Introduction Detection among supraventricular tachycardia (SVT), ventricular tachycardia (VT) and ventricular fibrillation (VF) is important because they are treated differently. In specific, VT can quickly deteriorate to VF or death whereas SVT is not as lethal as VT and VF. However, most of detection algorithms divided into only VT and VF or VT and SVT. Therefore, searching for a reliable detection method of these arrhythmias can help cardiovascular diagnosis system.
Method In this study, we classified 10 seconds segments of ECG into four classes (SVT, VT, VF, normal sinus rhythm, and other rhythms) from PhysioNet’s MIT-BIH datasets. We compared decision tree (DT), support vector machine (SVM), and convolutional neural network (CNN). We used the DT and the SVM by with and without the ensemble method. For CNN, we performed augmentation to solve high variance problem. All of the methods applied preprocessing step and 10 seconds window size that was based on the annotated episode of the database. We extracted a total of 10641 samples into a training set and test set based on the proportion of 7:3, among which 7449 training and 3192 test data were used to build the model. Each method performed 10-fold cross-validation to prevent overfitting.
Results An overall accuracy (OA) of The DT with ensemble obtained 98.2% 78.2%, whereas The DT without ensemble achieved 98.1% and 78.1% on the training and the test set, respectively. The SVM with ensemble obtained an OA of 98.6% and 78.0%, whereas the SVM without ensemble obtained an OA of 93.4 and 77.1% on the training and the test set, respectively. On the other hand, CNN with augmentation obtained an OA 98.1% on the training set, and 94.6% on the test set, respectively.