Multi-Subject fMRI Generalization with Independent Component Representation

Rasmus E. Madsen

AbstractGeneralizability in a multi-subject fMRI study is investigated. The analysis is based on principal and independent component representations. Subsequent supervised learning and classification is carried out by canonical variates analysis and clustering methods. The generalization error is estimated by cross-validation, forming the so-called learning curves. The fMRI case story is a motor-control study, involving multiple applied static force levels. Despite the relative complexity of this case study, the classification of the 'stimulus' shows good generalizability, measured by the test set error rate. It is shown that independent component representation leads to improvement in the classification rate, and that canonical variates analysis is needed for making generalization cross multiple subjects.
KeywordsIndependent Component Analysis (ICA), functional Magnetic Resonance Imaging (fMRI), Canonical Variates Analysis (CVA), Principal Component Analysis (PCA), Multiple Subjects
TypeTechnical report
Year2003
PublisherInformatics and Mathematical Modelling, Technical University of Denmark, DTU
AddressRichard Petersens Plads, Building 321, DK-2800 Kgs. Lyngby
SeriesIMM-Technical Report-2003-19
Electronic version(s)[pdf]
BibTeX data [bibtex]
IMM Group(s)Intelligent Signal Processing