Predicting subject-driven actions and sensory experience in a virtual world with Relevance Vector Machine Regression of fMRI data

Research output: Contribution to journalArticleAcademicpeer-review

18 Citations (Scopus)

Abstract

In this work we illustrate the approach of the Maastricht Brain Imaging Center to the PBAIC 2007 competition, where participants had to predict, based on fMRI measurements of brain activity, subject driven actions and sensory experience in a virtual world. After standard pre-processing (slice scan time correction, motion correction), we generated rating predictions based on linear Relevance Vector Machine (RVM) learning from all brain voxels. Spatial and temporal filtering of the time series was optimized rating by rating. For some of the ratings (e.g. Instructions, Hits, Faces, Velocity), linear RVM regression was accurate and very consistent within and between subjects. For other ratings (e.g. Arousal, Valence) results were less satisfactory. Our approach ranked overall second. To investigate the role of different brain regions in ratings prediction we generated predictive maps, i.e. maps of the weighted contribution of each voxel to the predicted rating. These maps generally included (but were not limited to) 'specialized' regions which are consistent with results from conventional neuroimaging studies and known functional neuroanatomy. In conclusion, Sparse Bayesian Learning models, such as RVM, appear to be a valuable approach to the multivariate regression of fMRI time series. The implementation of the Automatic Relevance Determination criterion is particularly suitable and provides a good generalization, despite the limited number of samples which is typically available in fMRI. Predictive maps allow disclosing multi-voxel patterns of brain activity that predict perceptual and behavioral subjective experience. (PsycINFO Database Record (c) 2011 APA, all rights reserved) (journal abstract)
Original languageEnglish
Pages (from-to)651-661
JournalNeuroimage
Volume56
Issue number2
DOIs
Publication statusPublished - 1 Jan 2011

Cite this

@article{d1e483f237fb4deb871ddc1cae72da75,
title = "Predicting subject-driven actions and sensory experience in a virtual world with Relevance Vector Machine Regression of fMRI data",
abstract = "In this work we illustrate the approach of the Maastricht Brain Imaging Center to the PBAIC 2007 competition, where participants had to predict, based on fMRI measurements of brain activity, subject driven actions and sensory experience in a virtual world. After standard pre-processing (slice scan time correction, motion correction), we generated rating predictions based on linear Relevance Vector Machine (RVM) learning from all brain voxels. Spatial and temporal filtering of the time series was optimized rating by rating. For some of the ratings (e.g. Instructions, Hits, Faces, Velocity), linear RVM regression was accurate and very consistent within and between subjects. For other ratings (e.g. Arousal, Valence) results were less satisfactory. Our approach ranked overall second. To investigate the role of different brain regions in ratings prediction we generated predictive maps, i.e. maps of the weighted contribution of each voxel to the predicted rating. These maps generally included (but were not limited to) 'specialized' regions which are consistent with results from conventional neuroimaging studies and known functional neuroanatomy. In conclusion, Sparse Bayesian Learning models, such as RVM, appear to be a valuable approach to the multivariate regression of fMRI time series. The implementation of the Automatic Relevance Determination criterion is particularly suitable and provides a good generalization, despite the limited number of samples which is typically available in fMRI. Predictive maps allow disclosing multi-voxel patterns of brain activity that predict perceptual and behavioral subjective experience. (PsycINFO Database Record (c) 2011 APA, all rights reserved) (journal abstract)",
author = "G. Valente and {de Martino}, F. and F. Esposito and R. Goebel and E. Formisano",
year = "2011",
month = "1",
day = "1",
doi = "10.1016/j.neuroimage.2010.09.062",
language = "English",
volume = "56",
pages = "651--661",
journal = "Neuroimage",
issn = "1053-8119",
publisher = "Elsevier Science",
number = "2",

}

TY - JOUR

T1 - Predicting subject-driven actions and sensory experience in a virtual world with Relevance Vector Machine Regression of fMRI data

AU - Valente, G.

AU - de Martino, F.

AU - Esposito, F.

AU - Goebel, R.

AU - Formisano, E.

PY - 2011/1/1

Y1 - 2011/1/1

N2 - In this work we illustrate the approach of the Maastricht Brain Imaging Center to the PBAIC 2007 competition, where participants had to predict, based on fMRI measurements of brain activity, subject driven actions and sensory experience in a virtual world. After standard pre-processing (slice scan time correction, motion correction), we generated rating predictions based on linear Relevance Vector Machine (RVM) learning from all brain voxels. Spatial and temporal filtering of the time series was optimized rating by rating. For some of the ratings (e.g. Instructions, Hits, Faces, Velocity), linear RVM regression was accurate and very consistent within and between subjects. For other ratings (e.g. Arousal, Valence) results were less satisfactory. Our approach ranked overall second. To investigate the role of different brain regions in ratings prediction we generated predictive maps, i.e. maps of the weighted contribution of each voxel to the predicted rating. These maps generally included (but were not limited to) 'specialized' regions which are consistent with results from conventional neuroimaging studies and known functional neuroanatomy. In conclusion, Sparse Bayesian Learning models, such as RVM, appear to be a valuable approach to the multivariate regression of fMRI time series. The implementation of the Automatic Relevance Determination criterion is particularly suitable and provides a good generalization, despite the limited number of samples which is typically available in fMRI. Predictive maps allow disclosing multi-voxel patterns of brain activity that predict perceptual and behavioral subjective experience. (PsycINFO Database Record (c) 2011 APA, all rights reserved) (journal abstract)

AB - In this work we illustrate the approach of the Maastricht Brain Imaging Center to the PBAIC 2007 competition, where participants had to predict, based on fMRI measurements of brain activity, subject driven actions and sensory experience in a virtual world. After standard pre-processing (slice scan time correction, motion correction), we generated rating predictions based on linear Relevance Vector Machine (RVM) learning from all brain voxels. Spatial and temporal filtering of the time series was optimized rating by rating. For some of the ratings (e.g. Instructions, Hits, Faces, Velocity), linear RVM regression was accurate and very consistent within and between subjects. For other ratings (e.g. Arousal, Valence) results were less satisfactory. Our approach ranked overall second. To investigate the role of different brain regions in ratings prediction we generated predictive maps, i.e. maps of the weighted contribution of each voxel to the predicted rating. These maps generally included (but were not limited to) 'specialized' regions which are consistent with results from conventional neuroimaging studies and known functional neuroanatomy. In conclusion, Sparse Bayesian Learning models, such as RVM, appear to be a valuable approach to the multivariate regression of fMRI time series. The implementation of the Automatic Relevance Determination criterion is particularly suitable and provides a good generalization, despite the limited number of samples which is typically available in fMRI. Predictive maps allow disclosing multi-voxel patterns of brain activity that predict perceptual and behavioral subjective experience. (PsycINFO Database Record (c) 2011 APA, all rights reserved) (journal abstract)

U2 - 10.1016/j.neuroimage.2010.09.062

DO - 10.1016/j.neuroimage.2010.09.062

M3 - Article

VL - 56

SP - 651

EP - 661

JO - Neuroimage

JF - Neuroimage

SN - 1053-8119

IS - 2

ER -