Ensemble Learning-Based Surrogate Models for Non-Invasive Estimation of Corneal Mechanical Properties

Accurate estimation of corneal mechanical properties is essential for advancing ocular biomechanics and improving the diagnosis of diseases like keratoconus. Conventional inverse finite element methods are often limited by their high computational cost and time requirement. In this study, surrogate models based on Random Forest, XGBoost, and LightGBM were developed to predict third-order Ogden material parameters from simulated corneal apex displacement data. Model performance was rigorously evaluated through mean squared error (MSE), coefficient of determination (R2), and a physics-informed stress-stretch error metric. The results demonstrated strong predictive accuracy in capturing the cornea’s mechanical behavior. Among the models, XGBoost achieved the closest match to the mechanical response, Random Forest provided robust overall accuracy, and LightGBM offered the fastest training. This machine learning-based approach effectively bridges the gap between clinical measurement data and intricate biomechanical properties, offering a fast, reliable, and non-invasive alternative to traditional inverse FEM methods.