Toward a neurometric foundation for probabilistic independent component analysis of fMRI data

Abstract

Improved fMRI data analysis methods hold promise for breakthroughs in cognitive and affective neuroscience. Group probabilistic independent component analysis (pICA), such as that implemented by MELODIC (Beckmann & Smith IEEE Transactions on Medical Imaging 23:137-152, 2004), is one popular technique that typifies this development. Recently pICA has been proposed to be a reliable method for studying connectivity networks (Zuo et al. NeuroImage 49:2163-2177, 2010); however, there is no "standard" way to complete a pICA, and the full impact of the options on neurometric properties of resulting components is unknown. In the present study, we sought to assess the robustness, reproducibility, and within-subject test-retest reliability of ICA in two data sets: The first included 30 subjects imaged 3 weeks apart while completing a cognitive control task, and the second included 27 subjects imaged 9 months apart during rest. In addition to examining the impact of analytic parameters on the neurometrics, this study was the first to simultaneously investigate within-subject reliability of ICA-derived components from rest and task fMRI data. Results suggested that for both task and rest, meta-level analyses using 25 subject orders optimized robustness of the components. The impact of dimensionality and voxel threshold for components was subsequently examined regarding properties of reproducibility and within-subject retest reliability. Component thresholds between 0.2 and 0.6 of the maximum value optimized reproducibility across multiple dimensionalities and produced generally fair to moderate reliability estimates (Cicchetti & Sparrow American Journal of Mental Deficiency 86:127-137, 1981). These guidelines strengthen the foundation for this data-driven approach to fMRI analysis by providing prescriptive findings and a descriptive set of neurometrics for resting-state and task fMRI.