Variational Bayesian mixed-effects inference for classification studies

摘要: Multivariate classification algorithms are powerful tools for predicting cognitive or pathophysiological states from neuroimaging data. Assessing the utility of a classifier in application domains such as neuroscience, brain-computer interfaces, clinical diagnostics necessitates inference on performance at more than one level, i.e., both individual subjects and population which these were sampled. Such requires models that explicitly account fixed-effects (within-subjects) random-effects (between-subjects) variance components. While this sort standard mass-univariate analyses fMRI data, they have not yet received much attention multivariate studies presumably because high computational costs entail. This paper extends recently developed hierarchical model mixed-effects introduces an efficient variational Bayes approach to inference. Using synthetic empirical we show is equally simple use as, than, conventional t-test subject-specific sample accuracies, computationally previous sampling permutation tests. Our independent type underlying thus widely applicable. The present framework may help establish future group analyses.