Interpretable Automated Arrhythmia Detection: An Assistive Framework for Clinicians

Cardiac arrhythmia due to changes in the rate or the site of origin of cardiac electrical impulses, manifests in altered electrocardiogram (ECG) characteristics. Automated arrhythmia detection in ECG often overlooks subtle pathological changes due to limitations in traditional static features. While deep learning features excel in classification, their interpretability remains challenging. For accurate automated arrhythmia detection, while ensuring clinical interpretability, this study proposes a unique feature selection algorithm to identify potent ECG biomarkers. Binary (One-vs-One) random forest (RF) classifiers were developed to discern key features for class distinction. The proposed feature selection technique outperformed popular methods, including the LASSO, F-test, and mRMR algorithms. The identified significant features enabled the multi-class inter-patient RF classifier to achieve an average F1 score of 0.98 on the test dataset, surpassing state-of-the-art results, including those obtained from a Residual Network trained on the same dataset. The proposed approach has the potential to assist clinicians with an adept and comprehensible tool for arrhythmia detection.