Has-Motion Resources

Can Principal Component Analysis Remove Lab Effects in Motion Capture Data?

If you have worked in a gait lab, then you already know that marker placement is hard, camera calibration isn’t always perfect, and lab effects are inevitable in long-term and/or multicenter studies. But what do we do about it? When our data looks good and gives us answers, do we always consider variability in the lab environment? This is something that continues to be a challenge to collaborative projects, especially those using multi-lab data collection. Could we use Principal Component Analysis (PCA) as a correction technique for cross-lab variability?

Photo by Unlimited Motion Ltd on Unsplash

Principal Component Analysis

PCA is a method of waveform analysis that reduces raw data into principal components. Principal components (PCs) are abstract mathematical constructions of new characteristics  that show as much variation across the dataset as possible, while retaining the important information about the dataset. To learn more about PCA, see our blog post.

Using PCA to compare subject groups and conditions in a single lab study is a common application of PCA in biomechanics. PCA quantifies the variability between groups while considering the entire waveform of the underlying data, rather than comparing individual metrics that summarize the wave (such as the mean or peak). But categorizing waveform patterns by comparing two study groups is just one example of how we can utilize PCA and waveform variability.

Cross-Lab Variability

So what does this have to do with cross-lab variability? Well, if we are able to pinpoint variation in gait measures between subject groups, can we identify variation from lab effects using PCA? Allison Clouthier and her colleagues suggested so in their paper Correcting waveform bias using principal component analysis: Applications in multicentre motion analysis studies

This study used PCA as a correction tool to remove variation that would be inevitable when combining waveform data from different research centers. Much of the variability due to lab effects come from marker placement, but also come from the software and hardware tools, lab environment, trial instruction and techniques, etc. 

When we perform PCA, we give each feature extracted from the waveform a Principal Component (PC) score. Clouthier et. al. compares the PC scores of these features for each dataset (in this case, from different lab centers). If they see statistically significant variance in any PCs, they consider these features to be lab effects. To remove these effects, Clouthier et. al. shows how to correct the waveform biases so that all PCs across the datasets from different centers would have the same means. By ensuring all PC means are equal, they can confidently combine the datasets since they have removed the bias or lab effects.

What is interesting about this method is that Clouthier et. al. apply the correction to each participant’s waveforms from one dataset. This approach maintains the variability within the specific dataset while reducing the overall variability in the combined dataset by removing lab effects. This ensures that they are not accidentally wiping away the natural variability across the subjects that would be of importance to the study. 

Clouthier et. al. tested this by inducing ‘lab effects’ by rotating the thigh segment by 10 degrees for one dataset, and used PCA on the knee adduction angle to observe the variance. By rotating the thigh, the resulting knee angles would differ significantly from the original dataset, providing an obvious overall variation. Results showed that the first two PCs differed between the two datasets. This indicates variability resulted from the induced ‘lab effects’ of the 10-degree rotated thigh segment. By correcting the individual strides of the red dataset (with ‘lab effects’), Clouthier was able to align the mean PC scores, reducing the overall variation between datasets.

Clouthier et. al. mentioned that in the induced variability study, it was obvious which dataset was inherently wrong. It made sense that if they found variability in PC means, they would make adjustments to the dataset with the altered thigh segment to align with the unaltered data. But, it may not always be the case where you know which datasets to ‘correct’. Discussions between centers would need to determine which dataset to choose for agreement among all parties.

Outro

We have seen PCA used in biomechanics to reduce high dimensional time varying signals, while maintaining enough variability to distinguish between study groups (see our blog here). We have observed how Clouthier et al. characterized and accounted for cross-lab variability in their data analysis, demonstrating additional PCA applications in biomechanics.  So this can work, but will researchers use it? Will PCA be a key step to pre-processing multi-center data? If you are working with multi-center data, we think you should try this out! We believe that this method could offer a great solution, but we question its adoption across the field. Try this out with your multi-center data, or your own induced variability, and let us know what you think!

Enjoyed this discussion? Follow us on linked-In to see more content and catch future blog posts.

Interested in seeing how you can run PCA in our application Sift? See our previous PCA blog post, or email info@has-motion.ca for more information.

Like this article? Share with friends!

Facebook
Twitter
LinkedIn

Return To Our Resources

HAS-Motion

Our Software Solutions

Latest Resources