On the features of spiking connectivity

1751 

Abstract Submission 

Joshua Faskowitz¹, Javier Gonzalez-Castillo¹, Daniel Handwerker¹, Peter Bandettini¹

¹Section on Functional Imaging Methods, NIMH, Bethesda, MD

Joshua Faskowitz, Ph.D.  
Section on Functional Imaging Methods, NIMH
Bethesda, MD

Javier Gonzalez-Castillo, PhD  
Section on Functional Imaging Methods, NIMH
Bethesda, MD

Daniel Handwerker, PhD  
Section on Functional Imaging Methods, NIMH
Bethesda, MD

Peter Bandettini, Ph.D.  
Section on Functional Imaging Methods, NIMH
Bethesda, MD

The concept of functional connectivity is pervasive in modern fMRI research. Resting state and task-based correlation has been applied in a litany of neuroimaging contexts, to demonstrate how discrete regions of the brain might be in or out of sync, potentially modulated by state [1], task context [4], or even clinical condition. The connectivity measure commonly used for fMRI analysis-Pearson or product-moment correlation-can be mathematically unwrapped into a series of values that record instantaneous similarity [3]. These data, which we call edge time series, record a dynamic pattern that when averaged, equals the time-averaged correlation. A longstanding finding in fMRI connectivity analysis is the identification that punctuated moments in time [7], identified via methods like co-activation maps [5], point processes [9], or high-amplitude edge events [10], contribute disproportionally to the time-averaged connectivity [2], such as modularity. These findings prompt us to further explore the characteristics of these punctuated moments using edge time series, which render connectivity patterns at the same temporal resolution as the collected data. Here, we simplify high-amplitude edge events as binary spikes and compare measures of these spikes to conventionally defined functional connectivity and its derivatives. Further, by examining the length of spikes, we seek to characterize how distinct regions of the brain correlate in different manners.

Resting-state fMRI data (14.4 minutes; 0.72 TR) was obtained from 50 random selected low-motion subjects from the Human Connectome Project. Preprocessed data included steps for motion correction, distortion correction, high-pass filtering, and ICA-FIX noise removal. The data were nuisance regressed further using traces from both CSF and WM masks. Time series were obtained for 400 nodes of the Schaefer parcellation [8] plus 55 subcortical and cerebellar nodes. Edge time series are constructed by taking the element-wise product between two z-score node time series. Enumerating this operation between all pairs of nodes renders the edge time series matrix. Edge times series were converted into spike time series (i.e., binary points) using a threshold of 2 (arbitrary units of co-fluctuation). Spikes were defined as contiguous above-threshold events, meaning that the length of each spike could be binned (short: 0.72-2.88 sec, medium: 2.88-5.76 sec, long > 5.76 sec).

First, a posterior cingulate cortex node was highlighted in a single subject to demonstrate its edge time series with another highly correlated node (also in the DMN) and an uncorrelated node (in the somatomotor system; Fig 1a-e). In this example, we observe that spike counts recapitulate functional connectivity, despite the loss of continuous co-fluctuation information (Spearman's rho: 0.88; Fig 1f). This pattern is replicated when using group-average data (rho: 0.87; Fig 1g). Next, we asked if different systems have different characteristic edge spike durations (Fig 2a). We observed that edges within the visual, somatomotor, and dorsal attention systems displayed the greatest counts of long spiking, whereas short spikes were generally pervasive. However, edges of the DMN to other systems displayed fewer short spikes. Finally, we compared nodes long and short spike counts at the node level (Fig 2b-c). Notably regions along the cingulate emanate short spiking edges and regions in the posterior temporal lobe emanate long spiking edges. This pattern did not relate to the principal gradient, a marker of hierarchical organization [6].

By virtue of unwrapping correlation, edge time series reveal the manner in which correlation values are realized. Some edges are marked by brief intermittent spikes whereas other edges have longer, more infrequent spikes. Future work should examine if distinct spiking profiles correspond to different cognitive processes, and if deviation from normal ranges can prove to be clinically relevant.

Connectivity (eg. functional, effective, structural) ²

fMRI Connectivity and Network Modeling ¹

FUNCTIONAL MRI

Other - connectivity; network; edge time series; spike;