Applying feature selection methods on fMRI data

Applying feature selection methods on fMRI data

Van Schooten, S.
Harel, R.
Ercan, S.
De Groot, E.

Schooten, S.

Electrical Engineering, Mathematics and Computer Science

Intelligent Systems

2014-03-29

In neuroscience, the ability to correlate and classify certain activity patterns of the brain to different physical and mental states of the subject is of high importance. Analysis of fMRI data is one of the venues in which this objective is being pursued. However data produced using fMRI technology is highly complex. To this end, machine learning becomes relevant. Prominent hurdles facing fMRI data analysis are their high-dimensionality (thousands of features per instance), low signal-to-noise ratio and interdependency. This motivates the use of feature selection methods in order to consolidate relevant information and discard noise. Many feature selection methods exist but only a few have been applied in the fMRI domain. In this paper we identify positive characteristics of feature selection algorithms that are beneficial when dealing with fMRI datasets. To do this, we evaluate representatives from each of the three main feature selection classes: Filters, wrappers and embedded methods. We have found probabilistic embedded methods to be the most suitable for fMRI data. We would therefore recommend using these (or similar) methods to process data with fMRI-like characteristics. Student report.

fMRI
feature selection

http://resolver.tudelft.nl/uuid:24680427-ae9e-4ddc-8e6a-1689a00a1cc9

Student project report

Student theses

student report

(c) 2014 Van Schooten, S.
Harel, R.
Ercan, S.
De Groot, E.