Multivariate Outlier Nominators for fMRI Data

• Activity Number: 33
• Type: Contributed
• Date/Time: Sunday, July 29, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): John F. Magnotti*+ and Nedret Billor
• Companies: Auburn University and Auburn University
• Keywords: multivariate outlier detection ; principal components analysis ; fMRI ; multivariate pattern analysis ; support vector machine

Abstract

Within the field of cognitive neuroscience, there is increasing interest in using multivariate pattern analysis (MVPA) to classify fMRI data into specific brain states. MVPA is also being applied to find patterns of activity that are diagnostic of certain clinical conditions (e.g., ADHD). Many classification algorithms are sensitive to multivariate outliers, but popular fMRI analysis software is mostly concerned with univariate outliers. The presence of outliers may lead to incorrect estimates of neuroanatomical regions diagnostic of specific brain states, as well as decreased classifier generalization. Problematic for some detection algorithms, fMRI datasets have orders of magnitude more dependent variables than observations and are large enough to make some approaches intractable. We adapted the BACON algorithm to fMRI data and compared its performance with other algorithms across simulated and real data. Results indicated that the BACON algorithm performed accurately and efficiently on both datasets. We recommend that studies using MVPA on fMRI data incorporate a multivariate outlier nomination step to increase both representativeness and generalizability of their classifiers.