Robust and Automated Detection and Classification of Premature Ventricular Contraction ECG Events Using Wavelet Transform and

Feed-forward Neural Networks

Abstract

Automatic ECG detection and classification is essential to the timely diagnosis of potentially lethal heart conditions in clinical settings. Specifically, accurate detection of premature ventricular contraction (PVC) beats in a patient with heart disease is necessary in order to prepare for the possible onset of ventricular tachycardia, ventricular fibrillation, or other deadly arrhythmias. Although many groups have developed algorithms for detecting PVC beats, none have shown reliability and high accuracy over a large set of data. In fact, many of the available methods have only been proved accurate over the same set of data used for training. We believe that the reason for this is the inherent and significant variation in morphology of the ECG waveform among different patients. We present a method which combines the wavelet decomposed ECG waveform and relevant timing information from the original ECG waveform in the feature set of our feed-forward MLP neural network classifier. The classifier achieves an accuracy of 95.59% over 40 files from the MIT / BIH database in differentiating normal, PVC, and other beats. The accuracy over 20 of these files separate from the neural network training set was 97.3%, demonstrating the robustness of our algorithm to unknown patients.