An adaptive bin-stream network based on frequency decomposition for classifying atrial fibrillation with low SNR data

Abstract

To detect atrial fibrillation (AF) in ECG signals with low signal-to-noise ratio (SNR), this study introduces the adaptive bin-stream network (ABNet) based on frequency decomposition. The ABNet offers notable advantages: it exhibits high robustness in identifying AF amidst noisy environments, it decomposes the ECG signals into 32-frequency channel recordings to refine frequency ranges for better identifying AF, and it designs an adaptive bin-stream network to gain the optimal results. The method utilizes a 5-level Haar wavelet packet decomposition to decompose the preprocessed ECG signals into their corresponding 32-frequency channel recordings, and the preprocessing signals and the recordings are fed into waveform stream and frequency stream of the bin-stream network, respectively. Finally, an adaptive approach is employed to obtain the optimal classification results. The ABNet was validated for the PhysioNet/Computing in Cardiology Challenge 2017 database (CinC 2017 Db) to classify 4 categories i.e., normal sinus rhythm (N), AF, other abnormal rhythms (O) and noise (P), and it achieved accuracy (acc) 93.08 %, precision (ppv) 78.68 %, sensitivity (sen) 81.84 %, specificity (spec) 94.00 %, and F₁ 0.8382. In addition, it achieved the acc 97.98, ppv 96.40, sen 98.37 %, spec 98.41 %, and F₁ 0.9595 for a synthetic Db consisting of Shandong provincial hospital AF database (SPH AF Db) and CinC 2011 Db for classifying 3 categories i.e., N, AF and P. These results underscore the effectiveness of the ABNet in capturing detailed information about waveform and different frequencies in ECG signals.