Adaptive Constrained Independent Vector Analysis: Application to Large-Scale fMRI Analysis


Author/Creator ORCID




Chemical, Biochemical & Environmental Engineering


Engineering, Chemical and Biochemical

Citation of Original Publication


Distribution Rights granted to UMBC by the author.
This item may be protected under Title 17 of the U.S. Copyright Law. It is made available by UMBC for non-commercial research and education. For permission to publish or reproduce, please see or contact Special Collections at speccoll(at)



Functional magnetic resonance imaging (fMRI) provides non-invasive indirect measures of neuronal activation in the brain. Analysis of large-scale fMRI datasets, acquired from a large pool of individuals, has become prominent in recent studies for the identification of global functional networks that summarize the population, and networks specific to individuals, groups, conditions, or modalities. Dynamic functional network connectivity (dFNC) analysis has gained popularity over recent years for extracting networks that are functionally correlated and continuously changing over the scanning period, due to their ability to identify distinct biomarkers in a variety of disorders such as schizophrenia, bipolar disorder, post-traumatic stress and in different stages of development. However, the methods used to extract dFNC patterns mostly capture the time-varying associations of the spatial networks, while assuming that the spatial network itself is stationary over the scanning period. Hence, a model that allows for the variability in both spatial and temporal domain, and jointly extracts networks specific to individuals while exploiting the dependence across a large group of individuals, provides an efficient way for analysis of fMRI data. Group independent component analysis (gICA) has been widely-used for the analysis of multi-subject fMRI data for nearly two decades, and has been applied to jointly analyze data from a large pool of subjects. However, gICA employs a significant group-level dimension reduction to estimate a common spatial subspace, which may cause loss of individual specific information making it limited in terms of preserving subject-specific information. Joint ICA, on the other hand, assumes a common temporal domain across multiple subjects, making it a model mismatch to analysis of resting-state fMRI data. Independent vector analysis (IVA) assumes variability in both temporal and spatial domain, and extracts subject-specific spatio-temporal patterns by jointly analyzing multi-subject fMRI data, effectively preserving subject variability in a data-driven manner. However, the performance of IVA depends on a number of key aspects of the data, namely the number of datasets, number of sources, number of samples and level of correlation across datasets. In this work we study the effect of each of these aspect, and observe that the performance of IVA degrades with increase in number of datasets and number of sources, and decrease in the level of correlation across datasets, for a fixed number of samples. In fMRI analysis, the number of samples for each subject is fixed, and the use of large number of datasets and sources is desirable, with the sources exhibiting low level of correlation across datasets. Hence, the application of IVA on large-scale fMRI data often gives undesirable results. In this work, we propose the adaptive constrained IVA (acIVA) technique that incorporates multiple reference signals into the IVA cost function and adaptively controls the effect of inaccurate and accurate references through an adaptive parameter-tuning technique. We study the performance of acIVA on high dimensional datasets, and demonstrate its superior performance in terms of its ability to meaningful patterns from large-scale fMRI datasets. We propose a sliding-window analysis technique using acIVA to extract dynamic functional network connectivity patterns that assume variability in both spatial and temporal domains. We demonstrate the significance of spatial dynamics through a classification technique, which shows an increase in prediction accuracy for spatial dynamic features, and also propose graph-theoretical metrics to quantify the variability in functional connectivity across networks and variability within each spatial network.