Overview of task and data structure.

(A) Battery of language tasks that all culminate in the same speech utterance. (B) Electrode coverage across participants; left hemisphere shown (right hemisphere is included in Supplementary Figure 1) (C) The sequence of the cognitive states of interest in a single trial.

Functional Connectivity analysis for an exemplar participant.

(A) Demonstration of flow of creating the functional connectivity feature and resulting vectors across all states. The values are ordered based on connections between different brain lobes (e.g. Frontal-Frontal, Frontal-Parietal, etc.). The features show clear differences in the connectivity patterns of the cognitive states. (B) Confusion matrix showing high accuracy classification of the connectivity vectors. This matrix was calculated using 5-fold cross-validation on the model for the exemplar participant. (C) Discriminative connections for each cognitive state. Orange lines depict connections whose median absolute deviation of the transformed SVM weights survive a Laplacian tail bound significance threshold; node positions correspond to electrodes.

State-Specific Functional Connectivity patterns across all participants.

(A) Confusion matrix obtained from five-fold cross-validated SVM classification, summed over 42 participants. Diagonal cells show per-class accuracies; off-diagonals show misclassification rates. (B) Distribution of participant-wise classification accuracies. The mean accuracy of 64.4% is well above the 20% chance level for five classes. (C) “Uniqueness ratio” for each cognitive state, comparing the sparsity of connections selected by our discriminative approach with those retained after a magnitude threshold (i.e. strongest connections). This ratio is closer to 1 when the connections are more unique in each state, and closer to 0 when they are overlaping across the cognitive states. Discriminative connections are significantly more state-specific than the strongest connections (average uniqueness ratio = 0.817, 0.273 respectively; two-sided permutation test, p < 0.0001 for all states). (D) Cortical distribution of discriminative connections for each state, across all participants. The patterns align with our current knowledge of involved regions in these states (e.g., STG-frontal links in auditory perception, motor–auditory decoupling in speech production). (E) Cortical distribution of connections selected solely by magnitude thresholding. The widespread, non-specific pattern contrasts with panel D, underscoring the greater specificity of the discriminative approach. (All brain surfaces are shown in left-hemisphere lateral view; right-hemisphere results are analogous and shown in the Supplementary Figure 2.)

Comparison of maps for Speech production using different methods.

(A) Electrode-wise map of high-gamma activity amplitude (70–150 Hz) during speech production, aggregated across 42 participants and averaged across trials and time. (B) Connections whose median Pearson correlation exceeds a p < 0.05 threshold. This magnitude-based connectivity thresholding approach yields a dense, overlapping network. (C) Connections identified by the discriminative SVM transformed weights, resulting in a sparse and state-specific distribution of the connections. (D) Ratio of electrodes identified in each region compared to the total number of electrodes recording in that region, using three different approach. The lower ratios for ‘Active Regions’ and regions involved in ‘Discriminative Connections’ indicate that these methods are more sensitive in identifying relevant electrodes compared to the ‘Strongest Connections’ approach. Stars mark regions where the proportion for discriminative connections differs from active regions (permutation test, p values listed: Superior Frontal p = 0.0114, Inferior Frontal p = 0.0793, Precentral p = 0.0001, Postcentral p = 0.0001, Parietal p = 0.0005, Superior Temporal p = 0.1920, Inferior Temporal p = 0.0319, Fusiform p = 0.0279, Occipital p = 0.0002, Temporal Pole p = 0.2697). (E) Venn diagrams summarizing how identified electrodes are distributed among brain regions.

Relationship between discriminative connectivity and local high-gamma activity.

(A) Histogram of FC values for all discriminative edges (pooled across states). While skewed toward stronger correlations, the distribution includes many weak connections that conventional magnitude thresholds would overlook. (B) The cortical distribution of state-specific discriminative connections. The color represents median functional connectivity values, representing the prominent connectivity patterns that remain stable across the trials of these state. (C) Histogram of high-gamma activity, averaged across trials and time at nodes participating in discriminative connections. (D) The cortical distribution of mean high-gamma activity, only for nodes participating in discriminative connections of each state. (E) Scatter-plot of mean connectivity strength versus mean high-gamma activity for the nodes participating in discriminative connections. A positive trend is present, yet many nodes deviate markedly, indicating regions with high connectivity but low signal activity and vice versa. (F) Connectivity–activity relative distance index projected onto the cortex (pink = nodes with relatively higher connectivity than activity; green = nodes with higher activity than connectivity). This map highlights regions whose functional importance is underestimated by activity-only analyses. (All brain surfaces are shown in left-hemisphere lateral view; right-hemisphere results are analogous and shown in Supplementary Figure 3.)

The spatial electrode coverage across participants, showing both right and left hemisphere.

Functional Connections identified using two methods, for participants with right hemisphere coverage.

(A) Cortical distribution of discriminative connections for each state. Orange lines depict connections whose median absolute deviation of the transformed SVM weights survive a Laplacian tail bound significance threshold; node positions correspond to electrodes. (B) Cortical distribution of connections selected solely by magnitude thresholding. The widespread, non-specific pattern contrasts with panel A, underscoring the greater specificity of the discriminative approach.

Relationship between discriminative connectivity and local high-gamma activity, for participants with right hemisphere coverage.

(A) The cortical distribution of state-specific discriminative connections, visualized on the right lateral view of an average brain. The color represents mean functional connectivity values, representing the prominent connectivity patterns that remain stable across the trials of these state. (B) The cortical distribution of mean high-gamma activity, only for nodes participating in discriminative connections of each state. (C) Connectivity–activity relative distance index projected onto the cortex (pink = nodes with relatively higher connectivity than activity; green = nodes with higher activity than connectivity). This map highlights regions whose functional importance is underestimated by activity-only analyses.