11 / 2022 – 04 / 2023
The importance of machine learning in in the field of biomechanics has grown in the past years and will increase even more in the future. Nevertheless, in most studies, in which machine learning is applied to human movement data, only a single dataset with few participants is used for training and testing the model, as data recording is expensive and time consuming. An optimal machine learning model should be generalizable to new unseen data though, which is very unlikely if it was only trained on a single, usually small, dataset. Combining existing datasets can be an opportunity to improve a machine learning model by training it on more data and more heterogeneous data without spending time and effort on recording new data . Yet, even if several datasets of the same task exist, the data will differ due to influencing factors like the marker placement, test instructions or experimental setup , which might bias the model result. Previous work has shown that when combining different datasets of overground walking for unsupervised clustering, the cluster result is in some cases influenced by the original data sources. The clustering should group together persons with the same anthropometric characteristics and walking pattern, yet it could be observed that in some cases the original datasets influenced the outcome . The marker placement of several markers, as well as varying test instructions were identified as influencing factors, but it could not be fully determined what part of the individual datasets (that contained marker and ground reaction force trajectories (GRFs)) was the most relevant factor for the observed result. In order to combine different datasets for future machine learning applications in biomechanics, it is essential to know what exactly biased the outcome and how we can overcome this.
The purpose of this thesis is to further investigate which factors in the gait datasets, that all measure the same task of overground walking, cause differences between original data sources that could influence a machine learning model. Specifically, the goal is to investigate if a machine learning model can differentiate between data sources, and find what causes exist for these differences. To do so, the datasets should first be used in a supervised machine learning task, where the model inputs are the individual marker and GRF trajectories, and the label to predict is the original data source. Secondly, the model predictions should be attributed back to the input variables, to understand what exactly the model learns or why it classifies a certain trial as belonging to a certain dataset. A possible way of achieving this is using a Layer-Wise Relevance Propagation technique. This method has been used in previous work to investigate which variables (marker and GRF trajectories) at what time windows of the gait cycle are most relevant for the characterisation of gait patterns from a certain individual . It is conceivable to apply the same technique to investigate which variables are especially relevant for the characterization of individual datasets. The code from  is publicly available and can be used as a basis / reference point for the implementation of the model.
 Ferber, R, et al.: Gait Biomechanics in the Era of Data Science. J Biomech 49.16, 2016.
 Bendetti, MG, et al.: Inter-laboratory consistency of gait analysis measurements. Gait Posture 38.4, 2013.
 Fleischmann, S, et al.: Feature extraction and clustering of motion capture and force plate data to improve personalized musculoskeletal model development. (Unpublished Master’s thesis), 2021
 Horst, F, et al.: Explaining the unique nature of individual gait patterns with deep learning. Sci Rep 9 2391, 2019.