Machine learning-based prediction of hemodynamic parameters in left coronary artery bifurcation: A CFD approach

Abstract

Coronary artery disease (CAD) is a leading cause of global mortality, often involving the development of atherosclerotic plaques in coronary arteries, particularly at bifurcation sites. Percutaneous coronary intervention (PCI) of bifurcation lesions presents challenges, necessitating accurate assessment of hemodynamic parameters such as wall shear stress (WSS) and oscillatory shear index (OSI) to predict acute coronary syndrome (ACS) risk. Computational fluid dynamics (CFD) provides valuable insights but is computationally intensive, prompting exploration of machine learning (ML) models for efficient hemodynamics prediction. This study aims to bridge the gap in understanding the influence of stenosis severity and location on hemodynamics in the left coronary artery (LCA) bifurcation by integrating ML algorithms with comprehensive CFD simulations, thereby enhancing non-invasive prediction of complex hemodynamics. An extensive dataset of 6858 synthetic LCA geometries with varying plaque severities and locations was generated for analysis. Hemodynamic parameters (TAWSS and OSI) were computed using CFD simulations and utilized for ML model training. Fourteen ML algorithms were employed for regression analysis, and their performance was evaluated using multiple metrics. The Decision Tree Regressor and K Nearest Neighbors models demonstrated the most effective prediction of TAWSS and OSI parameters, aligning well with CFD simulation results. The Decision Tree Regressor showed minimal prediction discrepancies (TAWSS: R2 = 0.998952, MAE = 0.000587, RMSE = 0.001626; OSI: R2 = 0.961977, MAE = 0.022264, RMSE = 0.041411) offering rapid and reliable assessments of hemodynamic conditions in the LCA bifurcation. Integration of ML algorithms with comprehensive CFD simulations provides a promising approach to enhance the non-invasive prediction of complex hemodynamics in the LCA bifurcation. The ability to efficiently predict hemodynamic parameters could significantly aid medical practitioners in time-sensitive clinical settings, offering valuable insights for coronary artery disease management. Further research is warranted to evaluate the effectiveness of deep learning models and address challenges in patient-specific applications.