Comparison of three machine learning methods to estimate myocardial stiffness

Myocardial mechanical remodeling occurs in several structural heart diseases altering the mechanical properties of the myocardium. Estimating mechanical properties of the myocardium has been suggested as a promising biomarker for assessing disease progression. Development of methods to estimate patient-specific myocardial mechanical properties from clinical data, without the use of expensive finite-element (FE) approaches, has been a longstanding challenge in biomechanical studies. In this paper, we evaluated the accuracy and feasibility of three machine learning (ML) models to predict myocardial material properties by directly employing anatomical and hemodynamical features, forgoing the use of a computationally intensive FE problem. A synthetic data set of 2000 heart examples was created for training, using MRI data of 25 rodent hearts, including healthy rats and rats with myocardial infarction and pulmonary hypertension. Anatomical features included cardiac geometry and fiber orientation, attainable from cardiac imaging, and the hemodynamic feature was the end-diastolic pressure-volume relationship (EDPVR), which can be estimated from a single pressure—volume point. We predicted myocardial material properties by training three ML models, namely, (i) random forest (RF), (ii) support vector machine (SVM) and (iii) artificial neural network (ANN), with a large data set of geometries, fiber orientations and EDPVR loading responses of rodents obtained from forward FE simulations as input features. Among the three models, ANN performed excellently (Raf2=99.12% and Rbf2=93.31%), followed by SVM (Raf2=93.76% and Rbf2=84.34%) and RF (Raf2=70.53% and Rbf2=13.04%) in estimating the fiber-stiffness parameters af and bf from an exponential constitutive model. The trained ANN model showed the least uncertainty in the predicted parameters. Furthermore, the importance of features compared to each other was evaluated in each model by applying suitable feature importance techniques, including mean decrease in impurity, permutation feature importance and Shapley additive explanation. Regions of the EDPVR curve corresponding to the largest loading pressures were found to be the most important features in all models. The ML models investigated here showed the ability to support the current clinical evaluation of heart diseases with structural information including accurate myocardial properties attainable from existing organ-level indices, ultimately improving the diagnosis and prognosis of cardiac diseases.