Atrial Fibrillation Detection with Low Signal-to-Noise Ratio Data Using Artificial Features and Abstract Features.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence in Healthcare

Background:

• Detecting atrial fibrillation (AF) from short, low signal-to-noise ratio (SNR) single-lead electrocardiogram (ECG) recordings is a significant challenge for wearable heart monitoring systems.
• Accurate identification of AF is critical for timely intervention and management of cardiovascular health.

Purpose of the Study:

• To propose and evaluate a novel AF detection method that effectively identifies AF rhythms from normal, other, and noisy ECG recordings.
• To enhance the performance of AF detection in challenging ECG signals commonly encountered in wearable devices.

Main Methods:

• A feature fusion approach combining 24 handcrafted features (waveform, interval, frequency-domain, nonlinear) with 38 deep features from a 13-layer 1-D Convolutional Neural Network (CNN).
• ECG recordings were standardized to 30-second segments using techniques like copying, cutting, and symmetry to handle variable lengths.
• Random Forest classification was employed on the fused feature matrix for rhythm classification.

Main Results:

• The proposed feature fusion method achieved a mean accuracy (Acc) of 0.857 across four categories: normal (N), AF (A), other rhythm (O), and noisy (∼).
• The F1-score for N, A, and O categories reached 0.837, indicating robust performance in distinguishing these rhythms.
• The method demonstrated satisfactory performance in identifying AF from short, low-SNR single-lead ECG recordings.

Conclusions:

• The feature fusion strategy effectively integrates diverse ECG signal characteristics for improved AF detection.
• The developed method shows promise for reliable AF detection in real-world wearable monitoring applications.
• This approach offers a viable solution for identifying atrial fibrillation in challenging, low-quality ECG data.