Table 1.
Individual characteristics of the 27 patients enrolled in the study.
Fig 1.
Two different classes of stimuli were used for visually cued generation of words: objects, and the control stimuli (scrambled versions of the target stimuli). Images were presented randomly, and subjects were asked to respond by saying, “scrambled” for control stimuli. Object stimuli were also ranked based on word frequency and selectivity (number of possible correct responses). Participants (n = 27) were implanted with sub-dural electrodes (SDEs), and electro-corticographic (ECoG) data were recorded during task performance at 1–2 kHz. ECoG data were fast Fourier transformed (FFT) and filtered in the high gamma band (70–110 Hz). A Hilbert transform was applied, and data were then transformed back using an inverse FFT. This process generates an amplitude envelope of the desired frequency band (here the gamma band). Functional connectivity between any two channels was assessed by correlating their envelopes of activity–the amplitude envelope correlation (AEC). The amplitude envelopes across trials on one channel are correlated with a second channel to estimate the connectivity. Activity between two channels can be correlated at various lags to estimate the directionality of information flow. The dotted line on each graph represents a lag of 0 ms, while connectivity above or below this line is lagged from one region to the next (activity in region A affects region B at a later time point). This allows for the assessment of unidirectional connectivity from A to B and from B to A, along with bi-directional connectivity.
Table 2.
List of all stimuli used in the analysis.
Fig 2.
Grouped reaction times per condition, across subjects (n = 27).
Using the audio recording, each patient’s reaction time was calculated based on the onset of their verbal response. Across the group, object naming had significantly higher latency than scrambled naming (p<0.001, two-sided, paired t-test. High-frequency objects had a significantly shorter latency than low-frequency words (p<0.001), and high-selectivity objects had higher latency than low-selectivity objects (p<0.001).
Fig 3.
Single individual time-frequency graphs of activity during object naming.
SDEs from a single, representative individual were anatomically localized to POr (red), PTr (green), POp (blue), and sCG (yellow). Time-frequency responses during object naming were computed and displayed percent power change relative to pre-stimulus baseline (two sided t-test, p<0.01 FDR corrected). Power changes at the individual level were consistent with those found across the group; including the POr gamma power deactivation and concurrent power increases in PTr, POp, and sCG. Using this same subject, connectivity analysis using the AEC method was also performed for these electrodes. (a) AEC was computed for two SDEs in a single subject over POr (Channel D) and PTr (Channel C). As in the group analysis, the dashed line represents a lag of 0 ms, areas above the dashed line represent activity on Channel C correlating with later activity on Channel A, and regions below the dashed line lag Channel D before Channel C. Confidence intervals were computed using trial reshuffling (contour lines are p<0.05 uncorrected, two-sided, 1000 resamples). (b) PTr and sCG, (c) PTr and POp, (d) POp and sCG, (e) POp and POr, and (f) POr and sCG.
Fig 4.
Group time-frequency activity plots.
(a) SDEs from 22 subjects with recording sites over the peri-sylvian cortex of the left hemisphere were anatomically localized to pars orbitalis (POr), pars triangularis (PTr), pars opercularis (POp), and sub-central gyrus (sCG), and then co-localized on a common brain surface. For both object (b) and scrambled image (c) naming, group time-frequency responses were computed by averaging the amplitude envelopes of gamma power for all SDEs across subjects in each sub-region (percent power change relative to pre-stimulus baseline, two-sided sign test, p<0.01 FDR corrected). During object naming, the first change from baseline was a decrease in high frequency power in POr, followed by concurrent increases in PTr, POp and sCG. The scrambled naming task showed weak activation of PTr, POp and sCG, although the time-points of these activations were similar for both tasks.
Fig 5.
Grouped gamma activation in IFG and sCG during each condition.
Across subjects, gamma (70–110 Hz) power changes were averaged (mean±1SD) in each sub-region for (a) both tasks (object and scrambled naming). Comparisons between tasks and conditions were computed at each time-point (p<0.01, paired t-test, two-sided, FDR corrected). (b) In the comparison of high- or low-selectivity objects, only PTr showed a sustained difference in activation (black bar). However, (c) for high- or low-frequency objects, no regions showed a significant difference in activation.
Fig 6.
Grouped gamma activation of IFG and sCG as a function of word length.
Short items contained 3 or fewer phonemes and long items contained 6 or more phonemes. No significant differences were observed in Gamma (70–110 Hz) power changes in the group (mean±1SD) for short vs. long items (p<0.01, paired t-test, two-sided, FDR corrected).
Fig 7.
Group time-frequency analysis of right hemisphere SDEs.
(a) SDEs from 10 subjects with right-hemisphere coverage were co-localized on a common brain surface as shown previously (Fig 4). Grouped time frequency responses were computed for (b) object and (c) scramble naming (p<0.01 FDR corrected, two-sided sign test). A direct, left vs. right hemisphere comparison was made in the gamma (70–110 Hz) frequency range (p<0.05, unpaired t-test, FDR corrected) for all four regions. During object naming, POr, PTr, and POp were all significantly more active in the left hemisphere (black box). De-activation of left POr was also significantly different. Only two significant differences were noted during scramble naming: right PTr was more active in the comparison (but not against pre-stimulus baseline), and left POp was significantly more active.
Fig 8.
Group functional connectivity of left IFG during object naming.
(a) Grouped PTr and POr connectivity was estimated by averaging Amplitude Envelope Correlations (AEC) calculated for each individual and the averaged across the group (n = 28 total pairs of SDEs, contour lines are p<0.05, two-sided t-test, FDR corrected). The dotted line on each graph represents a lag of 0 ms, while connectivity above or below this line is lagged from one region to the next (activity in region A affects region B at a later time point). See also methods and Fig 1 for further explanation. AEC across a range of lags (-250 to 250 ms, in 10-ms steps) are represented, with a dashed line depicting a lag of 0 ms. The area above the dashed line represents activity in PTr correlating with later activity in POr (directionality: PTr to POr), and the area below represents activity from POr correlating with later activity in PTr (directionality: POr to PTr). Warmer colors represent positive correlations between the two channels, and cooler colors indicate negative correlations. Unidirectional correlations appear as strong correlation either above or below the dotted line, while bidirectional correlations are centered on it. Given four sub-regions (POr, PTr, POp, sCG), there were six possible pairs of connectivity to evaluate: (b) PTr and sCG (n = 39 pairs), (c) PTr and POp (n = 29), (d) POp and sCG (n = 41), (e) POp and POr (n = 23), and (f) POr and sCG (n = 31).
Fig 9.
Group functional connectivity of left IFG during scramble naming.
Using the group AEC method in Fig 8, connectivity plots during scrambled imaging naming were made using the same electrode pairs as before. Overall, connectivity was weaker and shorter in duration during scrambled naming. (a) PTr and POr (n = 28 total pairs of SDEs, p<0.05, two-sided t-test, FDR corrected), (b) PTr and sCG (n = 39 pairs), (c) PTr and POp (n = 29), (d) POp and sCG (n = 41), (e) POp and POr (n = 23), and (f) POr and sCG (n = 31).
Fig 10.
Co-representation of grouped gamma power, connectivity and network states during word production.
(Top) The first change in gamma power during object naming for all subregions in the left IFG show is a decrease in power in POr. This is followed 100 ms later by increases in power in PTr and sCG, and finally POp. (Middle) Four inter-areal interactions were selected to highlight state transitions in the IFG and sCG, two positive correlations: POp and sCG (lag = 0 ms, Fig 8D) and POp to POr (lag = 0 ms, Fig 8C), and two negative correlations: POr to PTr (lag = 150 ms, Fig 8A) and sCG to PTr (lag = 150 ms, Fig 8B). Each interaction represents one lagged time series of the AEC plot (mean across groups ± 1 SD). Time-points corresponding to network-state transitions were identified using a K-means clustering analysis, and distinct clusters are highlighted in different shades of gray. Following stimulus onset, four different time epochs were identified (0 to 300 ms, 300 to 500 ms, 500 to 750 ms, and 750 ms until vocal response). (Bottom) Using changes in gamma power during object naming and the network connectivity, a schematic of network dynamics was synthesized for object generation. These dynamics are shown at baseline and for the four subsequent processing stages, matching the network state transitions shown above. Each node represents the gamma power change for a given region, and the connections between them (edges) depict the inter-areal functional connectivity. Epochs are shown on separate brain surfaces, and connectivity within the epoch is at 0 ms lag. Connectivity that extends between adjoining surfaces is at non-zero (150 ms) lag.
Fig 11.
(a) Functional connectivity between PTr and sCG were compared between high and low lexical frequency across the group (sCG to PTr, lag = 200 ms, mean±2 SD, n = 39 SDE pairs). Each line represents the correlation between sCG and PTr from pre-stimulus baseline through response onset at one lag. Just before articulation, this interaction becomes strongly negative; representing the feedback from sCG to PTr. Given the lagged timeline of correlations, the onset of sCG activity changes during this correlation occur 200ms prior to the time-scale for PTr shown here (b) When this correlation is compared for high frequency and low frequency objects, high frequency objects engage the feedback from sCG at a significantly earlier time-point (<800 ms) than low frequency objects (p<0.001, two-sided, paired t-test). This analysis was repeated for high- and low-selectivity objects. Given the lagged timeline of correlations, the onset of sCG activity changes during this correlation occur 200ms prior to the time-scale for PTr shown here (c) and (d), and demonstrated that the low-selectivity objects also engage the sCG feedback at a significantly earlier time-point.
Fig 12.
Group time-frequency activity plots relative to articulation.
(a) Using the SDEs located over POr, PTr, POp, and sCG, grouped time- frequency plots were recomputed during both object and scrambled naming relative to the onset of vocal response. In each individual, each trial was re- centered on the onset of articulation using the audio trace taken during the ECoG recording. The amplitude envelopes for all SDEs in each sub- region were then averaged across the group (percent power change relative to pre-stimulus baseline, p<0.01 FDR corrected, two-sided sign test). (b) In comparison with the stimulus onset locked analysis (Fig 1), the most prominent gamma power changes occurred in sCG and POp–regions that are most closely associated with motor processing in speech production. Power changes in PTr and POr were diminished in amplitude. (c) During scrambled naming, only sCG and POp had significant power increases relative to baseline.
Fig 13.
Group functional connectivity during object naming relative to vocal response.
Grouped connectivity analysis was recomputed relative to onset of articulation. This carried out by re- centering the data on an individual and trial-by-trial basis. From there, the group connectivity analysis was done as in Fig 3. (a) PTr and POr (n = 28 total pairs of SDEs, p<0.05, two-sided t-test, FDR corrected). AEC across a range of lags (-250 to 250 ms, in 20-ms steps) are represented, with a dashed line depicting a lag of 0 ms. The area above the dashed line represents activity in PTr correlating with later activity in POr (directionality: PTr to POr), and the area below represents activity from POr correlating with later activity in PTr (directionality: POr to PTr). (b) PTr and sCG (n = 39 pairs), (c) PTr and POp (n = 29), (d) POp and sCG (n = 41), (e) POp and POr (n = 23), and (f) POr and sCG (n = 31). Connectivity between POp and sCG, and PTr and POp were robust and bidirectional before articulation. The negative correlation from sCG and PTr was observed beginning around 500 ms before articulation.
Fig 14.
Group functional connectivity during scramble naming relative to vocal response.
Group connectivity analysis relative to onset of articulation was recomputed for the scrambled naming condition. (a) PTr and POr (n = 28 total pairs of SDEs, p<0.05, two-sided t-test, FDR corrected), (b) PTr and sCG (n = 39 pairs), (c) PTr and POp (n = 29), (d) POp and sCG (n = 41), (e) POp and POr (n = 23), and (f) POr and sCG (n = 31). Overall, magnitude of correlation was less also regions for the scrambled vs. the object naming conditions.