Tracking ongoing cognition in individuals using brief, whole-brain functional connectivity patterns

Abstract

Functional connectivity (FC) patterns in functional MRI exhibit dynamic behavior on the scale of seconds, with rich spatiotemporal structure and limited sets of whole-brain, quasi-stable FC configurations (FC states) recurring across time and subjects. Based on previous evidence linking various aspects of cognition to group-level, minute-to-minute FC changes in localized connections, we hypothesized that whole-brain FC states may reflect the global, orchestrated dynamics of cognitive processing on the scale of seconds. To test this hypothesis, subjects were continuously scanned as they engaged in and transitioned between mental states dictated by tasks. FC states computed within windows as short as 22.5 s permitted robust tracking of cognition in single subjects with near perfect accuracy. Accuracy dropped markedly for subjects with the lowest task performance. Spatially restricting FC information decreased accuracy at short time scales, emphasizing the distributed nature of whole-brain FC dynamics, beyond univariate magnitude changes, as valuable markers of cognition.

Trial registration: ClinicalTrials.gov NCT00001360.

Keywords: classification; cognitive states; connectivity dynamics; fMRI; functional connectivity states.

Conflict of interest statement

The authors declare no conflict of interest.

Figures

Fig. 1.

Experimental paradigm and summary of analysis for multitask scans.

Fig. S1.

Detailed schematic of experimental paradigm and analysis pipeline. (A) Timing of experimental paradigm for the continuous multitask experiments, as well as descriptions of the visuals presented to the subjects during each task. (B) Detailed depiction of how window-based FC connectivity patterns entering the classification analysis were computed. (C) Step-by-step diagram of the analysis pipeline used for the multitask paradigm. Data collected as subjects engaged in and transition between the different tasks were first preprocessed. We then extracted representative time series for the ROIs. These whole-length representative time series entered a PCA analysis used to reduce the dimensionality of the data. Selected PCA time series were then used to construct window-based connectivity patterns that entered the final clustering step. Groupings of windows based on connectivity patterns were finally compared against groupings of windows based on the mental task being performed.

Fig. 2.

Group-level FC state classification results.

Fig. 3.

Individual level FC state classification results. (A) Representative nonoutlier subject (S01). (B–F) Outlier subjects for one or more WLs.

Fig. S2.

Single-subject behavioral results. (A) Percentage of correct responses (PCorrect). (B) Discrepancy in PCorrect across blocks of the same task (a.b.s.t.). (C) Reaction time (RT). (D) Discrepancy in RT a.b.s.t. (E) Percentage of trials with missing responses (PMissing). (F) Discrepancy in PMissing a.b.s.t. In all panels, each subject is represented by four bars: transparent bar, overall value across all tasks; blue, memory; green, math; yellow, video. Subjects are sorted according to the panel’s specific metric. Outliers for more than one WL are marked in red.

Fig. S3.

Individual subject classification results (part 1). Classification results for nonoutlier subjects 2 (A), 4 (B), 6 (C), 7 (D), 9 (E), 10 (F), 11 (G), and 13 (H) are shown here. For each subject, we show classification results for WL = 180 s, 90 s, 45 s, 60 s, 30 s, and 22.5 s in the form of classification staffs. Correctly classified windows are marked with black dots whereas incorrectly classified windows are marked in red. To the right of the classification staffs, we report quantitative measures of classification in terms of classification accuracy and adjusted rand index (ARI). Below the classification staffs, we report values of percent correct responses, percent of missing trials, and response time for each active task block.

Fig. S4.

Individual subject classification results (part 2). Classification results for the remaining nonoutlier subjects: 15 (A), 16 (B), 17 (C), 18 (D), 19 (E), and 20 (F). The organization of results within each panel is the same as in Fig. S3.

Fig. 4.

Behavior versus FC classification (WL = 22.5 s). ARI outliers at multiple WLs marked in red. (A) ARI vs. PCorrect. (B) ARI vs. ΔPCorrect. (C) ARI vs. RT. (D) ARI vs. ΔRT. (E) ARI vs. PMissing. (F) ARI vs. ΔPMissing.

Fig. S5.

Localizer scan analyses. (A) Low order contrast maps for all active tasks (memory vs. rest; math vs. rest; and video vs. rest). Maps are thresholded at PFDR < 0.05. (B) High order contrast maps (task vs. task) for all possible active task pairs (memory vs. math; video vs. math; and video vs. memory) also thresholded at PFDR < 0.05. (C) Maps of ROIs ranked according to their activity-based discriminatory power across tasks, as determined by the F statistic for the task vs. task contrasts. Cooler colors are used for ROIs with the highest rank (lowest F and lowest discriminative power across tasks) whereas warmer colors are used for the ROIs with the lowest rank (highest F and highest discriminative power across tasks)

Fig. 5.

Change in accuracy when ROIs are selectively removed from the analysis for WL = 60 s (Top) and WL = 22.5 s (Bottom). Below the plots we show ROIs entering analyses when 0%, 25%, 50%, or 75% of ROIs are removed. Blue frames denote ROIs entering the analysis when most discriminative ROIs are removed first, and red frames denote the complementary analyses.

Fig. S6.

Group-level classification results for additional analyses with different atlases, levels of kept variance, clustering algorithms, and band-pass filtering criteria. In all panels, the combination of parameters reported in the main analysis is highlighted with a gray background. Bars represent average ARI across subjects; error bars represent SE. (A) Average group-level ARI for all window lengths when the number of ROIs in the atlas changes. Results are shown for versions of the Craddock atlas (26) with 30, 50, 70, 100, 150, 200 and 500 ROIs. (B) Average group-level ARI for all window lengths for the 200 ROI atlas when different levels of variance are kept in the PCA step (100%, 97.5%, 95%, 90%, and 75%). (C) Average group-level ARI for all window lengths for the 200 ROI atlas for two different clustering algorithms: k-means and hierarchical clustering. (D) Average group-level ARI for all window lengths for the 200 ROI atlas for two different band-pass filtering criteria: adaptive filtering based on WL and same filtering for all WLs.

Fig. S7.

Group-level classification results for all control analyses. Bars represent average ARI across subjects, and error bars represent SE. (A) Main analysis results for comparison purposes. (B) Average group-level ARI when classification is attempted after phase randomization of all ROI representative time series. (C) Average group-level ARI when classification is attempted after feature randomization. (D) Average group-level ARI after regressing out the task blocks. (E) Average group-level ARI when the features entering the classification are the ROI’s average signal level, instead of connectivity measures. (F) Average group-level ARI when the features entering the classification are the ROI’s SDs across time, instead of connectivity measures.