Fig. 4: Across-participant timepoint decoding accuracy varies with correlation order and cognitive engagement.

a Decoding accuracy as a function of order: PCA. “Order'' (x-axis) refers to the maximum order of dynamic correlations that were available to the classifiers (see Feature weighting and testing). The reported across-participant decoding accuracies are averaged over all kernel shapes and widths (see Identifying robust decoding results). The y-values are displayed relative to chance accuracy (intact: \(\frac{1}{300}\); paragraph: \(\frac{1}{272}\); word: \(\frac{1}{300}\); rest: \(\frac{1}{400}\); these chance accuracies were subtracted from the observed accuracies to obtain the relative accuracies reported on the y-axis). The error ribbons denote 95% confidence intervals of the means across cross-validation folds (i.e., random assignments of participants to the training and test sets). The colors denote the experimental condition. Arrows denote sets of features that yielded reliably higher (upward facing) or lower (downward facing) decoding accuracy than the mean of all other features (via a two-tailed t-test, thresholded at p < 0.05). Figure 5 displays additional comparisons between the decoding accuracies achieved using different sets of neural features. The circled values represent the maximum decoding accuracy within each experimental condition. b Normalized timepoint decoding accuracy as a function of order: PCA. This panel displays the same results as Panel a, but here each curve has been normalized to have a maximum value of 1 and a minimum value of 0 (including the upper and lower bounds of the respective 95% confidence intervals of the mean). Panels a and b used PCA to project each high-dimensional pattern of dynamic correlations onto a lower-dimensional space. c Timepoint decoding accuracy as a function of order: eigenvector centrality. This panel is in the same format as Panel a, but here eigenvector centrality has been used to project the high-dimensional patterns of dynamic correlations onto a lower-dimensional space. d Normalized timepoint decoding accuracy as a function of order: eigenvector centrality. This panel is in the same format as Panel b, but here eigenvector centrality has been used to project the high-dimensional patterns of dynamic correlations onto a lower-dimensional space. See Figs. S1 and S2 for decoding results broken down by kernel shape and width, respectively. Source data are provided as a Source Data file.