Interpretation of model prediction using SHAP coefficients for spine surgery simulations based on finite element modeling and machine learning

Severe cases of adolescent idiopathic scoliosis require surgery to prevent further deformation of the spine. In the absence of a clinical standard, finite element modeling (FEM) analysis is often used to predict the outcome of different surgical plans and identify the most promising strategy. However, this is a time-consuming process. Machine learning (ML) models have the potential to predict spinal deformity in real-time once trained on FEM data. In this project, we applied SHAP analysis on a ML model trained on two FEM models (target shape and rod shape) to explain how the ML model learns the FEM models. Feature importance analysis suggests that the ML model adapts to account for the properties of the FEM model. Understanding which features are the most important depending on the FEM model will help to develop a more advanced ML model. In particular, a graph neural network could be implemented to group the features based on their importance and improve prediction accuracy.