Introduction: : Right ventricle (RV) failure remains a leading cause of morbidity and mortality in patients with pulmonary hypertension (PH). In PH, chronic RV pressure overload triggers growth and remodeling (G&R) processes that may lead to either adaptive compensation or maladaptive changes resulting in eventual RV failure. Early prediction of remodeling trajectories could improve patient outcomes by enabling timely personalized interventions. However, accurately forecasting individual remodeling events from clinical data remains challenging due to the complex interactions between biomechanical, structural, and functional factors, as well as limited access to predictive, noninvasive biomarkers. Patient-specific finite-element modeling has emerged as a powerful tool that can integrate such details and simulate the longitudinal progression of remodeling in PH. Despite its potential to improve patient-specific care, the translational potential of FE modeling is limited by high computational demands and challenges in model personalization. Machine learning (ML) models trained on FE-simulated remodeling trajectories offer a promising pathway to leverage patient-specific clinical data for the rapid and scalable prediction of RV remodeling outcomes, potentially bridging the gap between high-fidelity computational modeling and real-time clinical decision-making.
Materials and
Methods: : We developed image-based biventricular FE computational cardiac models from cardiac magnetic resonance (CMR) imaging that incorporated subject-specific geometry, fiber architecture, and passive and active biomechanical behavior. These cardiac models were used to simulate G&R under chronic pressure overload conditions. A strain-driven growth framework was implemented, wherein the total deformation gradient (F) was decomposed into an elastic component (FE) and a growth component (FG), allowing for the independent characterization of mechanical deformation and biological remodeling. Forward simulations of the cardiac cycle were conducted using a range of parameterized inputs, including myocardial stiffness, myofiber orientation, and pressure overload. Approximately 500 synthetic remodeling trajectories were generated, encompassing a diverse span of pathological states. Key functional and structural parameters characterizing remodeling, such as ventricular dilation, RV wall thickening, and fiber reorientation, were extracted from each simulation. A multilayer feedforward perceptron (MLP) was then trained to predict the final remodeling state from clinically accessible baseline features, including hemodynamic indices of pressure overload and ventricular size/shape descriptors.
Results, Conclusions, and Discussions:: Preliminary synthetic data generation has demonstrated that variations in pressure overload and myocardial properties lead to diverse remodeling outcomes, consistent with clinical observations. We anticipate that MLP-based models will achieve high prediction accuracy; accuracy benchmarks are expected to included a strong correlation (R2>0.9) between predicted and simulated outcomes for key remodeling features. Importantly, ML surrogates can accelerate predictions by several orders of magnitude compared to traditional FE approaches, facilitating the transition to real-time clinical applications. While the current results are based on synthetic datasets, the proposed framework establishes a foundation for the future validation using patient-specific imaging and hemodynamic data.
Our proposed MLP-surrogate approach offers a novel strategy to predict biventricular G&R based on readily obtainable clinical metrics. By combining high-fidelity FE simulations with data-driven surrogate models, we aim to enable rapid, individualized predictions of cardiac remodeling trajectories in PH. This framework could ultimately support clinical decision-making in PH by forecasting disease progression and facilitating personalized therapeutic strategies.
