Opportunities to Reduce the Risk of Cardiovascular Death by Improving Machine Learning Methods

Abstract

Improving algorithms for automatic recognition of electrocardiograms requires increasing of training dataset, which is not always possible due to the rarity of certain cardiac diseases or ethical issues. It is possible to improve the algorithms for generating synthetic electrocardiograms using mechanistic models and generative-descriptive neural networks (GANs). At the same time, when evaluating the effectiveness of the proposed solutions, various authors offer different quality assessment metrics from subjective expert assessment to the squared mean error. We compare two approaches to generating synthetic electrocardiograms: pseudo-ECG generation using a one-dimensional cardiomyocyte model and GAN based on long-term short-term memory in terms of machine learning metrics: accuracy, recall, f1-score. We solve the problem of binary classification with bagging method, class 0 — normal sinus rhythm, class 1 — atrial fibrillation. We found that classifier trained on synthetic ECGs generated using GAN is slightly more effective compared to pseudo-ECGs generated using mechanistic models ones. At the same time, it turned out that GANs are not stable enough. The generation of synthetic electrocardiograms using both mechanistic models and GAN can be used to enrich the training set in case of recognition of rare cardiac diseases.