Multi-timepoint pattern analysis: Influence of personality and behavior on decoding context-dependent brain connectivity dynamics

Abstract

Behavioral traits are rarely considered in task-evoked functional magnetic resonance imaging (MRI) studies, yet these traits can affect how an individual engages with the task, and thus lead to heterogeneity in task-evoked brain responses. We aimed to investigate whether interindividual variation in behavior associates with the accuracy of predicting task-evoked changes in the dynamics of functional brain connectivity measured with functional MRI. We developed a novel method called multi-timepoint pattern analysis (MTPA), in which binary logistic regression classifiers were trained to distinguish rest from each of 7 tasks (i.e., social cognition, working memory, language, relational, motor, gambling, emotion) based on functional connectivity dynamics measured in 1,000 healthy adults. We found that connectivity dynamics for multiple pairs of large-scale networks enabled individual classification between task and rest with accuracies exceeding 70%, with the most discriminatory connections relatively unique to each task. Crucially, interindividual variation in classification accuracy significantly associated with several behavioral, cognition and task performance measures. Classification between task and rest was generally more accurate for individuals with higher intelligence and task performance. Additionally, for some of the tasks, classification accuracy improved with lower perceived stress, lower aggression, higher alertness, and greater endurance. We conclude that heterogeneous dynamic adaptations of functional brain networks to changing cognitive demands can be reliably captured as linearly separable patterns by MTPA. Future studies should account for interindividual variation in behavior when investigating context-dependent dynamic functional connectivity.

Keywords: behavior; brain network; dynamic functional connectivity; functional MRI; logistic regression; rest; task.

FIGURE 1

Schematic of the overall workflow for classifying task and rest using multi‐timepoint pattern analysis applied to dynamic functional connectivity. (a) Resting‐state and task‐evoked functional magnetic resonance imaging (MRI) data were analyzed for 7 tasks in more than 1,000 individuals. Dynamic functional connectivity was estimated between pairs of regions comprising an established parcellation atlas for each task. (b) For each pair of regions, a linear logistic regression model was trained to classify each of the seven tasks from rest based on dynamic functional connectivity. Training was performed using 10‐fold cross‐validation on 70% of the participants and classification accuracy was evaluated using an independent test set comprising 30% of the participants. Note that each prediction is accompanied by a corresponding probability estimate, represented by percentages here

FIGURE 2

Accuracy of classifying between tasks and rest using dynamic functional connectivity. Left panel—Matrices of prediction accuracies (148 × 148) prior to downsampling to canonical networks for (a) working memory task, (b) social cognition task, (c) language task, and (d) motor task. Regions are delineated based on the Craddock volumetric atlas, with two regions of the brain stem removed. They are grouped according to canonical networks demarcated by magenta boundaries and indicated by labels. Each yellow square within a matrix represents a region‐to‐region connection whose prediction accuracy exceeded the set threshold. Right panel—Circular graph representations (using toolbox from https://github.com/paul‐kassebaum‐mathworks/circularGraph) of internetwork and intranetwork connections yielding average prediction accuracies exceeding the predefined threshold for four tasks, that is, (a) working memory task, (b) social cognition task, (c) language task, and (d) motor task. Each node represents an intranetwork connection while each edge represents an internetwork connection. The edge weights and edge colors (indicated by color bars) represent out‐of‐sample classification accuracy. Yellow nodes represent intranetwork connections whose dynamics enable prediction with accuracy exceeding 70%, while pink nodes represent intranetwork connections whose dynamics enable weaker prediction at below‐threshold accuracies

FIGURE 3

Similarity between network configurations enabling classification between task and rest using DFC. Similarity was assessed using the Dice similarity index computed between every pair of thresholded 20 × 20 network matrices of prediction accuracies. Note that the other three tasks, that is, relational, emotion, and gambling, were not included in this analysis as they did not produce any network connections that exceeded the prediction accuracy threshold

FIGURE 4

Average accuracy of discriminating between task and rest based on time‐averaged (orange) and dynamic (blue) functional connectivity computed from all the region‐to‐region connections. The mean out‐of‐sample prediction accuracy is shown for each of seven tasks. The standard deviation (SD) across connections is indicated by the vertical line along the top of each bar

FIGURE 5

Qualitative and quantitative representations of the significant associations between behavioral measures and task identifiability measures. 1) Boxplots showing statistically significant correlations (p < .05, |r| > 0.1, Bonferroni corrected), with 75% confidence interval, between individual task identifiability measures and behavioral measures from (a) working memory task, (b) social cognition task, and (c) language task, after controlling for the influence of age and gender. 2) The word clouds at the bottom provide a qualitative representation of the correlations for (a) working memory task, (b) social cognition task, and (c) language task. The size of the word represents correlation strength, and the color represents behavioral category. Note that the font color in the word clouds matches the category colors defined for the boxplots. NT, nontarget trials; T, target trials; WM—working memory