Automated Extraction of Atrial Fibrillation Activity from the Surface ECG Using Independent Component Analysis in the Frequency Domain
Résumé
Background: Atrial fibrillation (AF) is the most common cardiac arrhythmia encountered by physicians. The analysis of AF from the surface electrocardiogram (ECG) requires the suppression of artifacts such as ventricular activity (VA) and noise corrupting the recordings. Independent component analysis (ICA) has recently been shown to tackle successfully the extraction of atrial activity (AA) in AF recordings. Methods: The present contribution puts forward a novel technique simultaneously exploiting the narrowband spectral character of AA and the statistical independence between AA and VA. The technique performs the iterative optimization of a sparsity and non-Gaussianity measure, the kurtosis, in the frequency domain. Results: On 35 ECG segments from 34 AF patients, the proposed one-stage technique obtains practically identical dominant frequency estimates than an existing technique based on two processing stages (Pearson's correlation equal to 0.9998). The proposed method extracts the AA with an average spectral concentration of 56 ± 17%, against 49 ± 17% for the existing method, requiring also fewer computations. Conclusions: The proposed ICA-based technique achieves a more accurate AA waveform estimation and appears more suitable for real-time DSP implementations in clinical environments.