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

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Atrial fibrillation (AF) detection from electrocardiogram (ECG) signals is challenging, especially with low signal-to-noise ratio (SNR).
• Existing methods may struggle with noisy ECG data, limiting accurate AF diagnosis.
• Advanced signal processing and machine learning are crucial for improving AF detection robustness.

Purpose of the Study:

• To introduce a novel deep learning model, the adaptive bin-stream network (ABNet), for robust AF detection in low SNR ECG signals.
• To enhance AF identification by utilizing frequency decomposition and an adaptive network architecture.
• To evaluate the performance of ABNet on established ECG databases for classifying normal sinus rhythm, AF, other rhythms, and noise.

Main Methods:

• ECG signals were preprocessed and decomposed into 32-frequency channels using 5-level Haar wavelet packet decomposition.
• A bin-stream network was designed with separate waveform and frequency streams to process the decomposed signals.
• An adaptive approach was employed within the network to optimize classification results for AF detection.

Main Results:

• On the PhysioNet/CinC 2017 database, ABNet achieved 93.08% accuracy, 81.84% sensitivity, and 94.00% specificity for classifying four categories (N, AF, O, P).
• For a synthetic database (SPH AF Db + CinC 2011 Db), ABNet demonstrated superior performance with 97.98% accuracy, 98.37% sensitivity, and 98.41% specificity for three categories (N, AF, P).
• The results highlight ABNet's capability in capturing detailed waveform and frequency information for effective AF detection.

Conclusions:

• The adaptive bin-stream network (ABNet) shows significant promise for accurate and robust atrial fibrillation detection, even in challenging low SNR ECG recordings.
• Frequency decomposition combined with an adaptive deep learning architecture is effective for improving the sensitivity and specificity of AF classification.
• ABNet's performance on multiple datasets validates its potential for clinical application in automated ECG analysis for AF screening.