Robust Neural-Network-Based Classification of Premature Ventricular Contractions Using Wavelet Transform and Timing Interval Features

Abstract

Automatic electrocardiogram (ECG) beat classification is essential to timely diagnosis of dangerous heart conditions. Specifically, accurate detection of premature ventricular contractions (PVCs) is imperative to prepare for the possible onset of life-threatening arrhythmias. Although many groups have developed highly accurate algorithms for detecting PVC beats, results have generally been limited to relatively small data sets. Additionally, many of the highest classification accuracies (> 90%) have been achieved in experiments where training and testing sets overlapped significantly. Expanding the overall data set greatly reduces overall accuracy due to significant variation in ECG morphology among different patients. As a result, we believe that morphological information must be coupled with timing information, which is more constant among patients, in order to achieve high classification accuracy for larger data sets. With this approach, we combined wavelet-transformed ECG waves with timing information as our feature set for classification. We used select waveforms of 18 files of the MIT/BIH arrhythmia database, which provides an annotated collection of normal and arrhythmic beats, for training our neural-network classifier. We then tested the classifier on these 18 training files as well as 22 other files from the database. The accuracy was 95.16% over 93,281 beats from all 40 files, and 96.82% over the 22 files outside the training set in differentiating normal, PVC, and other beats.