Motion-Robust Atrial Fibrillation Detection Based on Remote-Photoplethysmography

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Atrial fibrillation (AF) affects over 43 million worldwide and is a major stroke risk factor.
• Many AF patients are asymptomatic and undiagnosed due to lack of convenient screening tools.
• Remote photoplethysmography (rPPG) offers a potential non-contact solution but is challenged by motion artifacts.

Purpose of the Study:

• To develop and validate a non-contact AF detection method using rPPG.
• To address and mitigate motion disturbances inherent in rPPG signals.
• To create and utilize a comprehensive AF dataset for algorithm verification.

Main Methods:

• Development of NR-Net, ATT-Net, and SQ-Mask modules to handle motion noise and signal quality issues in rPPG.
• Utilizing a Convolutional Neural Network (CNN) for noise elimination.
• Employing channel-wise and temporal attention mechanisms to reduce the impact of poor signal segments.

Main Results:

• The proposed rPPG method achieved high accuracy (95.69%), sensitivity (96.76%), and specificity (94.33%) in distinguishing AF from normal sinus rhythm.
• The algorithm demonstrated superior performance against benchmark methods in AF vs. Non-AF and AF vs. Other arrhythmia scenarios.
• Accuracy improved significantly on slight motion data (up to 95.82%) and full motion data (over 3% increase).

Conclusions:

• The developed non-contact rPPG approach offers a promising and convenient tool for AF screening.
• The novel NR-Net, ATT-Net, and SQ-Mask modules effectively overcome motion-related challenges in rPPG-based AF detection.
• The large, diverse hospital-ward dataset supports the robustness and generalizability of the proposed method.