Deep Learning Approaches to Detect Atrial Fibrillation Using Photoplethysmographic Signals: Algorithms Development...

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Wearable devices utilize photoplethysmography (PPG) for atrial fibrillation (AF) screening.
• Existing algorithms struggle with paroxysmal AF due to premature atrial complexes (PACs).
• Rule-based and handcrafted machine learning methods lack diagnostic accuracy and reliability for AF detection with PACs.

Purpose of the Study:

• To develop deep learning (DL) classifiers for detecting AF from sinus rhythm (SR) using PPG data.
• To evaluate DL performance in the presence of PACs after cardioversion.
• To compare DL classifiers against traditional AF detection methods.

Main Methods:

• Utilized PPG and ECG data from 75 patients post-cardioversion.
• Developed 1-dimensional convolutional neural network (1D-CNN) and recurrent neural network (RNN) DL architectures.
• Introduced a PAC indicator and a confidence level (CL) metric for diagnostic reliability.

Main Results:

• DL classifiers achieved high diagnostic accuracy for AF vs. SR (e.g., 99.32% for 1D-CNN).
• Area under the curve values for DL classifiers significantly exceeded traditional methods (0.998 for 1D-CNN, 0.996 for RNN).
• DL performance improved with increased training data, especially with high PAC burden, and demonstrated higher CL for true diagnoses.

Conclusions:

• Deep learning classifiers demonstrate superior accuracy in detecting AF via PPG signals, even with frequent PACs.
• The confidence level (CL) metric enhances the reliability of AF/SR diagnoses from wearable devices.
• Validated DL classifiers show promise as reliable tools for widespread AF screening using wearable PPG sensors.