Beat-to-Beat Blood Pressure Estimation by Photoplethysmography and Its Interpretation

Figure 1. Visualization of the segmentation and decomposition of a PPG beat. The green circle marks the detection point from the algorithm of Lazaro et al. [49]. The black dashed lines display the interval around the detection point in which the minima (red crosses) are searched. The beat between these minima is then decomposed into two kernels (light grey). The sum of the kernels is the recomposed beat (black line).

Figure 2. Comparison of estimated blood pressure (orange line) and ground truth (blue line) for all samples of the test set. The top plot shows the CPT data, the middle plot shows the Queensland data, the bottom plot shows the PPG-BP data.

Figure 3. (a) shows a bar plot of mean absolute SHAP values that indicates global feature importance (mean absolute for each feature over all given samples). (b) depicts a beeswarm plot of SHAP values. For each sample, this plot shows a dot on each feature row. The features are ordered according to the mean absolute SHAP values for each feature. Depicted are the nine features with the highest mean absolute SHAP value and the sum of the remaining features.

Figure 4. Dependence plot for features skew and b/a showing the relationship of the value of the feature and its SHAP value for all samples. All samples with feature values in the first and 99th percentile are removed.

Figure 5. Beeswarm plot of SHAP values for subsets of samples of test set according to the MAE values of the prediction models. (a) error greater than two times the MAE, (b) error lower than half the MAE.

Figure 6. Beeswarm plot of SHAP values for subsets of samples of test set according to the predicted SBP. (a) prediction lower than 0.8 times the mean prediction, (b) prediction greater than 1.2 times the mean prediction.

Figure 7. Beeswarm plot of SHAP values for subsets of samples of test set according to the database.

Figure 8. Bar plot of mean absolute SHAP values for comparison models. Shows global feature importance (mean absolute for each feature over all given samples).

Figure 9. SHAP values for DBP prediction model on the whole dataset. (a) shows a bar plot of mean absolute SHAP values that indicates global feature importance (mean absolute for each feature over all given samples). (b) depicts a beeswarm plot of SHAP values. For each sample this plot shows a dot on each feature row. The features are ordered according to the mean absolute SHAP values for each feature. Depicted are the nine features with the highest mean absolute SHAP value and the sum of the remaining features.