PVC Detection Using a Convolutional Autoencoder and Random Forest Classifier

The accurate detection of premature ventricular contractions (PVCs) in patients is an important task in cardiac care for some patients. In some cases, the usefulness to physicians in detecting PVCs stems from their long-term correlations with dangerous heart conditions. In other cases their potential as a precursor to serious cardiac events may make their detection a useful early warning mechanism. In many of these applications, the long-term nature of the monitoring required and the infrequency of PVCs make manual observation for PVCs impractical. Existing methods of automated PVC detection suffer from drawbacks such as the need to use difficult to extract morphological features, domain-specific features, or large numbers of estimated parameters. In particular, systems using large numbers of trained parameters have the potential to require large amounts of training data and computation and may have issues generalizing due to their potential to overfit. To address some of these drawbacks, we developed a novel PVC detection algorithm based around a convolutional autoencoder to address these weaknesses and validated our method using the MIT-BIH arrhythmia database.