Quantifying inter-subject agreement in brain-imaging analyses.

In brain-imaging research, we are often interested in making quantitative claims about effects across subjects. Given that most imaging data consist of tens to thousands of spatially correlated time series, inter-subject comparisons are typically accomplished with simple combinations of inter-subject data, for example methods relying on group means. Further, these data are frequently taken from reduced channel subsets defined either a priori using anatomical considerations, or functionally using p-value thresholding to choose cluster boundaries. While such methods are effective for data reduction, means are sensitive to outliers, and current methods for subset selection can be somewhat arbitrary. Here, we introduce a novel "partial-ranking" approach to test for inter-subject agreement at the channel level. This non-parametric method effectively tests whether channel concordance is present across subjects, how many channels are necessary for maximum concordance, and which channels are responsible for this agreement. We validate the method on two previously published and two simulated EEG data sets.

Single-trial classification of MEG recordings.

While magnetoencephalography (MEG) is widely used to identify spatial locations of brain activations associated with various tasks, classification of single trials in stimulus-locked experiments remains an open subject. Very significant single-trial classification results have been published using electroencephalogram (EEG) data, but in the MEG case, the weakness of the magnetic fields originating from the relevant sources relative to external noise, and the high dimensionality of the data are difficult obstacles to overcome. We present here very significant MEG single-trial mean classification rates of words. The number of words classified varied from seven to nine and both visual and auditory modalities were studied. These results were obtained by using a variety of blind sources separation methods: spatial principal components analysis (PCA), Infomax independent components analysis (Infomax ICA) and second-order blind identification (SOBI). The sources obtained were classified using two methods, linear discriminant classification (LDC) and v-support vector machine (v-SVM). The data used here, auditory and visual presentations of words, presented nontrivial classification problems, but with Infomax ICA associated with LDC we obtained high classification rates. Our best single-trial mean classification rate was 60.1% for classification of 900 single trials of nine auditory words. On two-class problems rates were as high as 97.5%.