Improvements in task performance after practice are associated with scale-free dynamics of brain activity

Abstract

Although practicing a task generally benefits later performance on that same task, there are individual differences in practice effects. One avenue to model such differences comes from research showing that brain networks extract functional advantages from operating in the vicinity of criticality, a state in which brain network activity is more scale-free. We hypothesized that higher scale-free signal from fMRI data, measured with the Hurst exponent (H), indicates closer proximity to critical states. We tested whether individuals with higher H during repeated task performance would show greater practice effects. In Study 1, participants performed a dual-n-back task (DNB) twice during MRI (n = 56). In Study 2, we used two runs of n-back task (NBK) data from the Human Connectome Project sample (n = 599). In Study 3, participants performed a word completion task (CAST) across six runs (n = 44). In all three studies, multivariate analysis was used to test whether higher H was related to greater practice-related performance improvement. Supporting our hypothesis, we found patterns of higher H that reliably correlated with greater performance improvement across participants in all three studies. However, the predictive brain regions were distinct, suggesting that the specific spatial H↑ patterns are not task-general.

Keywords: Learning and performance; Practice effects; Scale-free activity; Working memory tasks.

Figure 1.

Left: Schematic of the dynamics of deactivation (λ) of an active node or propagation of activity (γ) to a neighboring node in a simple network exhibiting nonequilibrium phase transition. Right: Simulating propagation and decay in a one-dimensional lattice of nodes shows how different relationships between activation rate (γ) and deactivation rate (λ) results in the network to be in (A) subcritical (λ > γ), (B) critical (λ ≈ γ), or (C) supercritical (λ < γ) states.

Figure 2.

Hypothetical relationship between differences in fMRI Hurst exponent (H) and differences in task performance improvements with practice. The participant shown in red, whose brain activity is more scale-invariant when performing the task compared to their blue counterpart (left panel), is expected to improve their task performance to a greater degree than their counterpart (right panel).

Figure 3.

(A) Performance on the dual n-back task (A′) for all participants in the first and second runs of the DNB. Each line connects a participant’s performance in the first run to their performance in the second run. (B) Average Hurst exponent (H) for all participants in the first DNB run, the video run, and the second DNB run. Each line connects a participant’s average H across the runs. (C) Relationship of adjusted change in performance (adj. ΔA′) with the average H across participants in each of the three fMRI runs (dots are individual participants and each color is a separate run). The intercept of the regression of ΔA′ on initial performance (i.e., run1 A′) is added to adj. ΔA′ in this figure to center the spread around the mean of adj. ΔA′ rather than zero. (D) The primary latent variable from behavioral partial least squares (PLS) relating adj. ΔA′ to parcel-wise H in the DNB experiment shows a predominantly positive H pattern that is significantly positively expressed in all imaging runs, that is, the first DNB run, the video run, and second DNB run. In the left panel, y-axis shows the correlation of the PLS weight with adj. ΔA′ at each run and error bars show 95% confidence intervals as indicated by bootstrapping: *p < .05, **p < .001. All red parcels in the right panel show bootstrap ratio Z_BR values above +3 (total of five parcels) indicating a reliable positive association between H and the contrast in the left panel, i.e., higher H in those brain regions across all three brain imaging runs was related to greater performance improvements adjusting for baseline performance. There are no blue parcels with Z_BR < −3, indicating an exclusively positive direction for the H-to-adj. ΔA′ association. Cross-block covariance (σ_XY) shows the proportion of covariance between the left and right panel explained by this LV, and the p value is calculated from a permutation test for the eigenvalue for this LV.

Figure 4.

(A) Performance accuracy in the 2-back task for the HCP participants across the two NBK runs. Each line connects a participant’s performance in the two runs. (B) Average Hurst exponent (H) for all participants in the two NBK runs. Each line connects a participant’s average H across the runs. (C) Relationship of adjusted change in performance with their average H across participants in each of the two fMRI runs (dots are individual participants and each color is a separate run). The intercept of the regression of Δ Acc. on initial performance (i.e., run1 Acc.) is added to adj. Δ Acc in this figure to center the spread around the mean of adj. Δ Acc rather than zero. (D) The primary latent variable from behavioral PLS relating adj. Δ Accuracy to parcel-wise H in the NBK data shows a predominantly positive H pattern that is significantly positively expressed in the two runs. In the left panel, y-axis shows the correlation of the PLS weight at each run with adj. Δ Accuracy and error bars show 95% confidence intervals as indicated by bootstrapping: *p < .05, **p < .001. All red parcels in the right panel (nine total) show bootstrap ratio Z_BR values above +3 indicating reliable positive H association with the contrast in the left panel. There are no blue parcels with Z_BR < −3, indicating exclusively positive direction for the H-to-adj. Δ performance association. Cross-block covariance (σ_XY) shows the proportion of covariance between the left and right panel explained by this LV, and the p value is calculated from permutation testing for the eigenvalue for this LV.

Figure 5.

(A) Performance on the word completion CAST for all participants across all six runs. Each line connects a participant’s performance in the consecutive runs. (B) Average Hurst exponent (H) for all participants in the six CAST runs. Each line connects a participant’s average H across the runs. (C) Relationship of adjusted change in performance with their average H across participants in each of the six fMRI runs (dots are individual participants and each color is a separate run). The intercept of the regression of Δ Acc. on initial performance (i.e., run1 Acc.) is added to adj. Δ Acc in this figure to center the spread around the mean of adj. Δ Acc rather than zero. (D) The primary latent variable from Behavioral PLS relating adj. Δ Accuracy to parcel-wise H in the CAST experiment shows a predominantly positive H pattern that is significantly positively expressed in all six runs, that is, higher H in those areas was related with improvements in performance. In the left panel, y-axis shows the correlation of the PLS weight at each run with adj. Δ Accuracy and error bars show 95% confidence intervals as indicated by bootstrapping: *p < .05, **p < .001. All red parcels in the right panel (four total) show bootstrap ratio Z_BR values above +3 indicating reliable positive H association with the contrast in the left panel. There are no blue parcels with Z_BR < −3, indicating exclusively positive direction for the H-to-adj. Δ performance association. Cross-block covariance (σ_XY) shows the proportion of covariance between the left and right panel explained by this LV, and the p value is calculated from permutation tests for the eigenvalue for this LV.

Figure 6.

Audio-visual dual n-back task paradigm. In this example of the first three trials in a dual 2-back round, the participant correctly pressed their index finger for a match between the current and 2-back position of blue square, but falsely pressed their middle finger when they should not have (i.e., 2 does not match 9; so the correct response was only an index finger response and not an index finger and middle finger response). This is an example of one hit and one false alarm.