(H-292) Towards Machine-learning Enabled Hemodynamic Assessment for Total Cavopulmonary Connection

Introduction::
Single ventricle malformations are one of the most severe and complex heart defects that occur at birth, affecting approximately 2 in every 1000 live births in the United States. The Fontan procedure is a common palliative surgery for patients with single ventricle malformations. This procedure involves creating a total cavopulmonary connection (TCPC) to redirect venous blood flow directly to the pulmonary circulation.









TCPC hemodynamics is critical to the surgical outcomes and quality of life of Fontan patients. In recent years, a Fontan surgical planning paradigm has emerged that utilizes computational fluid dynamics (CFD) to predict TCPC hemodynamics, thereby improving TCPC hemodynamics prior to surgery. While this paradigm has yielded promising results in research settings, the high computational burdens of CFD have delayed the clinical use of Fontan surgical planning.









The aim of this study is to develop a data-driven flow modeling framework based on machine learning (ML) to enhance the computational efficiency of the hemodynamic assessment of TCPC.




Materials and Methods::
In this study, we focus on the ML-enabled flow model for an idealized TCPC with time-averaged flow rates. The TCPC diameter is 15mm, and the volumetric flow rate of all vessels associated with the TCPC ranges from 0.5 to 4.5 L/min, covering flow rates commonly observed in real patients.









CFD was performed to obtain clinically relevant hemodynamic metrics, including power loss (PL) and hepatic flow distribution (HFD). ML models were developed to predict the relationship between TCPC flow parameters and PL and HFD.









The performance of four machine learning algorithms, including Random Forest, Support Vector Regression, Extreme Gradient Boosting, and Neural Networks, was evaluated. Additionally, the study investigated the identification of a minimal simulation training set required to create a neural network model capable of capturing the relationship between flow parameters and hemodynamics.




Results, Conclusions, and Discussions::
Results and Discussion: All ML models demonstrated strong correlations between their predictions and CFD results (r2 = 0.95). The neural network (NN) model performed slightly better than the other models in predicting hepatic flow distribution (HFD) (r2=0.98), while the NN and XGBoost models had similar performance in predicting power loss (PL) (r2=0.99).









Furthermore, the study identified a reduced number of simulations necessary to maintain accurate PL predictions using the NN model. Specifically, the study successfully employed only 968 simulations in the training set to make adequately accurate PL predictions, compared to the previously required 7056 simulations, as shown in the attached figure.




Conclusion: This study showcases the potential of utilizing ML-enabled models to enhance the efficiency of hemodynamic assessment in TCPC. The research will be further expanded to include patient-specific TCPCs and flow conditions. The final model is expected to substantially reduce the CFD burden in Fontan surgical planning, ultimately aiding doctors in improving surgical outcomes and the quality of life for Fontan patients.








Acknowledgements (Optional): :

References (Optional): :