Generating electrocardiogram signals by deep learning

Abstract

Given the importance of a diverse and vast amount of realistic and labeled electrocardiogram (ECG) signals in improving the performance of biomedical signal processing algorithms, and the situation of severe lack of the signals, three generative models based on deep learning are introduced for the generation of ECG signals: The WaveNet-based model, the SpectroGAN model, and the WaveletGAN model. The WaveNet-based model adopts μ-law companding transformation as a preprocessing method and then is followed by a sequence of convolutional layers with dilation; SpectroGAN and WaveletGAN use short-term Fourier transform (STFT) and stationary wavelet transform (SWT) respectively to obtain suitable input form for the generative adversarial networks (GAN). Our proposed models are capable of generating ECG signals containing three different heartbeat types: normal beat, left bundle branch block beat and right bundle branch block beat. The synthetic ECG signals generated by our models are more realistic since deep artificial neural networks can discover intricate structure and characteristics of real ECG signals instead of manually setting specific parameters for synthesis. Besides, ECG signals produced by one of our proposed models could be naturally continuous and be up to more than 20 seconds. Furthermore, we first provide an evaluation approach for quantitatively demonstrating the performance of ECG generative models. The study demonstrates that deep learning is a feasible and effective method for ECG generation. Our proposed ECG generative models can be utilized to assess biomedical signal processing algorithms so as to improve their performance in clinical trials.