CODEWITHSUNDEEP
svm
David Lopezowski
David Lopezowski

Reputation: 1

MVPA related to the number of input features

I'm trying to run MultiVariate Pattern Analysis on fNIRS data. Generally, it has been used in fMRI where people search for patterns linked to activation with generally a high number of input features (hundreds of voxels). In fNIRS, however, we have very few input features (aprx. 24 per hemisphere). When running the analysis on the whole hemisphere i don't get any significant difference in the accuracy of the classifier but when i reduce the number of features to a specific region of 9 channels the classifier gives significant results. Can anyone explain to me why this is the case? When I vary one or more channel the results change a lot? Is it better to use a higher amount of channels and therefore those results are more robust?

In this particular case, MVPA uses an SVM linear classifier.

Upvotes: 0

Views: 30

Answers (1)

Sergio Novi
Sergio Novi

Reputation: 1

This is not a specific problem to fNIRS technology. Increasing the number of features does not necessarily lead to increasing the performance of the classifier. A few fNIRS channels may be introducing similar patterns among the classes which leads to poorer performance.

I recommend checking this fNIRS work for further details: https://www.sciencedirect.com/science/article/abs/pii/S0169260719324344

Best, Sérgio Novi

Upvotes: 0

Related Questions