Variational autoencoder–based neural electrocardiogram synthesis trained by FEM-based heart simulator

Figure 1. Overview of the proposed conditional variational autoencoder (VAE). The proposed method is based on a conditional VAE, which is composed of encoder and decoder neural networks. The encoder takes the electrocardiogram (ECG) x and cardiac parameters Y and outputs the distribution parameter of the latent variable z. The decoder takes the latent variable z and the cardiac parameters Y and outputs the ECG x.

Figure 2. Examples of the ground-truth and generated electrocardiogram (ECG) signals. Blue and orange lines represent the ground-truth and generated ECG signals, respectively. Gray vertical dotted lines represent the ground-truth or estimated positions. A: Example including the error of the S positions estimated by the NeuroKit2. B: Example with the lowest peak signal-to-noise ratio of lead I.

Figure 3. Mean absolute errors of individual electrocardiogram positions. Each heatmap corresponds to the position mean absolute error (PMAE) of each position. The vertical and horizontal axes of each heatmap represent the latent variable size and the number of layers. The numbers and colors of individual elements in the heatmaps represent the calculated PMAEs.

Figure 4. Peak signal-to-noise ratios (PSNRs) of individual leads. Each heatmap corresponds to the PSNRs of each position. The vertical and horizontal axes of each heatmap represent the latent variable size and the number of layers. The numbers and colors of individual elements in the heatmaps represent the calculated PSNRs.

Figure 5. Position mean absolute errors (PMAE) obtained for the different cardiac parameter values. The 8 graphs correspond to the first 8 rows of Table 1, respectively. The vertical and horizontal axes of each graph represent the PMAE and the values that the cardiac parameter can take. The blue, orange, green, and red lines show the PMAE differences of the Q, R, S, and T positions, respectively. The vertical lines on the dots of the line plots represent the standard deviations of PMAEs.

Figure 6. Peak signal-to-noise ratios (PSNR) of lead I obtained for the different cardiac parameter values. The 8 graphs correspond to the first 8 rows of Table 1, respectively. The vertical and horizontal axes of each graph represent the PSNR and the values that the cardiac parameter can take.