Introduction: Premature ventricular contraction (PVC) is a common cardiac arrhythmia triggered by ectopic heartbeats. This is harmless in most cases but can lead to life-threatening heart failure under particular circumstances. Automated PVC detection using electrocardiogram (ECG) data is a task of considerable importance to alleviate the heavy workloads of experts in the manual analysis of long-term ECGs. The aim of this research was to classify QRS complexes using deep neural networks in order to identify PVCs.
Method: Based on extracts from a database which is publicly available: the MIT-BIH arrhythmia database (MIT-BIH-AR). The initial dataset consisted of raw data in which each column of the dataset corresponded to one sample of the ECG signal. Samples were extracted from the database (30 before and 30 after each R peak annotation). Hence each complex was 169 milliseconds. Subsequently, an RGB images was created for each complex. A Long Short-Term Memory (LSTM) recurrent neural network was trained using the raw sample data and four convolutional neural networks (AlexNet, GoogleNet, Inception V3 and ResNet-50) were trained using the RGB images for classifying PVCs.
Results: During two different experiments (one with the entire dataset and the other with a balanced dataset) the final results showed high efficiency and reliability in the final diagnoses. Effectiveness was measured using precision, sensitivity, specificity, positive predictive value, F1 score, and AUC. According to the accuracy ResNet-50 was the best classifier in the first experiment achieving 99.8% accuracy (F1=99.2%) and Inception V3 was the best classifier in the second experiment (acc=99.7%, F1=98.8%). Interesting information was extrapolated from a review of the confusion matrix to conduct a "failure analysis" to explain where and why the classifiers made incorrect classifications.