04 Apr Novel interpretable Feature set extraction and classification for accurate atrial fibrillation detection from ECGs
The objective of this study is to introduce a novel approach for detecting atrial fibrillation (AFib) by analyzing Lead II electrocardiograms (ECGs) using a distinct set of features. To achieve this goal, we employed advanced signal processing techniques, including proper orthogonal decomposition, continuous wavelet transforms, discrete cosine transforms, and standard cross-correlation, to extract 48 features from ECG signals. Our method is designed to more effectively capture AFib-specific characteristics, such as beat-to-beat variability and fibrillatory waves, compared to conventional metrics. Additionally, the extracted features were developed to be physiologically interpretable. We then utilized an XGBoost-based classifier, training and evaluating it on the publicly available ‘Training’ dataset from the 2017 PhysioNet Challenge, which classifies ECGs into four categories: ‘Normal,’ ‘AFib,’ ‘Other,’ and ‘Noisy.’
Our model achieved an accuracy of 96% with an F1-score of 0.83 for AFib detection and 80% accuracy with an F1-score of 0.85 for normal ECG classification. We further compared our approach with the top-performing classifiers from the 2017 PhysioNet Challenge—ENCASE, Random Forest, and Cascaded Binary—which reported F1-scores of 0.80, 0.83, and 0.82 for AFib detection, respectively, as noted by Clifford et al. (2017).
Despite utilizing significantly fewer features than the leading algorithms from the competition (48 vs. 150-622), our model achieved a comparable F1-score of 0.83 for AFib detection. The physiologically interpretable features designed specifically for AFib detection enhance our method’s adaptability in clinical settings, enabling real-time arrhythmia detection, even for short ECG recordings (<10 heartbeats).
No Comments