Novel interpretable Feature set extraction and classification for accurate atrial fibrillation detection from ECGs

Area of Science:

• Cardiology
• Biomedical Engineering
• Signal Processing

Background:

• Atrial fibrillation (AFib) is a common arrhythmia requiring accurate detection.
• Traditional methods for AFib detection using electrocardiograms (ECGs) may not fully capture complex signal variations.
• Developing efficient and interpretable AFib detection algorithms is crucial for clinical applications.

Purpose of the Study:

• To introduce a novel method for detecting AFib from Lead II ECGs.
• To develop a unique set of physiologically interpretable features for AFib detection.
• To evaluate the performance of the proposed method against established algorithms.

Main Methods:

• Extracted 48 unique ECG features using signal processing techniques like proper orthogonal decomposition, continuous wavelet transforms, discrete cosine transform, and cross-correlation.
• Designed features to capture beat-to-beat variability and fibrillatory waves characteristic of AFib.
• Utilized an XGBoost classifier trained and validated on the 2017 PhysioNet Challenge 'Training' dataset.

Main Results:

• Achieved 96% accuracy and an F1-score of 0.83 for AFib detection.
• Attained 80% accuracy and an F1-score of 0.85 for normal ECG classification.
• The proposed method's F1-score for AFib detection (0.83) was comparable to top-performing models in the 2017 PhysioNet Challenge.

Conclusions:

• The novel method effectively detects AFib using a significantly smaller feature set compared to state-of-the-art algorithms.
• The interpretable features enable adaptability for real-time arrhythmia detection in clinical settings.
• The method demonstrates efficacy even with short ECG recordings (less than 10 heartbeats).