MLCGAN: MULTI-LEAD ECG SYNTHESIS WITH MULTI LABEL CONDITIONAL GENERATIVE ADVERSARIAL NETWORK

Electrocardiography(ECG) is a non-invasive tool used to identify the cardiovascular diseases. ECG classification studies have been concerned and made progress well. However, the problems about categories imbalance and absence of labelled clinic data are still dramatically hindered research development. Recently, generative models have been verified as a possible way to handle the data scarcity issues. For ECG synthesis, to the best of our knowledge, as the time sequences and multiple labels constraints no model can generate ECG corresponding to clinic data. In this paper, we present a novel multi-label conditional generative adversarial network, named MLCGAN. To synthesise reasonable long-term multi-lead data, multi-label mixing module is devised to combine with our improved WaveGAN. Moreover, the sampling strategy based on multilabels distribution is proposed. Comprehensive experiments demonstrate that MLCGAN can generate ECG data satisfied the clinic diagnose requirement and improve the performance of RestNet based ECG classifier.