronic AF. Several feature extraction methods were used to calculate 92 features from 429 real EGM records acquired during radiofrequency ablation of chronic AF. We selected the optimal subset of features by using a genetic algorithm, followed by K-nearest neighbors (K-NN) classification into four levels of fractionation. Sensitivity of 0.90 and specificity of 0.97 were achieved. Subsequently, the results of the classification were extrapolated to signals of a 3D human atria model and a 2D model of atrial tissue. The 3D model simulated an episode of AF maintained by a rotor in the posterior wall of the left atrium and the 2D model simulated an AF episode with one stable rotor. We used the K-NN classifier trained on a given set of real EGM signals to detect a specific class of signals presenting the highest level of fractionation located near the rotor's vortex. This method needs to be tested on real clinical data to provide evidence that it can support ablation therapy procedures.