Automated Atrial Fibrillation Detection Based on Feature Fusion Using Discriminant Canonical Correlation Analysis

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Atrial fibrillation (AF) is a prevalent cardiovascular disease associated with high morbidity and mortality.
• Early detection and treatment of AF are crucial for improving patient outcomes.
• Existing methods for AF detection from single-lead ECGs face challenges with accuracy and data redundancy.

Purpose of the Study:

• To develop and validate a multiple feature fusion method for accurate AF screening from short, single-lead ECG recordings.
• To address computational and information redundancy issues in traditional feature fusion techniques.
• To enhance the diagnostic performance by integrating expert-derived and deep learning features.

Main Methods:

• Proposed a discriminant canonical correlation analysis (DCCA) based feature fusion technique.
• Integrated traditional ECG features (expert knowledge) with deep learning features (ResNet, GRU).
• Evaluated the method on the Cardiology Challenge 2017 dataset.

Main Results:

• The proposed DCCA feature fusion achieved an F1 score of 88%.
• Achieved high diagnostic performance with 91.7% accuracy, 90.4% sensitivity, and 93.2% specificity.
• Demonstrated superior performance compared to single feature-based approaches.

Conclusions:

• The DCCA feature fusion method effectively screens atrial fibrillation from single-lead ECGs.
• Integrating diverse feature types significantly improves AF detection accuracy.
• This approach offers a promising tool for early and accurate diagnosis of AF.