A comparative analysis of kinematic simulation results obtained by manually and automated scaled OpenSim models during walking – preliminary findings

Musculoskeletal simulations provide a deeper understanding of human movement by not only estimating joint movements and external joint loadings but also muscle forces, and internal joint loading parameters [1]. To date, the scaling and personalization of musculoskeletal models are performed manually in an iterative and time-consuming process [2]. This might hinder their use in clinical settings, impedes evaluating a large amount of individuals with knee malalignment and, thus, also prevents usage of trending technologies such as Machine Learning and Big Data analysis [3]. To address this, a Matlab pipeline was developed to automate the scaling process and subsequent personalization steps in OpenSim. This study’s primary objective is to evaluate the kinematic differences between the two methods. To what extent does the kinematic output of automatically scaled models correspond with by-hand scaled model output? Sixteen children and adolescents with valgus malalignment of at least one lower limb participated in this study. Manual scaling was performed by an experienced user in OpenSim 3.3 via the GUI [2] with a modified generic model [4,5]. Frontal knee alignment was manually adjusted based on X-ray imaging data. Automatic processing was done through a custom Matlab workflow using the same generic model with one accompanied standardized markerset. A static trial was used to automatically adjust each model’s frontal knee alignment based on the knee valgus angle angles of the static trial [4], run initial marker registration, and finally OpenSim’s scaling tool. Time-normalized kinematic parameters of the pelvis, hip, knee, and ankle joints were compared using parametric t-tests from the Statistical Parameter Mapping package (SPM) [6]. Mean differences and root mean squared errors (RMSE) were calculated for all kinematic parameters. The RMSE varied between 1.3°±0.6 (pelvic rotation) and 5.7°±2.8 (hip flexion-extension) (Fig. 1). Download : Download high-res image (667KB)Download : Download full-size image The literature suggests that errors of 2 degrees or less are likely to be generally acceptable, while errors of up to 5 degrees require consideration in data interpretation [7,8]. Our observed RMSE varied between 1.8° and 5.7°. Greatest errors were found for the pelvis tilt and hip flexion, presumably caused by an inaccurate placement of the sacrum marker for the automated marker registration. Our data suggest that automated processing seems able to generate sufficiently enough accurate participant-specific simulations. Approaches such as recently described bilevel optimization for scaling and marker registration presumably could further improve the accuracy and should be considered in the near future for automated approaches. Effects on calculated joint and muscle loading parameters also need to be evaluated in a next step [9].