Frontiers in Bioengineering and Biotechnology (May 2022)
A Parameter Sensitivity Analysis on Multiple Finite Element Knee Joint Models
Abstract
The reproducibility of computational knee joint modeling is questionable, with models varying depending on the modeling team. The influence of model variations on simulation outcomes should be investigated, since knowing the sensitivity of the model outcomes to model parameters could help determine which parameters to calibrate and which parameters could potentially be standardized, improving model reproducibility. Previous sensitivity analyses on finite element knee joint models have typically used one model, with a few parameters and ligaments represented as line segments. In this study, a parameter sensitivity analysis was performed using multiple finite element knee joint models with continuum ligament representations. Four previously developed and calibrated models of the tibiofemoral joint were used. Parameters of the ligament and meniscus material models, the cartilage contact formulation, the simulation control and the rigid cylindrical joints were studied. Varus-valgus simulations were performed, changing one parameter at a time. The sensitivity on model convergence, valgus kinematics, articulating cartilage contact pressure and contact pressure location were investigated. A scoring system was defined to categorize the parameters as having a “large,” “medium” or “small” influence on model output. Model outcomes were sensitive to the ligament prestretch factor, Young’s modulus and attachment condition parameters. Changes in the meniscus horn stiffness had a “small” influence. Of the cartilage contact parameters, the penalty factor and Augmented Lagrangian setting had a “large” influence on the cartilage contact pressure. In the rigid cylindrical joint, the largest influence on the outcome parameters was found by the moment penalty parameter, which caused convergence issues. The force penalty and gap tolerance had a “small” influence at most. For the majority of parameters, the sensitivity was model-dependent. For example, only two models showed convergence issues when changing the Quasi-Newton update method. Due to the sensitivity of the model parameters being model-specific, the sensitivity of the parameters found in one model cannot be assumed to be the same in other models. The sensitivity of the model outcomes to ligament material properties confirms that calibration of these parameters is critical and using literature values may not be appropriate.
Keywords
