Evaluating the Quality of CycleGAN Generated ECG Data for Myocardial Infarction Classification

The demand for extensive annotated datasets in ECG
interpretation has led to the development of synthetic
datasets using generative neural networks. Our study is
aimed at assessing the quality of synthetic ECGs generated via a CycleGAN network by means of visual inspection (confidence bands and UMAP 2D plots), GAN-specific
evaluation methods (GAN-train and GAN-test scoring),
and statistical tests comparing ST segment amplitudes
(modified Hotelling T-squared test). To this goal, we utilized a selection of 12-lead ECGs from the PTBXL dataset
(available on Physionet) falling under three conditions:
normal sinus rhythm, anteroseptal myocardial infarction
and inferior myocardial infarction. Through the CycleGAN network we generated synthetic ECGs and compared
them with the original ones. The qualitative analysis,
by means of plots, showed that there was a difference in
the distributions of real and synthetic data. The GANtrain/test method provided results confirming this conclusion. Lastly, the ST-segments analysis showed distributions
which were dissimilar among all the conditions. In conclusion, our work demonstrated that generative networks
developed in the context of image processing cannot be
simply adapted to augment ECG datasets, and that proper
care should be enforced to verify the quality of the generated signals, before utilising such data in applications