Face prediction from fMRI data during movie stimulus: strategies for feature selection

Authors:
Jukka-Pekka Kauppi;Heikki Huttunen;Heikki Korkala;Iiro P. Jääskeläinen;Mikko Sams;Jussi Tohka
Affiliations:
Tampere University of Technology, Dept. of Signal Processing, Tampere, Finland;Tampere University of Technology, Dept. of Signal Processing, Tampere, Finland;Tampere University of Technology, Dept. of Signal Processing, Tampere, Finland;Aalto University School of Science, Dept. of Biomedical Engineering and Computational Science, Espoo, Finland and Aalto University, Advanced Magnetic Imaging Centre, Espoo, Finland;Aalto University School of Science, Dept. of Biomedical Engineering and Computational Science, Espoo, Finland and Aalto University, Advanced Magnetic Imaging Centre, Espoo, Finland;Tampere University of Technology, Dept. of Signal Processing, Tampere, Finland
Venue:
ICANN'11 Proceedings of the 21st international conference on Artificial neural networks - Volume Part II
Year:
2011

Citing 2
Cited 0

Learning to Decode Cognitive States from Brain Images

Machine Learning
Parameter determination of support vector machine and feature selection using simulated annealing approach

Applied Soft Computing

Quantified Score

Hi-index	0.00

Visualization

Abstract

We investigate the suitability of the multi-voxel pattern analysis approach to analyze diverse movie stimulus functional magnetic resonance imaging (fMRI) data. We focus on predicting the presence of faces in the drama movie based on the fMRI measurements of 12 subjects watching the movie. We pose the prediction as a regression problem where regression coefficients estimated from the training data are used to estimate the presence of faces in the stimulus for the test data. Because the number of features (voxels) exceeds the number of training samples, an emphasis is placed on the feature selection. We compare four automatic feature selection approaches. The best results were achieved by sparse regression models. The correlations between the face presence time-course predicted from fMRI data and manual face annotations were in the range from 0.43 to 0.62 depending on the subject and pre-processing options, i.e., the prediction was successful. This suggests that proposed methods are useful in testing novel research hypotheses with natural stimulus fMRI data.