Neural Network-Driven Finite Element Modeling for Estimating Knee Joint Cartilage Mechanical Responses

Abstract Low-fidelity, artificial intelligence (AI)-generated approaches are increasingly used to estimate kinematics and kinetics of human movement, which are traditionally measured by high-fidelity motion capture (Mocap) approaches. However, there are no methods to study knee joint tissue mechanics using such low-fidelity approaches. To do so, we investigated knee cartilage stresses and strains using finite element (FE) models driven by both high- and low-fidelity motion capture methods. We performed subject-specific FE modeling on nine healthy participants to evaluate tissue mechanical responses by two different approaches. High-fidelity kinematic and kinetic data were obtained through motion capture and musculoskeletal modeling, respectively, whereas artificial neural networks (ANNs) were used to estimate kinetic data from low-fidelity data (e.g., subject mass, height, age, gender, static knee abduction-adduction angle, and walking speed). These data were then used as loading inputs in the knee joint FE models that were generated from magnetic resonance images. The results indicated that the high- and low-fidelity approaches provided comparable estimates of maximum principal stress, maximum shear strain, and collagen fibril strain of tibial cartilage at the first peak of knee contact force ( p > 0.05). Significant differences between the methods in the peak values of the analyzed parameters were observed at the second peak of knee contact force ( p < 0.05), larger differences observed in the lateral joint compartment. However, most of these differences disappeared when comparing average values over the cartilage–cartilage contact areas. Our findings support the potential of the low-fidelity, AI-generated approach to assess tibial cartilage mechanics. This out-of-laboratory tool may enable cartilage failure prediction and improved management of osteoarthritis.