Machine Learning for Detecting Atrial Fibrillation from ECGs: Systematic Review and Meta-Analysis

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Atrial fibrillation (AF) is a prevalent arrhythmia linked to adverse cardiovascular outcomes.
• Early detection of AF is challenging, necessitating advanced diagnostic tools.
• Machine learning (ML) algorithms are increasingly utilized for AF detection in electrocardiogram (ECG) signals.

Purpose of the Study:

• To systematically evaluate and summarize the diagnostic accuracy of ML algorithms for AF detection.
• To compare the performance of different ML approaches, including traditional ML (TML) and deep learning (DL).

Main Methods:

• A meta-analysis of diagnostic accuracy was performed on studies retrieved from PubMed, Web of Science, Embase, and Google Scholar.
• Sensitivity and specificity were synthesized to assess the overall performance of ML algorithms.
• Included studies focused on ML-based AF detection from ECG signals.

Main Results:

• The meta-analysis included 14 studies, yielding pooled sensitivity and specificity of 97% for ML algorithms in AF detection.
• Deep learning (DL) algorithms achieved a higher sensitivity (98.1%) compared to traditional machine learning (TML) algorithms (91.5%).
• Utilizing multiple or public datasets slightly improved performance over single or proprietary datasets.

Conclusions:

• ML algorithms demonstrate high effectiveness in detecting AF from ECGs.
• Deep learning algorithms, especially convolutional neural networks (CNNs), outperform TML algorithms in AF detection accuracy.
• Integrating ML algorithms into wearable devices can facilitate earlier diagnosis of AF.