Machine learning has become increasingly important in biomechanics. It allows to unveil hidden patterns from large and complex data, which leads to a more comprehensive understanding of biomechanical processes and deeper insights into human movement. However, machine learning models are often trained on a single dataset with a limited number of participants, which negatively affects their robustness and generalizability. Combining data from multiple existing sources provides an opportunity to overcome these limitations without spending more time on recruiting participants and recording new data. It is furthermore an opportunity for researchers who lack the financial requirements or laboratory equipment to conduct expensive motion capture studies themselves. At the same time, subtle interlaboratory differences can be problematic in an analysis, due to the bias that they introduce. In our study, we investigated differences in motion capture datasets in the context of machine learning, for which we combined overground walking trials from four existing studies. Specifically, our goal was to examine whether a machine learning model was able to predict the original data source based on marker and GRF trajectories of single strides, and how different scaling methods and pooling procedures affected the outcome. Layer-wise relevance propagation was applied to understand which factors were influential to distinguish the original data sources. We found that the model could predict the original data source with a very high accuracy (up to >99%), which decreased by about 15 percentage points when we scaled every dataset individually prior to pooling. However, none of the proposed scaling methods could fully remove the dataset bias. Layer-wise relevance propagation revealed that there was not only one single factor that differed between all datasets. Instead, every dataset had its unique characteristics that were picked up by the model. These variables differed between the scaling and pooling approaches but were mostly consistent between trials belonging to the same dataset. Our results show that motion capture data is sensitive even to small deviations in marker placement and experimental setup and that small inter-group differences should not be overinterpreted during data analysis, especially when the data was collected in different labs. Furthermore, we recommend scaling datasets individually prior to pooling them which led to the lowest accuracy. We want to raise awareness that differences in datasets always exist and are recognizable by machine learning models. Researchers should thus think about how these differences might affect their results when combining data from different studies.
Read full abstract