Detection of paroxysmal atrial fibrillation from non-episodic ECG data using multi-dimensional feature representation and learning

Paroxysmal atrial fibrillation (PAF) detection based on routine electrocardiogram (ECG) signals is still one of the most challenging problems in research community, since non-episodic ECG fails to diagnose PAF. In this paper, a new PAF detection algorithm based on non-episodic ECG data using multi-dimensional feature representation and learning is proposed. Mean amplitude spectrum (MAS), mel-frequency cepstrum coefficients (MFCC), wavelet packet features (WPFS) and statistical wavelet packet features (SFS) are derived and represented as multi-dimensional image features. These four kinds of cardiac time frequency representations reflect dynamical characteristics during heart beating from four different aspects, which has shown to be more sensitive to detect latent PAF even before visible ECG pathologic changes can be observed. The extracted cardiac representations and deep learning technique are then incorporated, and a parallel DenseNet based feature learning scheme are proposed. Deep features underlying these four kinds of cardiac representations are fused on the decision level to improve classification performance. The appearing ECG test signals can be finally classified according to the min rule based decision-making principle. Experimental results show that accuracies of 81.66%, 85.41%, and 91.25% are achieved on PHY-PAF EEG database under two-fold, five-fold, and ten-fold cross-validations, respectively.