Fig. 1: GAN architecture for reconstructing ECGs using Lead I or Leads I and II. | Communications Medicine

Fig. 1: GAN architecture for reconstructing ECGs using Lead I or Leads I and II.

From: Evaluating the feasibility of 12-lead electrocardiogram reconstruction from limited leads using deep learning

Fig. 1

a The generator takes Lead I or Leads I and II as input and reconstructs the additional electrocardiogram (ECG) leads V1, V2, V3, V4, V5, and V6 as output. b The discriminator receives real and synthesized (fake) ECGs as input and classifies them as real or fake. Each block represents a layer type, with colors distinguishing them. The generator follows an encoder-decoder structure, using one-dimensional convolutional layers (Conv1d()), Leaky Rectified Linear Unit activation (LeakyReLU()) and Rectified Linear Unit activation (ReLU()), reflection padding (ReflectPadding), and up-sampling layers (Upsample()), ending with a hyperbolic tangent activation (Tanh()). The discriminator consists of stacked one-dimensional convolutional layers (Conv1d()) with Leaky Rectified Linear Unit activation (LeakyReLU()), followed by a linear layer and sigmoid activation (Sigmoid()).

Back to article page