An artificial intelligence-based noninvasive solution to estimate pulmonary artery pressure

Figures

• 

  FIGURE 1. Comparison design. Each model is configured by a set of learning parameters and supplied features, either raw signal or wavelet scattering features that were generated from waveform records with different window sizes and step lengths. The analyzed values of the window size ranged from 0.5 to 5 s, and the step sizes ranged from 0.1 s to the value of the window size. An example of the segmentation of waveform records is shown in Figure 2B. A total of 11 models were compared, including generalized linear regression, ridge regression, lasso regression, stochastic gradient descent regression, support vector machine regression, nearest neighbors regression, Gaussian process regression, random forest regression, extremely randomized trees regression, extreme gradient boosting tree regression, and residual convolutional neural network.

• 

  FIGURE 2. Demonstration of six signals and segmentation method. (A) ABP, CVP, respiration, PPG, and ECG are input predictors, and PAP is the outcome variable. (B) The sliding window method was adopted in this study to generate input samples.

• 

  FIGURE 3. Wavelet scattering transform and wavelet decomposition. (A) A segment of PAP signal was decomposed to seven level components by a symlet4 wavelet function, representing the frequency from high to low. (B) Four spectrumgrams present a segment of ECG signals after wavelet scattering transformation in which the four filter banks and the Gabor wavelet function were used.

• 

  FIGURE 4. Network architecture. The resNet50 design structure has 16 residual blocks, presenting 50 convolutional layers.

• 

  FIGURE 5. Prediction result compared with observational data. To inspect the prediction result visually, a randomly selected segment of observational PAP signal was compared with prediction ones given by AI model with five input signals and three ones, respectively. The red line presents the observational PAP signal collected by the catheter procedure; the green line indicates the prediction results using five input signals; the blue line shows the prediction results using three input signals. On average, the estimation by five input signals is higher than that by three input signals.