Cardiac parameter estimation from electrocardiograms based on inversion of physical simulator behavior

Fig 1. Overview of the network architecture of the proposed model. The proposed model is based on a DNN composed of an encoder network and a decoder network, which are a stack of one-dimensional convolutional layers and fully connected layers. The blue, green, and red boxes in the right figures represent data transformation layers with learnable parameters, activation function layers, and pooling layers, respectively. The transformation and pooling layers take several arguments, which are described in the box at the lower right of the figure. The proposed model is implemented with PyTorch 2.5.1, [34], and default values are used for the arguments not shown in the figure.

Fig 2. Dataset creation procedure. The training, validation, and test datasets used for the proposed method are created by splitting and shuffling the original dataset. The upper tables represent the original data folds, and the lower tables represent the data folds obtained through a cyclic data shuffle indicated by the five colors.

Fig 3. Evaluation results. The performance of the proposed method was measured by comparing the ground-truth cardiac parameters and the estimated cardiac parameters. The left line plot shows the mean absolute errors for the continuous cardiac parameters, and the right line plot shows the accuracies for the discrete cardiac parameters. The blue, orange, green, and red lines represent the performance of the proposed method, the decision tree, the XGBoost, and the random forest. The performance of the random forest is selected In this figure, we plot the results for the random forest with 100 trees and automatically determined depth. The horizontal axes in the plots indicate the kind of cardiac parameter.

Fig 4. Distributions of the estimated cardiac parameters. The violin plots and heatmaps represent the distributions of the estimated continuous and discrete cardiac parameters around the ground-truth values.

Fig 5. Evaluation results for noisy ECG signals. The performance of cardiac parameter estimation by the proposed method was measured on several ECG signals with different noise levels. In each graph, the horizontal axis represents the noise level, while the vertical axis represents the estimation performance.

Fig 6. Evaluation results for ECG signals with power-line interference. We evaluate the proposed method on ECG signals with varying levels of power-line interference (PLI) noise. In each graph, the horizontal axis represents the signal-to-noise ratio (SNR) in dB, which indicates the noise level, and the vertical axis represents the estimation performance. SNR=∞ indicates that the ECG signal contains no PLI noise. The blue and orange lines represent the performance under 50 Hz and 60 Hz power-line interference, respectively.