Introduction

Transcranial magnetic stimulation (TMS) represents a leading candidate as a circuit-based intervention to treat dysfunctional brain circuits in psychiatric and neurologic conditions1,2,3,4. In 2008, the FDA approved 10 Hz TMS stimulation for the treatment of depression, with initial randomized studies demonstrating 14–32% remission rates2,3,5,6,7. Recent advancements have shown that theta burst stimulation (or TBS), a patterned stimulation protocol involving high frequency (50–200 Hz) bursts spaced at theta rhythm (3–8 Hz), maybe a more effective alternative6,8,9,10,11,12,13,14,15,16. TMS-delivered TBS offers several advantages over the standard 10 Hz protocol, notably a much shorter treatment duration (3 min versus 37.5 min, respectively)6. Moreover, ‘accelerated’ TMS-delivered TBS approaches have been developed, allowing up to ten TMS-delivered TBS treatments to be delivered daily17,18,19. Despite these promising developments20,21,22, TMS clinical trials have shown variable results, with clinical response rates for depression remaining around 50% one month after treatment6,17,23,24. Factors contributing to this treatment heterogeneity include difficulty with selecting the optimal stimulation parameters from the vast search space of possible parameter combinations, such as stimulation intensity, frequency, duration, waveform, and stimulation location, among others, as well as a poor understanding of the neural effects underlying TBS. An improved understanding of how stimulation parameters influence how TBS directly alters brain activity can reveal pathways to improve treatment efficacy.

Recent studies indicate that repetitive TMS not only modulates neural firing11,25,26, but also induces enduring changes in neural activity patterns over several minutes27, suggesting a mechanism of neuroplasticity underlying its clinical effects27,28,29,30. Indeed, the rationale behind employing TBS-patterned stimulation in clinical treatment stems from its demonstrated efficacy in inducing long-term potentiation in slice physiology and rodent models31,32,33,34,35,36,37. Furthermore, the cumulative effect of TMS treatment across multiple sessions has been substantiated by preclinical data in the motor cortex, showing enhanced and prolonged changes in excitability with multiple sessions compared to single sessions13,38,39,40. Moreover, stimulation location is likely to be important for effective treatment, given motor TMS for stroke4 or the left DLPFC TMS for depression41 are needed in achieving specific therapeutic outcomes. Notably, TBS was also applied non-invasively using transcranial temporal inference principles, allowing targeting of the striatum with associated improvement in motor performance42. In summary, inducing neural changes that relate to clinical outcomes may depend considerably on stimulation location, intensity, and timing. However, elucidating these relationships on the neural level with highly resolved spatial and temporal dynamics have been difficult with non-invasive recordings such as EEG and fMRI. We propose uncovering the neural mechanisms underlying TBS-patterned stimulation and its effects on neural activity by using patterned intracranial direct electrical stimulation coupled with high spatiotemporal resolution intracranial brain recordings.

Direct electrical stimulation via intracranial leads has been a mainstay in modulating neural activity to uncover brain function and treat neurological and psychiatric disorders. Studies with deep brain stimulation (DBS) have led to significant progress in optimizing stimulation parameters such as current, frequency, and patterns in a tailored fashion43,44,45,46,47,48,49,50,51,52,53,54,55. This has translated to clinical success, whereby DBS is used for the treatment of neuropsychiatric conditions, notably medication-resistant major depressive disorder and OCD17,56,57. As such, intracranial stimulation coupled with intracranial EEG (iEEG) is emerging as a powerful method to study the mechanisms of TBS. This approach provides anatomically precise information about neuronal populations at a millimeter scale and temporally precise information on neural dynamics at a millisecond scale. Direct electrical TBS-patterned stimulation paired with iEEG measurements have found lasting entrainment of frequency-specific oscillations (in the theta band, 4–8 Hz) after TBS14, TBS-specific induced buildup of beta band coherence in the sensorimotor cortex58, and improvement in memory within a learning and memory task via TBS microstimulation of medial temporal lobe structures10. Crucially, the location of the TBS-specific responses were highly correlated with brain functional connectivity to stimulation14. Notably, the durability of TBS effects appears to correlate with the duration of the stimulation. A single burst of TBS can elicit post-stimulation changes lasting up to one second14 whereas multiple bursts for approximately 3.5 min in the motor cortex produced ongoing effects for approximately 1–2 min58. However, these studies predominantly focused on oscillatory changes, overlooking potential effects on voltage responses both within and across the TBS train (in the interval between bursts) or across stimulation trains (between rounds of TBS bursts). While non-human studies, such as those examining long-term potentiation (LTP)59 have historically examined the temporal evolution of voltage responses, equivalent investigations in humans have been sparse, particularly in brain regions pertinent to psychiatric disorders. Moreover, there remains a paucity of research addressing how neural changes unfold in response to repeated patterns of direct electrical TBS in intracranial human brain recordings60,61,62. Closing these gaps in understanding promises to deepen our grasp of the neural mechanisms underpinning TBS and its implications for neuropsychiatric disorders.

In this study, we evaluated the neural effects of electrically delivered focal TBS using iEEG voltage recordings in participants with refractory epilepsy. We studied how these effects varied as a function of stimulation location, time, and amplitude. Stimulation locations were chosen based on prior studies demonstrating stimulation-induced memory or neuropsychiatric improvements in individuals, including the anterior cingulate cortex (ACC), temporal lobe cortex, ventrolateral prefrontal cortex (VLPFC), and dorsolateral prefrontal cortex (DLPFC)1,63,64,65,66,67,68,69. Given TBS’s purported ability to induce persistent neural changes39,58,70, we hypothesized that TBS triggers discernible signatures of changing voltage responses both within and across repeated trains of intracranial TBS. We further posited that these changes are influenced by stimulation intensity, location specificity, and the baseline functional (resting state), structural, and stimulation-induced effective connectivity of the targeted brain network15,45,52,71,72,73,74,75,76,77,78. Our findings revealed robust voltage responses to single TBS bursts, with these acute responses exhibiting dynamic changes in amplitude both within and across trains of stimulation. These changes included both increasing and decreasing responses over stimulation trains, suggesting a form of short-term plasticity. Notably, the observed response dynamics were dependent on stimulation intensity and the targeted brain network. Furthermore, we demonstrated the predictive ability of baseline connectivity measures in forecasting TBS response patterns. Collectively, these results underscore the capacity of TBS-patterned electrical stimulation to induce region-specific short-term plasticity in the human brain.

Results

Theta-burst stimulation (TBS) induced voltage dynamics were recorded from 10 participants (median age = 27, ranging from 18 to 53 years old; five male and five female; one left handed individual and 9 individuals right-handed; Supplementaary Table 1). Participants were implanted with depth electrodes for clinical seizure monitoring, for a total of 4567 clinically implanted bipolar-referenced channels (Fig. 1A). Cortical sites chosen for TBS included the insula (n = 2 sites), anterior cingulate cortex (ACC, n = 5 sites), dorsolateral prefrontal cortex (DLPFC, n = 4 sites), orbitofrontal cortex (OFC, n = 2 sites), postcentral (n = 1 site), ventrolateral prefrontal cortex (VLPFC, n = 10 sites), and lateral temporal lobe (n = 5 sites) as identified using an automated parcellation algorithm79 (Fig. 1B). Per-region site number variability was due to the fact that site selection was from the original MRI and occurred before automatic parcellation and electrode localization. For the sake of consistency across participants, however, we defined the sites based on the parcellation for analyzes. Across participants, TBS-patterned stimulation was delivered at 29 unique bipolar pairs, with a median of 2.5 bipolar pairs of stimulated electrodes per participant (ranging from 2 to 4 bipolar pairs stimulated per participant; mean ± standard deviation = 2.9 ± 1.0 sites) across all 10 participants at two current amplitude levels (1 and 2 mA) for all participants except one individual (who received only 1 mA TBS; Supplemental Table 1). TBS patterns included five consecutive 200 Hz bursts (50 ms duration) per trial, with each trial (or train) spaced by ~20 seconds (Fig. 1D). For a given stimulation site, we computed functional and effective connectivity before TBS and the neural response to TBS across all channels (Fig. 1D; Methods).

Fig. 1: Experimental paradigm and analysis procedure.
figure 1

A 10 participants were enrolled with a combined total of 4567 bipolar-referenced channels. B Stimulation sites were selected across cortical regions including the anterior cingulate (ACC, 5 sites), the postcentral gyrus (1 site), the dorsolateral prefrontal cortex (DLPFC, 4 sites), the orbitofrontal cortex (OFC, 2 sites), the ventrolateral prefrontal cortex (10 sites), the insular cortex (2 sites), and the lateral temporal cortex (5 sites). C Trials of single pulse electrical stimulation (SPES) and theta-burst stimulation (TBS) were delivered at specific sites during each experimental session, while continuous iEEG was obtained at all other channels. D Schematic representation of analyzes performed. Resting iEEG data (3–10 min) prior to stimulation was used to construct a functional network using low-frequency amplitude and phase coupling. From SPES, cortical-cortical evoked potential (CCEP) mean and duration were quantified through parameterization to estimate stimulation-induced effective connectivity. Lastly, theta burst stimulation (TBS) consisted of ~4 min of 10 stimulation trains, with each train consisting of five theta-frequency bursts, separated by 20 s. Stimulation amplitude was applied first at 1 mA and then at 2 mA sequentially. TBS mean response was defined as the peak-to-trough response post-burst across all bursts (N = 50 bursts). Channels with TBS mean post-burst response above noise threshold are considered significant (TBS + ). In TBS+ channels, successive post-burst responses are analyzed with reference to the train number or the burst number. Repeated-measures ANOVA was used to determine if there was a significant burst number effect (within train plasticity) or a significant train number effect (across train plasticity). Figures in D are for visualization and schematic purposes only. Error bars represent +/- SEM.

Full size image

Theta-burst stimulation evokes consistent post-burst brain responses

In the neural activity following each TBS burst, we observed robust post-burst evoked responses in a subset of channels (Fig. 2A–C). Given that this post-burst TBS response exhibited a clear peak and trough pattern, we computed the peak-to-trough amplitude (Fig. 2B), which was significantly higher than the pre-train baseline (Fig. 2C; TBS + ; 2-samples T-test: t(98) = 31.8, P < 0.001). In contrast, other channels exhibited no response (Fig. 2D–F, see Methods). Quantifying the post-burst TBS response across all stimulation sessions and channels we found that 14.4% (1233/8540) of channels demonstrated a significant post-burst TBS response across both 1 mA and 2 mA current levels (TBS + , N = 10; n = 29 TBS sites; Fig. 2G; 2-samples t-test, P < 0.05 after FDR correction). Note the total channels sampled (n = 8540) is higher than the total number of channels across participants as the total channels sampled reflect the total number of stimulation sessions, since participants may undergo multiple stimulation sessions at different sites and different stimulation intensities (Supplementary Table 1).

Fig. 2: Theta-burst stimulation evokes consistent neural responses that are modulated over time.
figure 2

A, B Single trials and mean voltage trace in a TBS+ channel of (A) a train and (B) a single burst. The yellow shaded region denotes the post-burst quantification window. C Peak-to-trough quantification of the response time window in baseline and post-burst conditions (n = 50 bursts; two-sided two-sample T-test). D, E, F Same as A–C but for a TBS– channel. G Among aggregate of all channels across 10 participants, 1233/8540 (14.4%) of channels were TBS + . H Mean voltage trace of a train of bursts (collapsed across trains) in a TBS+ channel exhibiting within train plasticity. Bursts 1 to 5 are highlighted in different colors. I Mean voltage trace of different bursts within the train. Note in this channel successive bursts qualitatively are larger than the first burst. J ANOVA testing of mean response across five bursts showing within train plasticity (n = 10 trains per burst; two-sided repeated measures ANOVA). K Single-trial voltage traces and (L) mean voltage trace of the post-burst response across trains (collapsed across all bursts in a train). M Mean response across ten burst trains showing across train plasticity (n = 5 bursts per train; two-sided repeated measures ANOVA). N Among the 1233 TBS+ channels, 239 (19.4%) exhibited across train plasticity, 65 (5.2%) within train response plasticity, and 31 (2.5%) both types of plasticity. Error bars represent +/− SEM. For all panels, *denotes P < 0.05. **denotes P < 0.01, ***denotes P < 0.001.

Full size image

Theta-burst stimulation dynamically modulates brain responses with repeated stimulation

To quantify the potential effects of TBS on short-term neuroplasticity, we asked if the post-burst response changed as a function of stimulation presentations. We define LFP plasticity as a significant change in post-TBS response either within train or across trains. In a subset of sites with significant post-burst (TBS + ) responses (Fig. 2H–M), these responses increased with successive bursts (Fig. 2H–J; rANOVA: F(4,9) = 2.8, P = 0.037). We termed this phenomenon within train plasticity (see Methods for details). In parallel, in some locations, these responses changed as a function of the stimulation train (Fig. 2K–M; F(4,9) = 15.0, P < 0.001). We termed this phenomenon across train plasticity (see Methods for details). Across TBS+ channels (those demonstrating a post-burst response), 19.4% (239/1233) exhibited across train plasticity, 5.2% (65/1233) exhibited within train plasticity, and 2.5% (31/1233) exhibited both types of plasticity in the combined 1 mA and 2 mA current levels (Fig. 2N). In the subset of channels exhibiting either form of plasticity, and to better understand the temporal dynamics, we performed k-means clustering in both the within train and across train dimensions. Multiple clustering evaluation criterions converged at three clusters as the optimal number of clusters for within train response dynamics (Supplementary Fig. 1A). Each cluster represented a distinct response pattern across the five bursts (Fig. 3A). Cluster 1 exhibited increasing responses after successive bursts, accounting for 60% of channels exhibiting within train plasticity. Cluster 2 exhibited a rapid decrease in post-burst response which persisted, accounting for 26%. Cluster 3 was characterized by an initial increase and later decrease, accounting for 18% of channels (Fig. 3B). K-means clustering on response dynamics across train dimension did not reach an optimal solution as different criterion diverged from each other (Supplementary Fig. 1B), highlighting high variability of across train response patterns. In the absence of reliable clusters, we sought to understand the proportion of channels which demonstrated either an upwards trend or a downwards trend. We compared the mean post-burst response in the first two trains against the last two trains and identified channels that showed significant difference in either direction of change (see Methods). We identified channels demonstrating clear uptrend or downtrend in the post-burst response across trains (Fig. 3C), which accounted for 21% and 13% of total channels with across train plasticity (Fig. 3D). Note that a large proportion (66%) of channels demonstrated response patterns where the initial and ending set of stimulation trains had similar post-burst responses.

Fig. 3: Post-burst responses exhibit diverse patterns including response facilitation and habituation.
figure 3

A K-means clusters of within-train post-burst responses (n = 65 channels with significant within-train responses). Cluster 1 showed an increasing post-burst response, cluster 2 a decrease in post-burst response, cluster 3 an initial increase and subsequent decrease. B Amongst channels with significant within-train plasticity, 60% were in cluster 1, 26% in cluster 2 and 18% in cluster 3. C K-means clustering criterion did not converge at an optimal cluster number for across-train plasticity dynamics. Up-trending and down-trending post-burst responses were identified by comparing the first two trains and the last two trains (n = 239 channels with significant across-train responses). D 21% of channels with plasticity had a positive trend, 13% had a negative trend, and 66% had a trend that did not differ in initial and final post-burst responses. Error bars reflect the standard error of the mean.

Full size image

Increased stimulation current amplifies TBS-induced plasticity

We next examined how changing the stimulation current intensity influenced TBS responses to determine if there is a dose-dependent effect on either the post-burst response or on within/across train plasticity (Fig. 4A). The post-burst responses were morphologically similar, with the 2 mA condition demonstrating significantly larger responses (Fig. 4B,C; Paired t-test: t(98) = −4.23, P < 0.001). Further, while 1 mA TBS led to a significant post-burst response but no plasticity, 2 mA lead to both within and across-train plasticity (Fig. 4D,E; rANOVA-Within-Train: F(4,9) = 7.46, P < 0.001; rANOVA-Across-Train: F(4,9) = 3.12, P = 0.006). At the group level, 2 mA TBS led to a higher proportion of brain regions with significant post-burst responses (Fig. 5A; 1 mA median: 4.5%; 2 mA median: 20%; signed rank test: Z = −4.6, P < 0.001) as well as stronger post-burst responses (Fig. 5B; 1 mA median: 3.03Z; 2 mA median: 3.31Z; Z = 4.29, P < 0.001). With regards to plasticity, larger amplitude stimulation resulted in a higher proportion of channels that demonstrated either type of plasticity (within or across train, Fig. 5C; 1 mA median: 1.6%; 2 mA median: 3.9%; Z = −4.05, P < 0.001). Specifically, this increase was driven by a higher proportion of channels exhibiting across train plasticity (Fig.5D–F; 1 mA median: 0.7%; 2 mA median: 3.0%; Z = −3.67, P < 0.001). Taken together, these results indicate that higher stimulation current substantially increases cortical response to TBS, both in the proportion of post-TBS responses and dynamic modulation of the post-TBS response.

Fig. 4: Theta-burst stimulation responses are dependent on stimulation dose.
figure 4

A Location of the stimulating and recording channels for B–E. B Mean voltage trace of the post-burst response for 1 mA and 2 mA stimulation (n = 50 bursts per current). C Post-burst response was significantly different between 1 mA and 2 mA stimulation (n = 10 trains per burst; two-sided two-sample T-test). D Mean response across five bursts with significant across burst plasticity noted only for the 2 mA condition (n = 5 bursts per train; two-sided repeated measures ANOVA). E Mean response across ten stimulation trains. For each train, the five bursts are collapsed. Significant across-train plasticity is noted for the 2 mA condition (two-sided repeated measures ANOVA). Error bars represent +/− SEM. For all panels, *denotes P < 0.05. **denotes P < 0.01, ***denotes P < 0.001.

Full size image
Fig. 5: Dose-dependent effects of theta-burst stimulation on a group level.
figure 5

A Proportion of TBS+ channels was significantly higher in the 2 mA but not the 1 mA condition. Each line represents a stimulation session, while each color represents a different participant (n = 29 matched sessions, two-sided signed rank test, applies to A–F). The higher mean post-burst response across channels B and higher proportion of TBS+ channels C with 2 mA stimulation, D Proportion of channels showing within train plasticity was not different between 1 mA and 2 mA stimulation conditions but E proportion of channels showing across train plasticity was significantly higher. F Proportion of channels showing both types of plasticity were not different between 1 mA and 2 mA stimulation conditions. For all boxplots, the center is the median, the box is the interquartile interval (25 and 75 percentile), and the whiskers extend from the box to either the 2.5 percentile or the 97.5 percentile. For all panels, *denotes P < 0.05. **denotes P < 0.01, ***denotes P < 0.001.

Full size image

Spatial distribution of TBS responses depends on stimulation location

We next divided TBS responses based on the brain region stimulated as we hypothesized that different regions would respond differentially. Across participants and combining the 1 mA and 2 mA responses, we observed that TBS produced post-burst responses in distinct regions of the brain depending on the stimulation target (Fig. 6A). For example, DLPFC stimulation resulted in a high proportion of significant post-burst responses in surrounding frontal regions and the cingulate. VLPFC stimulation drove responses similarly in prefrontal regions but also included lateral and mesial inferior frontal regions. Lateral temporal lobe stimulation resulted in primarily temporal and parietal responses. Anterior cingulate stimulation resulted in significant responses in the cingulum, in the prefrontal and in parieto-occipital regions. We subsequently quantified the proportion of channels exhibiting post-burst responses (TBS + ) and types of response plasticity over common anatomic divisions (Fig. 6B). For DLPFC stimulation, the top three regions exhibiting TBS responses were ACC (50% TBS + ; 5% across train plasticity), DLPFC (34% TBS + ; 3% across train plasticity; 1.4% within train plasticity; 1.4% both types of plasticity) and PCC (29% TBS + ; 2% across train plasticity; 5% within train plasticity). For VLPFC stimulation, the top responses were DLPFC (28% TBS + ; 4.4% across train plasticity; 3.6% both plasticity), VLPFC (33% TBS + ; 14% across train plasticity; 1% within train plasticity; 1% both plasticity) and OFC (38% TBS + ; 14% across train plasticity; 1.2% both plasticity). For lateral temporal stimulation, the top responses were parietal cortex (40% TBS + ; 20% across train plasticity), lateral temporal cortex (30% TBS + ; 5.6% across train plasticity; 2.6% within train plasticity; 1.2% both plasticity) and the PCC (23% TBS + ). And for ACC stimulation, top responses were DLPFC (48% TBS + ; 13% across train plasticity; 6.5% within train plasticity; 4.3% both plasticity), PCC (100% TBS + ; 6.3% across train plasticity; 44% within train plasticity) and occipital cortex (52% TBS + ; 6.3% across train plasticity; 4.2% within train plasticity).

Fig. 6: Spatial specificity of theta-burst stimulation response and plasticity by stimulation site.
figure 6

A Surface heatmap and bar chart B showing the percentage of TBS+ local channels for a particular stimulation site within an anatomical region (gray), within train response plasticity (blue), across train plasticity (green) and both types of plasticity (red). The top three sites for each stimulation location are noted with an asterisk (*). Note reciprocal responses in the DLPFC and ACC when stimulated. ACC stimulation elicits widespread TBS responses. DLPFC: dorsolateral prefrontal cortex; VLPFC: ventrolateral prefrontal cortex; VMPFC: ventromedial prefrontal cortex; OFC: orbitofrontal cortex; ACC: anterior cingulate cortex; PCC: posterior cingulate cortex.

Full size image

Structural, effective, and functional connectivity constrain and predict theta-burst responses

Finally, we hypothesized that the underlying brain connectivity shapes the location and spatial extent of post-burst responses and plasticity during TBS15,71. As a proxy for structural connectivity, we measured the ratio of proximity to gray and white matter as well as the distance to the stimulation site for each recording channel15,71,72,73,80. To measure other types of connectivity, phase locking value (PLV) and voltage correlations during the resting state before TBS were calculated to measure functional connectivity while cortico-cortical evoked potential (CCEP) amplitude and duration were measured to evaluate stimulation-induced effective connectivity14,74 (Fig. 7A; Supplementary Fig. 3, see Methods for details). Across participants and stimulation sites, we compared these structural and functional measures in regions with and without significant post-burst responses (TBS+ vs TBS−). We found that TBS+ regions (regions with significant TBS responses following stimulation) tended to reside more in white matter (signed-rank test: Z = 5.6; P < 0.001), have higher functional (low-frequency PLV; Z = −6.2; P < 0.001) and stimulation-induced effective connectivity (CCEP amplitude; Z = −6.36; P < 0.001), and are located closer to the stimulation site (Z = 6.89; P < 0.001; Fig. 7B). Within TBS+ regions, we observed that those regions exhibiting any form of plasticity had higher PLV (Z = 3.29; P < 0.001), CCEP amplitude (Z = −2.12; P = 0.03) and were closer to the stimulation site (Z = 4.22; P = < 0.001; Fig. 7C). In addition, in differentiating types of plasticity (within train, across train or both), PLV (Kruskal–Wallis test; Chisqr = 6.8; P = 0.03), CCEP amplitude (Chisqr = 16.2; P < 0.001) and distance to stimulation site (Chisqr = 17.7; P < 0.001) were significantly different across these channel groups (Fig. 7D). Of note, sites showing across train plasticity were further away from the stimulation site than sites exhibiting within train plasticity (Fig. 7D). We obtained additional baseline metrics including CCEP duration and resting voltage correlations (see Methods). TBS+ channels had longer induced CCEPs after SPES at the same site (Z = −3.3; P < 0.001) and higher voltage correlations (Z = −6.0; P < 0.001; Supplementary Fig. 3). Channels with any type of plasticity had higher voltage correlations (Z = −3.3; P < 0.001) but no difference in CCEP response duration. Lastly, among the types of plasticity, CCEP duration was different across the types of response plasticity (Chisqr = 9.8; P = 0.007), but not voltage correlations (Supplementary Fig. 3).

Fig. 7: Baseline structural, resting, and effective connectivity characteristics underlying TBS-evoked responses and plasticity.
figure 7

A Exemplar brains from a participant depicting variations in four baseline characteristics: (1) gray matter to white matter proximity ratio for a given channel in log-scale, (2) phase locking value (PLV) to the stimulation site for a given channel, (3) CCEP amplitude for a given channel and (4) Euclidean distance to the stimulation site for a given channel. B Differences in baseline characteristics in channels with and without significant TBS response (n = 63 total stimulation sessions in both groups; two-sided signed rank test). C Differences in baseline characteristics in channels with and without response plasticity (n = 60 stimulation sessions with any plasticity responses; two-sided signed rank test). D Differences in baseline characteristics in channels with different types of response plasticity (n = 18 stimulation sessions with any within-train response; n = 48 stimulation sessions with any across-train response; n = 18 stimulation sessions with both types of responses; two-sided signed rank test). Significant differences were observed for PLV, CCEP amplitude and distance to stimulation site, but not for gray matter to white matter proximity ratio. Each line represents a stimulation session, while each color represents a different participant. For all boxplots, the center is the median, the box is the interquartile interval (25 and 75 percentile), and the whiskers extend from the box to either the 2.5 percentile or the 97.5 percentile. For all panels, *denotes P < 0.05. **denotes P < 0.01, ***denotes P < 0.001.

Full size image

We next asked if structural and functional measures can be used to predict the spatial distribution of TBS responses (Fig. 8A–F). We stratified the recording channels into local (< 30 mm) and distant (> 30 mm) groups relative to the stimulation location to take into account the effects of volume conduction and as local vs distant sites may have differential responses71,80. We constructed multivariate logistic regression models with a ten-fold cross-validation scheme using all computed structural and functional measures. We found that local TBS+ channels were predicted more reliably than distant TBS+ channels (Fig. 8A; Local AUC 0.80, 95%CI: 0.79–0.81; Distant AUC 0.74 95% CI: 0.72–0.76). CCEP amplitude had the largest feature coefficient T-statistic for prediction of TBS+ channels (Fig. 8B). Note the T-statistic is positive indicating that larger CCEP amplitude is associated with TBS+ prediction. We also generated additional predictive models using the individual features (Supplementary Fig. 4A). CCEP amplitude outperformed all other features in the prediction of TBS+ channels (Local AUC 0.75, 95%CI: 0.73–0.86; Distant AUC 0.76 95% CI: 0.74–0.79). Limiting the selection to only TBS+ channels, we found that whether or not a channel undergoes plasticity can be predicted with above chance discrimination, whether local or distant (Fig. 8C; Local AUC 0.60 95%CI: 0.58–0.63; Distant AUC 0.59 95% CI: 0.56–0.62). Distance to the stimulation site exhibited the highest feature coefficient T-statistic for predicting plasticity (Fig. 8D). Note the T-statistic is negative for distance indicating that decreasing distance to the stimulation site is associated with plasticity. Distance to the stimulation site remains a leading feature in predicting sites with any plasticity compared to other individual features (Supplementary Fig. 4B; Local AUC 0.62, 95%CI: 0.59–0.66; Distant AUC 0.60 95% CI: 0.56–0.65). Lastly, in the subset of channels that exhibited plasticity, we found that these baseline measures can also be used to classify the type (within vs both types) of plasticity (Fig. 8E; Local AUC 0.59 95%CI: 0.52–0.66; Distant AUC 0.70 95%CI: 0.60–0.79). CCEP amplitude again had the largest feature coefficient T-statistic for predicting the type of plasticity (Fig. 8F). CCEP amplitude also outperforms other features in predicting plasticity when using individual features (Supplementary Fig. 4C; Local AUC 0.60, 95%CI: 0.50–0.71; Distant AUC 0.71 95% CI: 0.63–0.78). Therefore, measures such as distance to the stimulation site and connectivity can be used to predict whether we observe TBS responses as well as short-term plasticity at different time scales.

Fig. 8: Structural, resting and effective connectivity at rest predict sites of TBS-evoked responses and plasticity.
figure 8

A Receiver Operating Characteristic (ROC) curves derived from baseline characteristics used to predict presence of a significant TBS post-burst response at a given channel. The curves are further stratified by channels either closer than or further from 30 mm of the stimulation site. B The mean T-statistic for baseline features used in models to construct the ROC curves. Note CCEP amplitude is the most significant predictor for TBS+ channels (n = 20 model runs, 10-fold cross-validation per each distance cutoff). C ROC curves for prediction of plasticity using only TBS+ channels. D The mean T-statistic for baseline features used in models to construct the ROC curves. Note distance to the stimulation site is the most significant predictor for channels with plasticity (n = 20 model runs, 10-fold cross-validation per each distance cutoff). E ROC curves for prediction of types of plasticity using only channels that demonstrated plasticity. F The mean T-statistic for baseline features used in models to construct the ROC curves. Note CCEP amplitude is the most significant predictor for type of plasticity (n = 20 model runs, 10-fold cross-validation per each distance cutoff). Error bars represent +/− SEM.

Full size image

Discussion

We used direct electrical stimulation (DES) to identify neural responses to spaced, intermittent theta-burst stimulation (TBS) delivered across 29 sites in ten individuals. First, we characterized the effects of TBS by evaluating voltage deflections in response to each burst, focusing on temporal changes that may signify a form of short term plasticity60,61. Second, we identified neuromodulatory effects of TBS resembling short-term plasticity such as facilitation and habituation that varied with dosage and stimulation. Finally, we examined the relationship between these effects to underlying anatomical and functional connectivity.

We initially confirmed the reliable elicitation of evoked responses from the theta-frequency bursts. These post-burst evoked responses were detectable in 14.4% of regions and were increased with higher stimulation currents. Successive application of TBS bursts modulated these evoked responses in amplitude over time, indicating short-term plasticity occurring both within and across trains. Among the regions exhibiting post-burst responses, 27% demonstrated some form of short-term plasticity, with 19% exhibiting across-train changes, 5% within-train changes, and 2.5% experiencing both. Interestingly, across-train but not within-train plasticity was increased with higher stimulation current. Furthermore, leveraging baseline structural and resting functional connectivity profiles enabled accurate prediction of TBS response locations and their changes across bursts and trains, surpassing chance level. TBS responses exhibited high predictability (AUC = 0.75−0.80) whereas changes in TBS responses were less predictable (AUC = 0.60−0.70). Factors such as CCEP amplitude and anatomical distance between the recording and the stimulation sites played important roles in predicting both TBS responses and their changes within and across trains.

Understanding how TBS-patterned direct electrical stimulation induces plasticity in humans, remains critical but limited in humans in vivo, particularly in the frontal lobe. Previous studies in epilepsy patients evaluated the effects of TBS in the sensorimotor cortex, revealing modifications in beta-frequency coherence58. While other iEEG studies have examined TBS-induced oscillatory changes, notably increasing theta band power14,81, the voltage effects of TBS and the dynamics of these neural responses within and across stimulation trains remain unexplored. This is critical as these features – voltage changes after stimulation – underlie neuroplasticity studies in animal models59,60,61,62. Our results indicate the reliable quantification and temporal tracking of evoked response immediately following each high-frequency burst (TBS response). The neuronal origin of this response is likely complex, reflecting an interaction of both monosynaptic and polysynaptic responses52,82. The observation that these post-burst responses dynamically change over time indicates a form of short-term plasticity, potentially linked to underlying neuronal responsiveness or synaptic modifications. Our prior work with 10 Hz direct brain stimulation demonstrated that voltage changes observed during the stimulation protocol itself is highly predictive of subsequent post-stimulation short-term changes (on the scale of minutes) in network connectivity45,83. In this study, we tested how varying different stimulation parameters such as current amplitude and stimulation site impact both the immediate post-burst response as well as changes in the post-burst response (short-term plasticity). We found that stronger TBS (higher amplitude) resulted in broader engagement of cortical regions and drove changes in post-burst responses over time. In addition, each stimulation site exhibited a distinct spatial profile of TBS-responsive brain regions, emphasizing the importance of understanding of how stimulation parameters modulate TBS responses to each burst as well as changes in the TBS response across bursts and trains.

The short-term plasticity observed between bursts (within-train) and across-trains in this study could potentially serve as an acute indicator for optimizing TBS treatment. By leveraging observed indicators in a closed-loop fashion, clinicians could aim to maximize plasticity effects after a single treatment, metaplasticity effects after multiple treatments, and ultimately, clinical outcomes84,85,86. This approach would involve real-time monitoring of neural responses during TBS and dynamic adjustment of stimulation parameters to enhance the desired plasticity effects. However, it is important to note that the within-train and across-train plasticity occurs at two distinct time scales (~0.100 s vs 20 s), which may suggest different underlying mechanisms of induced plasticity. The within-burst plasticity, for instance, may reflect processes such as facilitation and habituation, while the much more variable response changes observed across trains, occurring on the order of tens of seconds, could indicate other plasticity mechanisms59,61,87. Understanding these temporal differences and their associated mechanisms is crucial for developing targeted optimization strategies.

Our observation that repeated TBS trains led to dynamic changes in voltage responses could help explain why repeated sessions of TMS-delivered TBS lead to improved clinical outcomes in depression and other psychiatric disorders88,89,90. This type of state-dependent accumulated brain changes that may result from spaced and repeated stimulation is referred to as metaplasticity and is well-known in non-human models84,85,86. Indeed, intermittent TBS – delivering repeated and spaced TBS trains (as in this study)—has been shown to produce neural changes that could reflect plasticity while continuous TBS (no spacing between trains) did not—an effect which could be NMDA-receptor dependent16,70,91. In other words, intermittent TBS could be more conducive to triggering plastic changes in the human brain.

One consideration when interpreting our findings is that we performed intracranial iTBS using parameters that were aimed to maximize neural change. These parameters are different from the different from the iTBS parameters delivered with TMS when treating medication-resistant depression in several regards. First, TMS-delivered iTBS for depression uses triplets of 50 Hz stimulation for the stimulation bursts, whereas we opted for 200 Hz stimulation. This was motivated by prior studies that have often used high frequency stimulation (> 100 Hz) and have shown consistent neural responses14,15,44,47. Second, while the TMS-delivered iTBS uses an inter-train interval of 8 seconds, we chose an inter-train interval of 20 s to ensure that any responses induced from the stimulation train reliably returned to baseline, minimizing potential signal build-up differences across trains. Third, while the typical TMS-delivered inter-burst interval for iTBS depression treatment is 200 ms (5 Hz), we utilized an inter-burst interval of 167 ms (6 Hz), which is similar but was noted in a prior study to lead to greater post-stimulation-induced theta power14. While intracranial stimulation patterned to match TMS-delivered iTBS for depression treatment may have similar neural effects as those elicited from our chosen iTBS parameters, we cannot exclude that these changes in parameters can lead to unique and independent neural effects. As such, future studies will be necessary to evaluate other TBS parameters including iTBS pattern used for depression treatment (triplet at 50 Hz, 10 bursts/train at 5 Hz, 8 s inter-train interval), as well as other less commonly applied stimulation pattern paradigms such as beta-burst stimulation. Further, as epilepsy is frequently comorbid with depression and other neuropsychiatric disorders, future studies could address the neural effects of iTBS in these sub-populations92,93.

A major finding in this work is that functional, effective, and structural connectivity all contributed to an ability to predict the post-burst response as well as the type of plasticity induced by TBS. This is consistent with multiple past studies which have indicated that underlying connectivity shapes stimulation responses, though few of those studies examine whether this connectivity can predict changes in the brain15,45,52,71,72,73,74,75,76,77. Since TBS responses and even plasticity could be predicted based solely on pre-TBS metrics, this opens up several exciting potential applications. In the future one could select the optimal stimulation target by considering the pre-treatment structural and functional connectivity of a set of possible stimulation sites. The two most predictive features we observed were (1) the distance to the stimulation site and (2) CCEP amplitude. Translating this to the non-invasive space, diffusion-tensor imaging (DTI) could be employed to define anatomical tracts, while single pulse stimulation using TMS and recording with EEG can be used to index stimulation-induced effective connectivity94,95,96,97,98,99. Based on these two parameters, a gradient of how likely the target brain area would undergo plasticity from stimulation at different locations can be constructed, prior to any treatment. Further, these TBS results shown here may be useful in predicting which circuits could be engaged with stimulation in certain regions non-invasively. For instance, DLPFC stimulation largely induced ACC responses with TBS in our study. These results may similarly inform non-invasive targeting of deep brain structures, which is being made possible through use of temporal inference42.

Inherent limitations exist in any iEEG study, such as the clinical constraints of sparse brain tissue sampling with sEEG, limited sampling of stimulation parameters due to time and safety constraints, and the presence of epileptic tissue in patients with epilepsy. To minimize these issues, we followed prior guidelines for iEEG studies100,101 by including participants with different epilepsies and seizure onsets, stimulating outside of epileptic regions, and acquiring data many hours away from seizure windows. However, this study did not include behavioral measures, which are necessary to link plasticity to potential impacts on behavior, such as cognitive function. Incorporating behavioral tasks in future studies will be critical to understand how plasticity-inducing protocols can causally alter human function and for what duration, but these tasks require time and would benefit from an already optimal set of stimulation parameters that can maximize neural change. Another consideration is the lack of sham stimulation in this study. Although all simulations performed were not perceived by the patient, we cannot rule out the possibility of non-specific factors causing changes in neural activity, which will be particularly important when evaluating behavioral effects in the future. We also did not specifically examine the duration of post-stimulation changes, which could be achieved through assessing evoked responses (CCEPs) after single electrical pulses, or by evaluating the voltage or spectral responses in the post-stimulation spontaneous recordings83. Given the time constraint (only minutes before the next stimulation) and the multiparameter design of our study, we did not assess this aspect of TBS. As such, we cannot discount the fact that there may be some carry-over from the initial 1 mA TBS into the 2 mA TBS, although a prior study has shown that the effects of 2–3 min of intracranial TBS typically do not last longer than 2 min58. Finally, stimulation sites in the insula, orbitofrontal cortex, and motor region have few replicates, thus precluding definitive effects of TBS from these regions.

In sum, we observed that direct electrical intermittent TBS induces reliable immediate voltage responses which are modified by the history of stimulation in a predicable subset of brain regions based on functional connectivity. These results offer further insight in how patterned neurostimulation alters neural activity, and how stimulation therapies can be optimized to maximize neuroplasticity and, ultimately, therapeutic effects.

Methods

Ethics statement

All patients voluntarily participated after fully informed consent as monitored by the Massachusetts General Brigham Institutional Review Board covering Massachusetts General Hospital (MGH). Participants were informed that involvement and engagement in the stimulation tests would not alter their clinical treatment in any way, and that they could withdraw at any time without jeopardizing their clinical care. Electrode placement and anatomical localizations were placed for seizure localization determined by a multidisciplinary clinical team purely for clinical indications and that research participation played no role in the decisions for electrode placement.

Human participants and recordings

We recorded intracranial neural activity from 10 participants with intractable epilepsy undergoing evaluation through invasive monitoring. In all cases, participants underwent stereo-electroencephalography, with implantation of multi-contact depth electrodes to locate epileptogenic tissue in relation to the essential cortex (Supplementary Table 1). Depth electrodes (PMT, Chanhassen, MN, USA) with diameter 0.8 and 4–16 platinum/iridium-contacts (electrodes) 1–2.4 mm long with inter-contact spacing ranging from 4–10 mm (median 5 mm) were placed stereotactically, based on clinical indications for seizure localization determined by a multidisciplinary clinical team independent of this research. Following implant, the preoperative T1-weighted MRI was aligned with a postoperative CT using volumetric image coregistration procedures and FreeSurfer scripts79,102,103,104,105,106 (http://surfer.nmr.mgh.harvard.edu). Electrode coordinates were manually determined from the CT in the patients’ native space79,105 and mapped using a surface based electrode labeling algorithm (ELA79,105,106;) and a volume-based electrode volume labeling approach79 that registered each contact to the DKT atlas107. To map the electrode locations to common brain locations in MNI (Montreal Neurological Institute) space, we used MATLAB and Fieldtrip tools (http://www.ru.nl/neuroimaging/fieldtrip)108. Surface representation of quantified measures on a common pial surface was performed using the ECoG/fMRI visualization and data plotting toolbox for Matlab toolbox (https://edden-gerber.github.io/vis_toolbox/; plot_data_on_mesh.m; patch size = 30; overlap method = mean).

In all cases but one, participants had received their normal antiepileptic medications prior to stimulation to minimize the risk of seizure. Recordings used a Blackrock system with FrontEnd amplifiers with a sampling rate of 2 kHz (Blackrock Microsystems, Salt Lake City, UT, USA). Depth recordings were referenced to an EEG electrode placed on skin (C2 vertebra or Cz or mastoid scalp electrode) or a chest EEG surface contact.

For a subset of participants (N = 4), neural activity during single pulse electrical stimulation (SPES, Supplementary Table 1) has been presented in previous publications with different analyzes80,109.

Design of direct electrical stimulation

We applied direct electrical stimulation to bipolar contact pairs across the brain. Tentative stimulation locations were chosen from sites in or near the anterior cingulate cortex (ACC) or lateral temporal lobe cortex, or the ventrolateral prefrontal cortex (VLPFC) and dorsolateral prefrontal cortex (DLPFC) based on MRI colocalization of channels1,63,64,65,66,67,68,69. The gray-matter white matter boundary was specifically targeted when possible as direct electrical stimulation has been shown to engage both local circuits and distant brain networks80. Stimulation locations must also be outside of areas of seizure onset as judged by the monitoring epilepsy clinicians. Sites near the corpus callosum as well as sites which were too medial in the cingulate were not used to avoid potential discomfort from direct dural stimulation. The final TBS site chosen was aided by examining responses to single pulse electrical stimulation (SPES). We performed SPES-induced cortico–cortico evoked potentials (CCEPs) mapping across several cortical regions, including putative TBS sites. A mean 22.1 ± 8.3 STD sites were used for SPES per participant. SPES responsiveness was used to inform the location of TBS delivery. Specifically, we prioritized contact pairs that showed response to SPES outside of the local region, since these reflect highly connected brain regions80. Although we had intended to target the VLPFC, ACC, DLPFC, and lateral temporal lobe, due to contact availability and sites chosen based on initial visual inspection of the coregistered channel locations on the MRI (before automated mapping to parcellated brain regions in each participants’ native space), some selected stimulation sites were ultimately classified to postcentral or OFC or insula on post-hoc automatic parcellation analyzes. As such, final labels of stimulation sites classified were based on the automatic anatomical parcellations.

Following CCEP mapping, we performed TBS. We designed our TBS paradigm based on prior studies to maximize neural change. Specifically, the intracranial iTBS paradigm comprised of five bursts of 200 Hz stimulation across 50 ms with 10 charge-balanced biphasic symmetrical pulses. We chose 200 Hz as the frequency of each burst as it has been shown to produce more consistent responses across brain regions and individuals44,46,47. Bursts were spaced 117 ms apart for an inter-burst frequency of 6 Hz. This was done given a prior study demonstrating significantly increased post-stimulation induced theta power at 6 Hz or higher14. Five sequential bursts were used to keep the per-trial total duration of stimulation closer to 0.5 s and to mimic previous studies14,44. The five bursts together were considered a single ‘train’. Then, ten trains (each with 5 bursts of 200 ms trains, with 10 trains chosen for the sake of time and to be able to stimulate multiple sites per participant) were spaced 20 sec apart with a jitter in timing pulled from a random distribution of time with a maximum of  ± 2 s. The 20-s interval was chosen to ensure that any responses induced across the stimulation trains had reliably returned to baseline.

First, ten trains were performed at a single bipolar pair at 1 mA. Next, usually around a minute later, another ten trains at the same bipolar pair at 2 mA were performed if the monitoring epileptologist allowed it and there was no reported sensation from the participant. In total, 29 total unique sites were stimulated across participants. Five of these bipolar pairs had TBS stimulation at 1 mA only while all other sites had both 1 mA and 2 mA stimulation. Only one participant received solely 1 mA TBS stimulation. All ten participants had both SPES and TBS stimulation testing at the same (s = 25), or neighboring (s = 4) bipolar pairs. Across all ten participants, a median of 2.5 sites were stimulated with TBS, ranging from 2 to 4 bipolar pairs stimulated per participant (mean ± standard deviation = 2.9 ± 1.0 sites, Supplementary Table 1).

A CereStim stimulator (Blackrock Microsystems, Salt Lake City, UT) was used to deliver TBS or SPES stimulation. Current injection and return paths used neighboring electrodes in a bipolar configuration43,55. Stimulation was controlled via a custom CereStim API via MATLAB or a custom C + + code (https://github.com/Center-For-Neurotechnology/CereLAB). For both TBS and SPES, the same 233 µs duration waveforms were used: 90 µs charge-balanced biphasic symmetrical pulses with an interphase interval of 53 µsec at 7 mA between 10 and 25 trials (for SPES)71,76,80,109,110,111,112,113 and TBS at 1 mA and 2 mA. The interval of 53 µs within the biphasic waveform was required as a hardware-limited minimum interval between square pulses with the CereStim stimulator. SPES involved only one set of the charge-balanced biphasic symmetrical pulses spaced 4–5 s and at different locations. Responses were recorded in a total of 4567 clinically implanted bipolar-referenced channels across participants. Channels, in this case, are defined as the bipolar-referenced signal from pairs of electrodes.

A trained electroencephalographer monitored ongoing recordings for epileptiform activity and asked participants if they experienced any sensations. In only one case did a participant report some sensation at which point we stopped stimulation and did not use data from stimulation at that site in our analyzes. No stimulation had to be stopped for clinical reasons such as inducing epileptiform or seizure-like activity. Participants were awake and were aware that they were being stimulated but were blind to the stimulation timing and parameters.

Electrophysiologic analysis

Data analysis was performed using custom analysis code in MATLAB and Fieldtrip (http://www.ru.nl/neuroimaging/fieldtrip)108. Channels with excessive line noise or without clear neural signal (determined by visual inspection) were removed from the analysis. The remaining electrodes were demeaned and bipolar re-referenced relative to the nearest electrode neighbors to generate a signal represented on the channel level101,114,115.

Resting phase synchrony to measure functional connectivity

To estimate functional connectivity through the oscillatory synchrony of two brain regions, we computed phase locking value (PLV) between all possible electrode pairs using FieldTrip (ft_connectivityanalysis)101,108,116. PLV provides a measure of inter-regional synchrony based on phase difference between the paired signals. To calculate PLV, we divided the resting, pre-stimulation period ranging from 2 to 10 min into 2 s epochs. We used a 4th-order Butterworth filter to obtain the analytical signal between 5–13 Hz, and performed Hilbert transform to obtain the instantaneous phase of the signal. The phase difference between the two signals is used to compute PLV. We chose a low-frequency range (5–13 Hz) to assess functional connectivity as it has been previously used to measure functional connectivity in the setting of studying TBS15. As a separate measure of resting connectivity, we also computed the Pearson’s correlation between the voltage trace of pairwise channels in the same 2 s epochs. The correlation was subsequently averaged across the entirety of the resting data recording.

CCEP mapping to measure stimulation-induced effective connectivity

To examine the effective connectivity of the stimulated network, we used responses to SPES to perform corticocortical evoked potential (CCEP) mapping45,46,52,71,74,76,80,83,109,110,111,112,113. Since we were comparing the SPES-induced CCEP responses with the TBS responses, we applied SPES in the same or nearby contact pairs as the TBS stimulation. Only 4 sites had SPES applied in a neighboring pair of sites compared to TBS, whereas the remaining 25 sites had SPES and TBS applied to the same region. A median of 20 SPES trials (range: 4–40) were applied. As stated above, if a site induced a sensation, we stopped stimulation at that site (which occurred in 2 SPES sites). If a site induced sensation with SPES, we did not use that site for TBS. CCEPs from each channel were first epoched from −1000 to 1000 ms. The epoch was subsequently standardized using Z-scores against the pre-CCEP baseline period (−150–1000 ms; Fig. 1D; Panel 2). We used a basis profile curve parameterization approach to quantify the mean CCEP amplitude as well as the mean CCEP duration117. Whereas conventional CCEP quantification can be dependent on the CCEP waveform and selection of the quantification time, this approach utilizes a machine learning framework that allows for general quantification of CCEP responses, irrespective of the shape of the response.

Theta-burst stimulation response and dynamics quantification

To examine cortical responses to TBS, we first aligned the iEEG signal to the start of each theta-burst and epoched the data −1 s to 1 s to the offset of the burst, producing 50 observations (5 theta-bursts per train *10 trains; Fig. 1D Panel 3). Each observation or epoch was subsequently standardized using Z-scores against a pre-train baseline (−2 s to −1.5 s prior to the start of each train). A pre-train baseline period was chosen because the pre-burst time period, which is part of the previous burst’s post-burst response, has not returned to ‘resting’ voltage levels. The post-burst evoked response was quantified by taking the peak-to-trough amplitude from 0.01 s to 0.1 s post-burst. The 10 ms delay was to avoid contamination from the stimulation artifact, which had a rapid drop off by 2 ms when we aligned each epoch to the offset of the burst stimulation. To compare to the post-burst response, a ‘baseline’ response was also quantified, by measuring the peak-to-trough amplitude from −0.1 s to −0.1 s (as a measure of maximum variance) prior to the start of each train. A two-sample t-test was used to compare the post-burst evoked response against the pre-train ‘baseline’ response. The post-burst evoked response was considered significant at an alpha of 0.05, after FDR correction for multiple channels comparison. A channel with a significant post-burst evoked response is abbreviated as TBS + . To evaluate the temporal dynamics of the post-burst evoked response as a function of burst order, we used a repeated-measures ANOVA (Fig. 1D; Panel 3).

To quantify the dynamic nature of the TBS response, we define short-term LFP plasticity (referred to as plasticity in this manuscript) as a significant change in post-TBS response either across TBS bursts within train or across trains. Only channels that demonstrated significant post-burst evoked responses as defined above were considered for analysis of change in the TBS response, indicating plasticity. This thresholding procedure (only using channels that exhibited strong TBS responses at the single burst level), was to ensure the response being studied over time reflects a change in the underlying amplitude of the evoked response, and not due to a drift in a weak post-burst signal. In the ‘burst’ or ‘within-train’ dimension (five consecutive bursts delivered at theta frequency; see above), we fitted a repeated measures ANOVA model whereby bursts 1 to 5 are the repeated measures, and stimulation train 1 to 10 is the predictor variable (Burst Responses ~ Stimulation Trains). A channel was considered to have demonstrated within-train plasticity if the coefficient for the repeated measure was significant at an alpha of 0.05 after FDR correction for multiple channels comparison. Similarly, in the stimulation train dimension, we fitted a repeated measures model whereby stimulation trains 1 to 10 are the repeated measures, and burst order 1 to 5 is the predictor variable (Stimulation Train Responses ~ Burst Order). A channel was considered to have demonstrated across-train plasticity if the coefficient for the repeated measure was significant at an alpha of 0.05 after FDR correction for multiple channels comparison.

Binary classification analysis

To determine if baseline structural and functional features can be used to predict TBS responses on a per-channel basis, we used multivariate logistic regression. Three sets of data were used. (1) To predict spatial distribution of significant post-burst responses, all channels across participants and sessions were pooled, and categorized by presence or absence of a significant post-burst response. (2) To predict presence of TBS-induced plasticity, only channels with significant post-burst responses were selected, and are categorized by presence or absence of plasticity (with-train or across-train plasticity). (3) To predict the type of plasticity, only channels demonstrating TBS-induced plasticity were used, and were categorized as within-train plasticity, across-train plasticity, or both forms of plasticity. Channel subselection was done for plasticity prediction (dataset 2 and 3) since only channels with significant post-burst responses were considered for analysis of plasticity. To control for possible effects of pure volume conduction, we further stratified channels as being local (< 30 mm) or distant (> 30 mm) to the stimulation site. Features used in the classification analysis included both structural and functional metrics. These included proximity to white matter, Euclidean distance to the stimulation site, CCEP amplitude, CCEP duration, PLV, and voltage correlation. Using these five features, we performed logistic regression with ten-fold cross-validation. Receiver operating characteristic (ROC) curves were constructed, and we quantified the area under the curve (AUC) to evaluate model performance. Mean AUC and 95% confidence interval were constructed based on variance from the ten-fold cross validation scheme.

Statistical analysis

We conducted binary group comparisons of the neural data (e.g. comparison of TBS+ versus TBS- channels) utilizing either the Mann–Whitney test or two-sample t-test for independent samples, and the signed rank test or one-sample t-test for paired samples. For analyzes involving more than two groups (e.g. comparison of within train, across train and both modulation types), we employed ANOVA tests or Kruskal Wallis tests. Parametric tests were selected based on the observation of approximate normal distribution within comparison groups. K-means clustering was performed on dynamics of post-burst responses across either the burst or the train dimension. To determine the optimal number of clusters, we evaluated four clustering criteria which includes the Calinskin-Harabasz method, DaviesBouldin method, gap method, and the silhouette method (evalclusters.m; Matlab 2022b). If the evaluation criterions did not converge at an optimal cluster number, we did not perform clustering and instead evaluated the approximate direction of change across channels. For across train response patterns, the criterions did not converge. To assess approximate trends in the channels with across train plasticity, we compared post-burst responses in the first two trains (5 bursts per train; n = 10 bursts) against post-burst responses in the last two trains (n = 10 bursts) using the Mann–Whitney test. Channels demonstrating a positive trend were identified with P < 0.05 and a Mann–Whitney Z-value that was positive, whereas channels with a negative trend were identified with P < 0.05 and a Mann–Whitney Z-value that was negative.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.