Page 235 - Artificial Intelligence for Computational Modeling of the Heart

208 Chapter 6 Additional clinical applications

Table 6.5 Results of the Deep Neural Network for time-dependent quantities on the test dataset.

Circulation Parameters MAE RMSE
Aortic pressure [mmHg] 1.46 0.35
Ventricular pressure [mmHg] 1.33 0.18
Systemic Atrial pressure [mmHg] 0.52 0.02
Ventricular volume [ml] 1.85 0.52
Aortic Flow rate [ml/s] 6.22 0.08
Pulmonary artery pressure [mmHg] 0.56 0.06
Ventricular pressure [mmHg] 0.57 0.01
Pulmonary Atrial pressure [mmHg] 0.52 0.01
Ventricular volume [ml] 1.33 0.54
Pulmonary artery Flow rate [ml/s] 3.13 0.07

Figure 6.13. Correlation between predictions and ground-truth. (A) Predicted vs.
Ground-truth time at max. elastance; (B) Predicted vs. Ground-truth resistance.

The relatively large runtime required for personalizing the
whole body circulation model could represent an important draw-
back for the clinical adoption of closed-loop hemodynamic mod-
els in contexts in which their contribution could be relevant,
such as for continuous monitoring of patients in intensive care
units. Hence, a method for efficient estimation of clinically rele-
vant measures of interest such as the PV loop could represent a
powerful diagnostic tool in specific clinical workflows. This work
demonstrates the definition of a new model based on two deep
neural networks trained on a synthetically generated database of
clinically observable quantities (as input) and clinically relevant
measures of interest (as ground truth values) estimated by a WBC