Beatwise ECG Classification for the Detection of Atrial Fibrillation with Deep Learning.

Area of Science:

• Cardiology
• Artificial Intelligence
• Signal Processing

Background:

• Atrial fibrillation (AF) is the most common cardiac arrhythmia, and early intervention is crucial for treatment success.
• Current manual classification of AF types from ECGs is subjective and may not accurately reflect disease severity.
• Estimating AF burden (percentage of AF beats) is vital for assessing disease progression and guiding treatment.

Purpose of the Study:

• To develop and validate a novel machine learning method for beat-wise ECG classification to accurately estimate AF burden.
• To improve the objective assessment of AF severity by analyzing individual heartbeats.
• To differentiate between Normal Sinus Rhythm (SN), AF, Noises (NO), and Others (OT) in ECG recordings.

Main Methods:

• A novel deep learning architecture combining a 1D U-Net and a Recurrent Neural Network (RNN) was developed.
• The 1D U-Net processed ECGs to identify fiducial points and segment heartbeats.
• The RNN enhanced temporal classification of individual heartbeats for accurate AF detection.

Main Results:

• The combined U-Net and RNN model achieved high testing accuracies for the four classes: SN (0.86), AF (0.81), NO (0.79), and OT (0.75).
• The study demonstrated the feasibility of using this deep learning approach for beat-wise ECG classification.
• The proposed method showed superior performance in estimating AF burden compared to existing methods.

Conclusions:

• The novel machine learning network effectively performs beat-wise ECG classification for AF burden determination.
• Combining U-Net and RNN architectures offers a promising approach for objective AF assessment.
• Further validation is recommended to confirm the clinical utility of this deep learning approach for AF management.