Early detection of paroxysmal atrial fibrillation from non-episodic ECG data using cardiac dynamics features and different classification models.

Objective.Intelligent computer-aided diagnosis techniques enable inspection of invisible electrocardiogram (ECG) pathological changes for early detection of latent heart diseases. This study concentrates on latent pathological changes within non-episodic ECG data, describes a cardiac dynamics based methodology for the detection of paroxysmal atrial fibrillation (PAF).Approach.Three-dimensional dominated components of routine 12-lead ECG signals are extracted without complex signal segmentation operations. Cardiac dynamics features are captured using deterministic learning algorithm and represented as the three-dimensional graphic. This kind of nonlinear dynamics representation is shown to have high discriminative power for PAF detection even before pathologic changes can be observed visibly in ECG signals. Nonlinear dynamics measures are extracted and finally fed into different machine learning methods for the PAF detection task. Suspected PAF patients undergoing Holter monitoring are studied. Cardiac dynamics measures are calcuated simultaneously with routine rest ECG examination, in which Holter monitoring results are collected as the gold standard.Main results.The proposed method yielded a sensitivity of 97%, a specificity of 91%, and an overall accuracy of 92%.Significance.Abnormal cardiac dynamics induced by PAF can be detected using cardiac dynamics features and different classification models before obvious pathological changes are present. The proposed method is expected to provide a complementary tool to the commonly used ECG examination for PAF detection, which are crucial for identifying patients at risk of latent PAF.