Brainpack: a suite of advanced statistical techniques for multi-subject and multi-group fMRI data analysis

Functional magnetic resonance imaging (fMRI) data based on blood oxygenation level dependent (BOLD) signal have become widely available, leading to exponential growth in the number of published studies reporting on human brain function. fMRI data have also posed challenges, including a low signal to noise ratio, various noise sources, correlation between observations, and size of the data set. Also, researchers are interested in drawing conclusions from a sample of subjects to a relevant population, and in comparing the performance between groups of people. Our motivating fMRI data involve both block and event-related runs, multiple tasks, scanning sessions, and groups of subjects. The objective of this study is to identify brain regions associated with performance of cognitive tasks and to observe the effects of practice as measured by BOLD signal across different tasks and contexts. To accomplish the goal, we develop a suite of reliable and robust statistical tools, called BrainPack, that is composed of aggregation, decorrelation, data volume reduction, cluster analysis, and comparison of group clustered maps. The proposed approach does not require a specific model, can detect signals from noisy data, and take temporal correlations into account compared to model-based analysis. Through use of the BrainPack suite, we find practice-induced BOLD signal attenuation across groups and tasks in regions associated with sensorimotor and cognitive control processes. The BrainPack application improves existing between-group analysis methods to solve persistent problems in fMRI data analysis using the following advancements: (i) robust, effective, and powerful analyses for identifying neural circuits across any group using statistical learning methods and (ii) optimized multiple group analysis methods using simultaneous comparison of group maps.