A Hybrid Hierarchical Method for Electrocardiogram (ECG) Heartbeat Classification

What is it about?

This paper proposes an automatic reliable two-stage hybrid hierarchical method for ECG heartbeat classification. The heartbeats are segmented dynamically to avoid the consequences of the heart rate variability. Discrete Wavelet Transform (DWT) is utilized to extract morphological features that describe the segmented heartbeat. The extracted features are then reduced by using Principle Component Analysis (PCA). Subsequently, the resulted features along with four RR features are fed into a Support Vector Machine (SVM) to classify five categories (first stage).Thereafter, the heartbeats are further classified to one of the classes belonging to the assigned category (second stage). Two different strategies for classification have been investigated: One versus All and One versus One. The proposed method has been applied on data from lead 1 and lead 2. A new fusion step is introduced, where a stacked generalization algorithm is applied and different types of classifiers have been examined. Experiments have been carried out using a MIT_BIH database. The best overall and average accuracies obtained by the first stage are 98.40% and 97.50% respectively. For the second stage, 94.94% and 93.19% are the best overall and average accuracies obtained respectively. The best results are achieved using SVM with one versus one classification strategy for both stages and decision trees classifier for the fusion step.

Featured Image

Why is it important?

An automatic hybrid hierarchy method of two stages has been developed. The first stage classifies a heartbeat to one of five main categories. Thereafter, the heartbeat moves to the second stage to know which class in this category it belongs. A Discrete Wavelet Transform is applied on dynamically segmented heartbeats. The resulted wavelet coefficient structures have been projected to subspace of 18 PCA components. After concatenating the four RR features, the final feature vectors are fed into the SVM neural network for classification. Two strategies for classification, One versus One and One versus All have been examined. The proposed method has been applied to ECG data from both lead 1 and lead 2. A new fusion step based on stacked generalization in the first stage is suggested to get the advantage of considering the information derived from both leads 1 and 2. The new step has been examined and compared with the rejection method utilized by the existing studies. All experiments have been conducted using a MIT-BIH database. The results reveal the significance of RR interval features for characterizing some categories. In addition, simplifying the classification process to a One versus One strategy achieved the best accuracies for both lead 1 and lead 2. Furthermore, not only the overall accuracy has been considered by this study (as all studies in the literature) but also the average accuracy. Moreover, the proposed fusion step based on the stacked generalization method with decision trees, overwhelms the rejection method and provides best benefit of both leads information regarding the first stage. In addition, to the best knowledge of the authors, none of the existing studies provides information on both category and class in the hierarchical approach. If only the category is needed the first stage will be enough. On the other hand, if more specific information on the nature of the heart beat is needed, it will be the role of the second stage. Finally, the proposed method has achieved in the first stage a best result of 98.4 % overall accuracy and a best average accuracy of 97.5%. Regarding the second stage, 94.94% overall accuracy and 93.19% average accuracy have been achieved using One versus One classification strategy. The average accuracy achieved for each class provides evidence on the robustness of the proposed method to classify all classes (accuracy of most classes is above 90%) and not only some classes at the expense of the others. In the future, we are looking forward to having enough data to gain the advantage of applying stack generalization in both classification stages.

Perspectives