Association of Neuroimaging Data with Behavioral Variables: A Class of Multivariate Methods and Their Comparison Using Multi-Task FMRI Data

Abstract

It is becoming increasingly common to collect multiple related neuroimaging datasets either from different modalities or from different tasks and conditions. In addition, we have non-imaging data such as cognitive or behavioral variables, and it is through the association of these two sets of data-neuroimaging and non-neuroimaging-that we can understand and explain the evolution of neural and cognitive processes, and predict outcomes for intervention and treatment. Multiple methods for the joint analysis or fusion of multiple neuroimaging datasets or modalities exist; however, methods for the joint analysis of imaging and non-imaging data are still in their infancy. Current approaches for identifying brain networks related to cognitive assessments are still largely based on simple one-to-one correlation analyses and do not use the cross information available across multiple datasets. This work proposes two approaches based on independent vector analysis (IVA) to jointly analyze the imaging datasets and behavioral variables such that multivariate relationships across imaging data and behavioral features can be identified. The simulation results show that our proposed methods provide better accuracy in identifying associations across imaging and behavioral components than current approaches. With functional magnetic resonance imaging (fMRI) task data collected from 138 healthy controls and 109 patients with schizophrenia, results reveal that the central executive network (CEN) estimated in multiple datasets shows a strong correlation with the behavioral variable that measures working memory, a result that is not identified by traditional approaches. Most of the identified fMRI maps also show significant differences in activations across healthy controls and patients potentially providing a useful signature of mental disorders.

Keywords: ICA; IVA; data fusion; fMRI; neuroimaging; neuropsychology.

Figure 1

Three different ways to identify associations between imaging and behavioral features using (a) one-to-one, (b) one-to-many, and (c) many-to-many association techniques. Here, $\mathrm{r}(·,·)$ is the function that denotes the association between the profiles and behavioral features.

Figure 2

Step 1: IVA step for the fusion of neuroimaging datasets. IVA estimates interpretable components independent within the datasets while maximally correlated with the same index components across the datasets. Subject profiles corresponding to these components represent each subject’s contribution to the components and can be used to identify multivariate associations with behavioral features.

Figure 3

Step 2: Two different techniques to identify the association across imaging and behavioral features. Here, (a) represents the ac-IVA technique where a single behavioral feature is used to constrain imaging features (subjects profiles), and (b) represents the IVA-G approach where all the behavioral features are taken into account to identify the correlated imaging and behavioral features. In both cases, “black dot”s in the ${\mathbf{B}}_k$ matrices represent the identities or indices of the correlated profiles or behavioral features in ${\widehat{\mathbf{A}}}_k^T$ and ${\mathbf{Y}}^{\mathsf{T}}$, which, in turn, represent indices of the components in ${\mathbf{U}}_k$’s associated with behavioral features.

Figure 4

Estimation performance of individual ICAs, jICA, DS-ICA, and IVA for different (a) number of subject and (b) Step-heights. Here, correlated profiles and behavioral features are identified using the one-to-one correlation analysis technique. All the results are averaged over 100 independent runs.

Figure 5

Accuracy performance of correlation, regression, ac-IVA, and IVA-G to identify the profiles correlated with the behavioral features for different (a) numbers of subjects and (b) step-heights. The components and their corresponding profiles are estimated using IVA. Performances of all the methods improve with the number of subjects and step-heights. All the results are averaged over 100 independent runs.

Figure 6

The four behavioral features used in this study are collected using the letter-number sequence subtest of WAIS-III (WAIS), the face recognition subtest of WMS-III (WMS1), and the logical memory test of WMS-III (WMS2) and HVLT (HVLT). The behavioral feature WAIS measures the working memory, WMS1 measures the visual memory, and WMS2 and HVLT measure the verbal memory and learning. All four features show strong group differences ($p<0.05$) between healthy controls and patients with schizophrenia.

Figure 7

FMRI components estimated by the IVA-L-SOS algorithm. Here, the colormap of the components is adjusted so that the colors red, orange and yellow mean higher relative activation in controls and blue means higher relative activation in patients. The estimated components show significant group differences even after strict Bonferroni correction ($(p\times 25)\le 0.05$) in DMN, FP, motor, and auditory cortex areas for AOD and SM tasks; and visual cortex and AG areas for the SIRP task.

Figure 8

Behavioral features and subject profiles identified as correlated by ac-IVA. Here, (a) WAIS is correlated with the profiles associated to motor, RFP, auditory, and DMN components in the AOD task; (b) WMS1 is correlated with the profiles associated with visual and RFP components in the SIRP task; and finally, (c) WMS2 is correlated with the profiles associated to motor, and FP components in the SM task.

Figure 9

Behavioral features and subject profiles identified as correlated by IVA-G. Here, WAIS and WMS1 are correlated with the profiles associated to (a) motor cortex, RFP, and DMN components in AOD task; (b) DMN, visual cortex, and AG components in SIRP task; and (c) DMN, visual, and auditory cortex components in SM task.

Figure 10

Behavioral features and neuroimaging components identified as correlated using one-to-one Pearson correlation technique across (a) AOD, (b) SIRP, and (c) SM task datasets. Here, we are showing p-values of the correlation and use $p<0.05$ as a threshold to identify the significantly correlated behavioral and neuroimaging features.