Detection of Paroxysmal Atrial Fibrillation from Dynamic ECG Recordings Based on a Deep Learning Model.

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Atrial fibrillation (AF) is a common clinical arrhythmia, with risk increasing with age.
• AF often co-occurs with comorbidities like coronary artery disease (CAD) and heart failure (HF).
• Accurate AF detection is challenging due to its intermittent and unpredictable nature.

Purpose of the Study:

• To develop and validate a deep learning model for precise atrial fibrillation detection.
• To enable the identification of AF onset and offset from electrocardiogram (ECG) data.
• To improve diagnostic accuracy and reduce false alarms in AF monitoring.

Main Methods:

• A deep learning model incorporating residual blocks and a Transformer encoder was developed.
• The model was trained on dynamic ECG data from the CPSC2021 Challenge.
• The method did not distinguish between atrial fibrillation (AF) and atrial flutter (AFL).

Main Results:

• The model achieved high accuracy (98.67%), sensitivity (87.69%), and specificity (98.56%) in AF rhythm testing.
• Onset and offset detection showed high sensitivity (95.90% and 87.70%, respectively).
• A low false positive rate (FPR) of 0.46% minimized false alarms. Noise stress tests confirmed robustness.

Conclusions:

• The deep learning model effectively discriminates AF from normal heart rhythms.
• The model accurately detects the onset and offset of AF episodes.
• Feature visualization confirmed the model's interpretability, focusing on key ECG waveform characteristics.