Feature fusion for imbalanced ECG data analysis

World Health Organization (WHO) indicates that cardiovascular disease remains challenging in diagnosis and treatment. The electrocardiogram (ECG) is a very important diagnostic assistant for cardiac diseases. Traditionally, most of the ECG analysis methods are evaluated by their intra-patient performance, which however may not suitable for inter-patient cases. Here, we propose a complete classification system with excellent generalization ability. We first extract the 2D-convolutional and PQRST features of a single heartbeat after preliminary processing. We then balance the data with the Random Over Sampler algorithm after comparing several imbalanced algorithms. Finally, we use a Random Forest (RF) classifier to classify the data according to the Association for the Advancement of Medical Instrumentation (AAMI) standards (1988). Results show that RecallM (MR), PrecisionM (MP) and FscoreM (MF) of our proposal are all above 99%. In order to evaluate the performance of different methods, we designed inter-patient and intra-patient experiments separately. To further demonstrate the robust and adaptability of our model, we then transferred it to another data set and performed the experiment. In our experiments, the values of macro- and micro-metrics are up to 99%. All of the results are averages of five experiments, and the Average Accuracy (AA) of experiments applied here are above 99%, which illustrates that our proposal is a promising alternative and superior to most of the state-of-the-art methods.