SeqAFNet: A Beat-Wise Sequential Neural Network for Atrial Fibrillation Classification in Adhesive Patch-Type...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Devices

Background:

• Adhesive patch electrocardiographs are widely used for arrhythmia screening due to their convenience.
• Accurate detection of atrial fibrillation (AF) is crucial for preventing related complications.
• Existing methods may require improvement for enhanced AF classification performance.

Purpose of the Study:

• To develop a reliable deep learning model for improved AF classification using patch-type ECG devices.
• To align AF diagnosis with the 2020 European Society of Cardiology (ESC) guidelines.
• To enhance the precision and sensitivity of AF detection in real-world clinical settings.

Main Methods:

• Developed SeqAFNet, a two-stage bidirectional Recurrent Neural Network (RNN) with a many-to-many architecture.
• Utilized RR interval frames for beat-wise classification of ECG signals, processing sequential data.
• Employed an ensembling technique to combine outputs from various temporal sequences for enhanced prediction accuracy.
• Trained and validated the model on clinical trial data from the MEMO Patch and public ECG databases.

Main Results:

• The SeqAFNet model achieved high performance metrics on the patch dataset: accuracy (0.986), precision (0.981), sensitivity (0.979), specificity (0.992), and F1 score (0.98).
• The model demonstrated consistent performance across different public datasets, indicating robustness.
• SeqAFNet proved effective in capturing local features and long-term dependencies relevant to AF detection.

Conclusions:

• SeqAFNet is a robust and accurate tool for classifying atrial fibrillation from adhesive patch-type ECG recordings.
• The developed deep learning approach shows significant potential for real-world applications in arrhythmia screening.
• This method offers a reliable way to improve AF diagnosis in line with current clinical guidelines.