Heartbeat Dynamics: A Novel Efficient Interpretable Feature for Arrhythmias Classification

Arrhythmias are a significant class of cardiovascular diseases, and timely and accurate detection is critical in preventing high-risk events such as sudden cardiac death. Despite the attention that automatic detection of arrhythmias based on electrocardiogram (ECG) has received, static features used in traditional methods fail to adequately describe the various weak changes of the ECG, resulting in significant but weak pathological information being overlooked. Although deep learning (DL) extracted features demonstrate efficiency in arrhythmia classification, the interpretability of DL methods remains challenging. In this study, we propose a novel and efficient interpretable feature for arrhythmia classification, heartbeat dynamics. It models morphological changes in the heartbeat, and is more sensitive to weak heartbeat variations, and reflects underlying dynamical changes throughout the cardiac cycle at the electrophysiological level. To evaluate its efficiency for arrhythmia classification, we conducted experiments on the MIT-BIH arrhythmia database, using three classical classifiers: k-nearest neighbor (KNN), random forest (RF), and support vector machine (SVM). Our proposed method achieves 99.41% accuracy, 99.10% precision, 98.84% recall, and 0.9897 F1 score with KNN as the classifier, comparable to or better than most DL-based methods. These results indicate that heartbeat dynamics has a strong ability to discriminate between different classes of heartbeats. We anticipate that the heartbeat dynamics feature will enhance the generalization capacity of the arrhythmia detection algorithm when integrated with other static features.