Associate Professor Washington State University Pullman, Washington, United States
Introduction:: Excessive flexion angles of the neck that are intensified by mobile device use are associated with neck pain [1]. The forward position of the head causes an increased gravitational moment [2], resulting in increased spinal loads and muscle fatigue. Musculoskeletal models can be used to predict spinal loads, but previous work has shown that using common assumptions for determining the pose of the cervical vertebrae (position and orientation) from external markers does not provide accurate model results compared to using vertebral pose acquired from X-ray images [3]. External marker data is an attractive alternative due to accessibility and the elimination of radiation exposure, but a significant amount of information about the vertebral pose needs to be estimated. Here, we explore a variety of methods aimed at accurately estimating cervical vertebral pose using 5 external markers. We investigated the application of prediction models (linear regression, neural networks) and the feasibility of OpenSim’s [4] inverse kinematics (IK) with varying degrees of freedom (DOF). We hypothesized that linear regressions would perform best due to their ability to fit on small amounts of data. We were less certain about the ability of the NN to properly fit, and the efficacy of OpenSim’s inverse kinematics to return correct neck configurations given the lack of markers on C1-C6 vertebrae.
Materials and Methods:: The data were collected in an ergonomics study on tablet use [2]. Simultaneous lateral X-rays and photos were taken in 5 conditions (postures) for 30 subjects. For all models, predictors were x-y position (cm) of reflective markers on the canthus, tragus, C7 spinous process, and iliac crest, relative to the sternal notch, from photographs. Predicted variables were the x-y position and angle of C1-C7. We compared the following models: multivariate linear regression (LR), multivariate regression with covariance estimators (MRCE) [5], a shallow neural network (NN), 24 DOF neck model for OpenSim IK (independent rotations at each vertebral level), and a 3 DOF neck model for OpenSim IK (kinematically constrained rotations). Performance was assessed by mean absolute error (MAE) from the X-ray ground truth, and whether the returned neck pose was a valid configuration.
The multivariate linear regression and neural networks were implemented in Python using scikit-learn and Keras from TensorFlow, respectively. Various neural network architectures were explored, but here we used a neural network with 3 hidden layers with 64 units each and trained 100 iterations. MRCE was done in R. To overcome the limited dataset (n=150), we employed a leave-one-subject-out methodology, where we set aside each subject’s 5 postures as validation and averaged the MAE.
We used Python to interface with the OpenSim API to run IK. We first used the ScaleTool to register our markers and scale the model, locked the ground movement of the model, performed inverse kinematics, then compared the vertebral pose to the ground truth.
Results, Conclusions, and Discussions:: We found that the linear regression models were the most accurate (Table 1). MRCE and LR had comparable MAE and outperformed every other model. The NN model commonly predicted nonphysiological neck shapes. The dataset was small in both the number of samples and markers used. The size of the dataset and the nature of discontinuous static postures likely made it difficult for the NN to properly fit. Other neural network architectures were explored, but none could consistently output neck shapes that were physiologically reasonable.
OpenSim’s inverse kinematics is widely used for human gait, but we have reservations about the application for spine models. We wanted to explore its use for returning accurate cervical vertebral poses. In terms of MAE, IK resulted in much worse predicted positions and angles. We hypothesize that given 1) time-series data and 2) more markers may allow inverse kinematics solvers and neural networks to perform better.
Curating a larger dataset would most likely improve performance for the prediction models. However, few studies have used external markers from photographs to predict vertebral pose from X-rays simultaneously, which is the main method explored here. The external markers we used, except for the iliac crest, should all be identifiable on most cervical spine X-rays. Even without photographs, there may be value in taking the external marker locations from X-rays to predict the vertebral pose, as it would enable using a much larger dataset, although this is outside the scope of this project.
Getting accurate inputs to biomechanical models is crucial to receiving accurate outputs from the model. In the case of the neck, the inputs of cervical vertebral pose are most accurate when acquired through X-ray imaging. Predicting these inputs without X-ray will minimize the cost and risk to individuals and enable studies in more natural environments. In addition, this study only evaluated sagittal plane neutral and flexed postures, and the results would be different for continuous three-dimensional postures in a larger range. Current plans for this project are to use outputs of these predictive models in OpenSim to investigate the effects on spinal loading.
Acknowledgements (Optional): : We want to thank Ellis Hughes, Rebecca Hsieh, Charles Leahy, John Einsten, Dr. Francis Pascual, and Dr. Xianglong Wang for their contributions to this project. We thank the Office Ergonomics Research Committee for funding support for the subject data collection. This material builds upon work done for the 2021-2022 WSU Bioengineering Senior Capstone.