1
$\begingroup$

Just a bit of context: We are applying Machine Learning algorithms in the field of Human Biomechanics. In a previous project, my colleagues identified three different subgroups in an injured population performing a movement utilising hierarchical clustering. In order to be able to classify new subjects, we developed a classification algorithm (using the same dataset) that was able to allocate each case to the predefined cluster.

We currently are working in another study that involves new comparing injured and healthy subjects performing the same task in which the clusters were found.

My questions are, is there any way to know to what extent clusters found in injured athletes apply to healthy athletes? If I used the same algorithm and features, what could be the effects/consequences? can I justify it? What are the limitations?

Any answer or directions to literature will be greatly appreciated as I am struggling to find any similar situation in the literature/internet.

Thanks everyone.

Cheers!

$\endgroup$
  • $\begingroup$ Elaborate on the question, what precisely do you want to do? $\endgroup$ – user2974951 Dec 4 '18 at 10:24
  • $\begingroup$ are you trying to find the healthy people among the injured population? $\endgroup$ – Bharath Kumar L Dec 5 '18 at 3:52
  • $\begingroup$ We're trying to assess whether healthy and injured athletes are equally consistent in their movement strategy selection. In other words, do injured and healthy patients get classified in the same subgroup over multiple repetitions of the same task or do they keep changing subgroups instead? Probably the best solution would be having a healthy and injured model and test them separately but this is not the case for us as we don't have access to a large set of healthy athletes. Either way, I think @Mark.F 's answer is right and there are more relevant biomech implications than ML ones. $\endgroup$ – YoungResearcher Dec 5 '18 at 8:42
0
$\begingroup$

I'm not sure this is the best place to post this question. This looks more like a question for Bio-mechanics experts / doctors / physiotherapists...

From a machine-learning prospective, you are attempting to use a model that was trained on a specific dataset of injured athletes and then use it to predict the classification of another (maybe) different dataset. In doing so, there is a hidden assumption that both populations share similar distributions.

You could try to train the model on data gathered from the new healthy athletes and attempt to classify the injured population and see if there are significant differences between the results of the 2 models.

$\endgroup$
0
$\begingroup$

The dataset (population) that you used to define the clusters was a different one from the one you are looking at now (healthy population). That means you can’t use your clusters to categorize healthy people just like that, but your research question is not what cluster each person belongs to.

If your question is to see whether people are consistent in their movement strategy then you can use the information from the clusters you have, or any other clusters you just made up. The cluster assignment is a proxy (a simplified representation) of each measurement, and you’re trying to see if they are consistent from the cluster numbers of each measurement.

This seems to me to lead to results that are difficult to interpret. What if healthy people are consistently very close to a border between two clusters? Consistent, but jittering between the clusters.

I would say you need to drop the idea of using a cluster as a proxy. Let’s say your measurement is how fast people can run a mile, time after time. To know if they’re consistent, you would just look at the standard deviation for each person.

I guess your problem is that you have many measurements in each example, and you’re trying to simplify by using a cluster number rather than the actual measurement. That’s dimensionality reduction, to a single dimension, which is extreme, and I wouldn’t use clustering for it. Is it really necessary to reduce the dimensions, or would it be possible to define a consistency metric on the original data? If it’s not and you need to reduce the dimensions, I would first turn to PCA and if that’s not enough, maybe an embedding from an auto-encoder.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.