Informational connectivity: identifying synchronized discriminability of multi-voxel patterns across the brain

Abstract

The fluctuations in a brain region's activation levels over a functional magnetic resonance imaging (fMRI) time-course are used in functional connectivity (FC) to identify networks with synchronous responses. It is increasingly recognized that multi-voxel activity patterns contain information that cannot be extracted from univariate activation levels. Here we present a novel analysis method that quantifies regions' synchrony in multi-voxel activity pattern discriminability, rather than univariate activation, across a timeseries. We introduce a measure of multi-voxel pattern discriminability at each time-point, which is then used to identify regions that share synchronous time-courses of condition-specific multi-voxel information. This method has the sensitivity and access to distributed information that multi-voxel pattern analysis enjoys, allowing it to be applied to data from conditions not separable by univariate responses. We demonstrate this by analyzing data collected while people viewed four different types of man-made objects (typically not separable by univariate analyses) using both FC and informational connectivity (IC) methods. IC reveals networks of object-processing regions that are not detectable using FC. The IC results support prior findings and hypotheses about object processing. This new method allows investigators to ask questions that are not addressable through typical FC, just as multi-voxel pattern analysis (MVPA) has added new research avenues to those addressable with the general linear model (GLM).

Keywords: MVPA; connectivity; fMRI; method; multivariate; networks; pattern discriminability.

Figure 1

The relationship between Informational Connectivity and other fMRI measures.

Figure 2

Pattern discriminability over time in real data. Top: The underlying basis for the pattern discriminability metric—shown here for the bottle condition in one seed in one subject. The blue line represents each time-point's Fisher z-scored correlation with the training pattern for the correct class. The green lines show the correlation values with mean training patterns for the three other classes. Bottom: Pattern discriminability is calculated by taking the correlation with the correct class's mean training pattern and subtracting the correlation strength of the strongest incorrect class (see text for details). When a time-point's value surpasses zero, it would reflect a classifier successfully predicting that time-point's condition. The arrow shows the corresponding values between the plots.

Figure 3

Significantly connected regions in IC and FC analyses for three of the seeds. A group t-test (p < 0.001 with minimum cluster size from permutation testing) determined significance (described in the “Materials and Methods”). Connectivity strength is displayed between green (lower values) and red (higher values). Each seeds region is shown in blue.

Figure 4

Connectivity strengths before cluster-based thresholding for three of the seeds. The displayed regions have connectivity above zero from the group t-test at p < 0.001 prior to thresholding in cluster-based permutation tests, to visualize sub-threshold connectivity for both methods. Connectivity strength is displayed between green (lower values) and red (higher values). Each seed region is shown in blue.

Figure 5

Venn diagrams of voxels significantly connected to each seed through IC (dark gray) and FC (light gray). Searchlights that overlapped with the relevant seed region have been removed. Here, FC results come from an analysis using the timeseries of searchlights' (rather than voxels') mean values, to give a suitable comparison with the searchlight-based IC results.

Figure 6

Connectivity strengths of all searchlights with a seed in the left fusiform gyrus (present in both the GLM and MVPA searchlight results). The IC and FC results for every brain searchlight are displayed relative to the searchlight's mean univariate activation to the objects and decoding accuracy in a 4-way classification of object-types. Searchlights that overlapped with the seed region have been removed. The FC values reflect the described FC approach, using each searchlight's mean timeseries (rather than each voxel's timeseries) to give a suitable comparison with IC (which reflects information in a searchlight volume). The empty space visible in the top-left octant of the FC graph for searchlights with low response levels (despite high decoding accuracy) highlights connectivity that is inaccessible to univariate FC.

Figure 7

Searchlights with significant informational connectivity to at least one of the three left hemisphere seeds (top) and at least one of the three right hemisphere seeds (bottom), shown against MVPA accuracy and mean functional activation. The green, yellow, and red colors each represent searchlights that are connected with just one seed. Blue points show searchlights that are connected to two seeds and black points show searchlights connected to three seeds. Searchlights overlapping with one of the three seeds regions were removed from each scatterplot.