Classification of intracavitary electrograms in atrial fibrillation using information and complexity measures

•Our selected features in AEGMs characterization are robust to the used classifier.•Classification performance in each proposed task overcome previous results.•The estimation of selected features is not time consuming. Classification of complex fractionated atrial electrograms is crucial for the study of atrial fibrillation and the development of treatment strategies, because these electrophysiological phenomena represent a common substrate for radiofrequency ablation in treatment of this arrythmia. The objective of this work is the characterization of short term atrial electrograms using nonlinear dynamics measures, helping in the automatic classification of electrograms. The dataset consists of 113 atrial electrograms recordings from left-atrial endocardial mapping. These signals were classified by three expert electrophysiologists into four classes, from C0 (non fractionated) to C3 (high degree of fractionation). The calculated features were Approximate entropy, Dispersion entropy, Fuzzy entropy, Permutation entropy, Tsallis entropy, Shannon entropy, Renyi entropy, and Lempel-Ziv complexity. Features were selected for classification using Neighborhood Component Analysis. Different classifiers were tested using selected features, and the one with maximum sensitivity and specificity in each task was reported. We obtained a classification performance that overcome previous works on this database and are comparable to the results of studies performed over bigger datasets. Separation between C3 signals from (C0, C1, C2) signals was performed with 99.98% sensitivity and 96.61% specificity. Non-fractionated signals (C0 + C1) were separated from fractionated signals (C2 + C3) with 96.72% sensitivity and 94.51% specificity. Moreover, the estimation times of the selected features are low enough to consider the online application of this scheme. Classification performance obtained using information and complexity measures shown better results than previous works over this dataset, encouraging the application of these features to characterize atrial electrograms.