Weaker Brain Dynamics during Sustained Working Memory Task: Perspectives from Co-activation Patterns

Target Audience: Neuroscientists interested in brain resting-state dynamics Introduction: Brain resting-state (RS) is believed to be a varying mixture of arousal, levels of attention, mind wandering, etc., various on-going conscious processes may cause considerable fluctuations in the observed network patterns. We therefore have hypothesized that weaker connectivity dynamics should be observed when subjects are under less variable conscious states, for example, when their attentions are primarily engaged by sustained external tasks. Using sliding-window approach, we have demonstrated less variability of Pearson correlations with respect to the default-mode network (DMN) during working memory (WM) state compared to rest (Figure 1 (a)). Very recently, it was found that conventional correlation analysis may indeed be driven by activity at only a few critical points, and that multiple spatial patterns (referred to as co-activation patterns (CAPs)) could be further obtained via temporal decomposition of these critical frames. These reproducible CAPs reflect a repertoire of brain patterns, and the corresponding temporal information allows us to track state alternations at each individual frame, thus offering a more direct way to unveil neural information carried by the source signals. Here, we attempt to employ CAPs analysis to examine the fundamental changes in brain repertoires that underlie the macroscopic decrease in correlation variations during WM task, as shown by sliding-window analysis. Methods: fMRI Images from 17 healthy subjects were acquired at 3T (GE Signa 750, spiral-in/out sequence, TR=2s). Respiration and cardiac (pulse oximetry) data were recorded using the scanner’s built-in physiological monitoring system. Experiments: Each subject underwent (1) a resting-state scan (8 min, relaxed & closed eyes); (2) a continuous 2-back WM task (8min, inter stimulus interval = 3s). Preprocessing: Physiological noise correction, slice time correction, detrending, spatial smoothing (Gaussian FWHM = 4mm), nuisance regression (six head motion parameters, signals from the white matter and the CSF, subjects’ behavioral data) and < 0.1 Hz low-pass filtering were applied. Datasets were further transformed to z scores and normalized to the MNI template. Brain dynamics with respect to the DMN and executive control network (ECN) were examined here. Sliding-window Correlation Analysis: standard deviations of the sliding-window Pearson correlation (window size = 1min, window step = 4s) sequences with respect to two network ROIs [DMN, MNI(-6, -58, 28), ECN, MNI(34, -54, 46)] were computed and utilized to reflect the macroscopic variability of functional connectivity with DMN/ECN across a single scan; CAPs analysis: (1) frames with ROI signal intensity amongst the top 30% were selected and temporally concatenated; (2) the extracted time frames were further temporally decomposed into multiple CAPs based on their spatial similarity using the k-means clustering method (cluster number ranges from 2 to 16). To account for the instability of single trial clustering result, for each cluster number, we repeated k-means clustering 1000 times, then synthesized the 1000 classification results into one via Normalized-cut method (the weighting between two time points i, j were defined as wi,j=∑ 1{frame i and j are within the same cluster} 1000 1 ); (3) For each cluster number n, only those ‘dominant CAPs’ were preserved for later comparisons (see Figure 2).