Distinct temporal patterns of beta bursts differentiate successful and failed movement cancellation Authors: Eunseo Oh, Yang Seok Cho, Sung-Phil Kim Published: 2026 ABSTRACT SummaryBeta bursts, transient increases in beta-band activity in the brain, have been implicated in action stopping. However, it remains unclear how their temporal dynamics directly map onto the stop process. By analyzing human electroencephalography (EEG) recordings from the stop-signal task, we found that after the stop signal, frontal beta burst rates increased earlier when movement was successfully canceled than when it was not. Temporal decoding and low-dimensional neural trajectory analyses further corroborated this finding, showing that multi-channel beta burst patterns diverged between stop and go trials earlier for successful than failed stops. While averaging burst rates over a wide window from stop-signal presentation to stop-signal reaction time (SSRT) obscured these dynamics, narrowing window near SSRT uncovered clearer neural signatures of inhibitory control. Notably, traditional beta amplitudes did not capture the timing differences observed in beta bursts. Overall, beta burst dynamics suggest that stop process onset timings may underlie behavioral variability in movement cancellation. Introduction Movement cancellation, the ability to withhold undesired movements before execution, is a fundamental component of inhibitory control, implemented through response inhibition.1,2 A leading theoretical framework for response inhibition is the race model, a competition between two cognitive processes: the “go” and “stop” processes.3,4,5,6,7 The model suggests that movement is successfully canceled when the stop process reaches its threshold before the go process completes. A primary paradigm to test this framework is the stop-signal task (SST).1,7,8 SST includes frequent go and infrequent stop trials where a stop signal is presented after a variable stop-signal delay (SSD) following a go cue. Behavioral measures such as reaction time (RT) and stop-signal RT (SSRT) provide quantitative estimates of the go and stop processes. Neuroscience studies have highlighted the fronto-basal ganglia (FBg) network as the central neural circuit for implementing response inhibition.9,10,11,12 The FBg network connects the frontal cortex (FC), particularly the right inferior FC (rIFC)13,14 and the pre-supplementary motor area (pre-SMA),15 to the subthalamic nucleus (STN) of the basal ganglia.16,17,18,19 This network comprises multiple pathways, including movement initiation-related direct and inhibition-related indirect and hyperdirect pathways. While the indirect pathway involves striatum and globus pallidus, the hyperdirect pathway involves direct signaling from the FC to the STN, which acts as a “brake”14,20,21 on thalamic output and induces inhibition-related neural activity in the primary motor cortex (M1).19,20 Moreover, prior studies have consistently suggested that transmission of inhibitory signals across the FBg network is mediated by beta oscillations (13–30 Hz).10,12,14,18 While the amplitude of beta oscillations traditionally represent motor function, showing event-related desynchronization (ERD) during movement and event-related synchronization (ERS) after movement,22,23 recent studies recognize short-lasting “beta bursts” as a key neural signature of response inhibition.24,25,26,27,28,29,30,31 Frontal beta bursts decrease during movement initiation and increase during movement cancellation, being higher between SSD and SSRT in successful stop than in failed stop or go.27,32 Importantly, the likelihood and timing of beta bursts predict single-trial behavior more accurately than beta power,25,28 indicating that beta bursts better capture inhibitory control. According to the race model, an earlier onset of the stop process increases the likelihood of winning the race, either by gaining a temporal advantage over the go process in the independent race model or by rapidly suppressing the go process before it reaches the threshold in the interactive race model.1,33,34 Thus, the timing of the stop process onset is fundamental to successful movement cancellation. However, previous studies have mainly varied SSD to find the relationship of stop process onset with successful cancellation, leaving unclear what determines success or failure given a certain SSD. Although accumulator-based race models predict behavioral variability through stochastic drift rate fluctuations,5 whether success also depends on stop process onset after the stop signal remains unknown. To address this, it is crucial to examine time-resolved beta bursts from SSD to SSRT. Prior studies reporting higher beta burst rates in successful stops compared to other conditions (e.g., failed stops or correct go [CG]) have typically relied on aggregate measures computed across relatively broad SSD-SSRT windows. Also, not all studies have reported greater beta burst rates in successful stops compared with failed stops within the windows (e.g., see Jana et al.32). Such temporal averaging may conflate earlier engagement with greater overall burst probability, making it difficult to disentangle timing effects from magnitude differences across the FBg networks.27,32 While previous studies have investigated beta burst timing-related indices25,26,28 or P3 onset in event-related potential (ERP) during stopping,31,35,36 the present study extends these findings by directly contrasting time-resolved beta burst rates between successful and failed stops. Here, we investigated how beta burst dynamics differ between successful and failed stops to find neural representations of the stop process. We hypothesized that frontal beta burst rates would increase earlier in successful than in failed stops, as both burst rates and timing have been linked to inhibitory success and response slowing.25,28 We further expected bursts to appear first in the FC, followed by the contralateral motor cortex (MC) based on downstream effects on the M1 through the FBg network. We also compared beta bursts and oscillatory amplitudes to determine which better distinguishes successful from failed stops. By analyzing electroencephalography (EEG) data during SST, we found that beta burst rates increased earlier in successful stops than in failed stops with similar SSD via multiple analysis approaches, originating in the FC and spreading to the MC. Compared with conventional beta amplitudes, beta bursts provided a stronger and more temporally precise neural marker of the stop process. In summary, this study directly tested whether differences in beta burst onset timing distinguish successful from failed stopping by analyzing SSD-locked dynamics rather than focusing solely on burst rates or SSRT-aligned engagement. Our work provides insights into the neural mechanisms that support successful inhibitory control during movement cancellation. Results Behavioral results Forty-four right-handed participants (23 females, mean age 23.0 ± 3.38 years old) performed SST while their EEG was recorded. In go trials, participants responded to the direction of the go cue, whereas in stop trials, they withheld responses in response to a stop signal presented at a variable delay (SSD; Figure 1A) after the go cue. The SST comprised 800 trials (600 go trials and 200 stop trials), with left and right responses equally distributed in random order. The SSD was adjusted using the staircase method8,37 to make the proportions of successful and failed stop trials nearly equal.Figure 1Stop-signal task design and beta burst detection(A) Stop-signal task (SST) paradigm.(B) Beta burst detection procedure. (Top) Beta band (15–29 Hz) time-frequency power map. (Middle) Local maxima of power values on the map exceeding six times the median power were identified as beta bursts. (Bottom) Beta oscillations aligned with beta burst detection.(C) Beta burst rate. Burst rates were estimated across trials using a 50 ms sliding window. The example shows beta burst rates in successful stop from one participant, time locked to SSRT (sky blue line). To illustrate the temporal proximity of beta burst increases to the estimated timing of inhibition (SSRT), SSRT-aligned plots were drawn rather than SSD-locked only in this case. Stop-signal task design and beta burst detection (A) Stop-signal task (SST) paradigm. (B) Beta burst detection procedure. (Top) Beta band (15–29 Hz) time-frequency power map. (Middle) Local maxima of power values on the map exceeding six times the median power were identified as beta bursts. (Bottom) Beta oscillations aligned with beta burst detection. (C) Beta burst rate. Burst rates were estimated across trials using a 50 ms sliding window. The example shows beta burst rates in successful stop from one participant, time locked to SSRT (sky blue line). To illustrate the temporal proximity of beta burst increases to the estimated timing of inhibition (SSRT), SSRT-aligned plots were drawn rather than SSD-locked only in this case. In go trials, all participants completed the task with low omission (mean = 0.004, SD = 0.005) and error rates (mean = 0.003, SD = 0.004), where the omission rate was defined as the ratio of go trials with no response, while the error rate denoted the ratio of go trials with incorrect directional responses. In stop trials, the average probability of responding, p(respond|signal), across participants was 0.495 ± 0.053 (mean ± SD) well aligning with the intended SST design of p(respond|signal) = 0.5. The mean SSD and SSRT were 245 ± 47 and 261 ± 21 ms, respectively, where SSRT indicated the estimated time required to inhibit a response following the stop signal. The RT in the failed stop trials (462 ± 40 ms) was significantly faster than that in the CG trials (521 ± 49 ms) (t(43) = −21.70, p = 8.70 × 10−25, paired t test), as predicted by the race model. There was no significant difference between left and right trials, except for CG RTs (t(43) = −5.17, p = 5.81 × 10−6, paired t test) where participants responded significantly faster in the left go trials. Behavioral results are summarized in Table 1 (individual behavioral results can be found in Table S1) and Figures S1 and S2.Table 1Mean group data for behavioral indices with standard deviationsMeasureaMean ± SDP(respond|signal)0.495 ± 0.053Go omission rate0.004 ± 0.005Go error rate0.003 ± 0.004SSD (ms)245 ± 47 SSDL (ms)246 ± 48 SSDR (ms)245 ± 45SSRT (ms)261 ± 21 SSRTL (ms)258 ± 26 SSRTR (ms)263 ± 24Failed-stop RT (ms)462 ± 40 Failed-stop RTL (ms)462 ± 41 Failed-stop RTR (ms)463 ± 39Correct-go RT (ms)521 ± 49 Correct-go RTL (ms)515 ± 50 Correct-go RTR (ms)527 ± 49aAbbreviations: P(respond|signal), probability of responding; SSD, stop-signal delay on stop trials; SSRT, stop-signal reaction time; Failed-stop RT, reaction time on failed stop-signal trials; Correct-go RT, reaction time on go trials; L, left trials; R, right trials. Table 1Mean group data for behavioral indices with standard deviationsMeasureaMean ± SDP(respond|signal)0.495 ± 0.053Go omission rate0.004 ± 0.005Go error rate0.003 ± 0.004SSD (ms)245 ± 47 SSDL (ms)246 ± 48 SSDR (ms)245 ± 45SSRT (ms)261 ± 21 SSRTL (ms)258 ± 26 SSRTR (ms)263 ± 24Failed-stop RT (ms)462 ± 40 Failed-stop RTL (ms)462 ± 41 Failed-stop RTR (ms)463 ± 39Correct-go RT (ms)521 ± 49 Correct-go RTL (ms)515 ± 50 Correct-go RTR (ms)527 ± 49aAbbreviations: P(respond|signal), probability of responding; SSD, stop-signal delay on stop trials; SSRT, stop-signal reaction time; Failed-stop RT, reaction time on failed stop-signal trials; Correct-go RT, reaction time on go trials; L, left trials; R, right trials. Mean group data for behavioral indices with standard deviations Abbreviations: P(respond|signal), probability of responding; SSD, stop-signal delay on stop trials; SSRT, stop-signal reaction time; Failed-stop RT, reaction time on failed stop-signal trials; Correct-go RT, reaction time on go trials; L, left trials; R, right trials. Temporal dynamics of beta burst rates in relation to movement cancellation In each trial, we detected beta bursts and calculated time-resolved beta burst rates across trials (Figures 1B and 1C, see STAR Methods section and Figure S3 for details). The averaged beta burst rates were normalized using a z-transform computed separately for each behavioral type (correct go [CG], failed stop [FS], and successful stop [SS]). To investigate the temporal dynamics of beta burst rates, we first examined individual temporal dynamics at representative right frontal (F8), frontocentral (Fz), and motor channels (C3 and C4) associated with the FBg network (Figure 2), consistent with previous beta burst studies and EEG topography conventions. Although frontocentral beta burst activity has often been assessed at the FCz channel in high-density recordings, Fz was used here as a frontocentral representative channel in the present montage. We compared beta burst rate analyses using the conventional SSD-SSRT window with analyses in which the interval was subdivided into shorter windows before and after SSRT, a model-derived estimate based on race model assumption, to identify time periods most informative for inhibition (Figure S4). Also, we analyzed multichannel burst patterns in two ways: supervised decoding to determine when stop- and go-related patterns became distinguishable, and unsupervised dimensionality reduction to characterize the evolution of their low-dimensional neural trajectories.Figure 2Temporal profile of normalized beta burst ratesThe grand average z-normalized beta burst rates across participants, separated by response direction (left or right), are depicted for three conditions at selected electrode sites: (A) F8, (B) Fz, (C) C3, and (D) C4. The mean stop-signal reaction time (SSRT) across participants is marked by vertical gray lines. Significant differences between conditions are denoted by colored bars below each graph (p < 0.05, FDR corrected), excluding the first 100 ms after SSD to avoid early perceptual responses. Shadings surrounding the graphs or the SSRT indicate standard error of the mean (SEM). See also Figures S4 and S5. Temporal profile of normalized beta burst rates The grand average z-normalized beta burst rates across participants, separated by response direction (left or right), are depicted for three conditions at selected electrode sites: (A) F8, (B) Fz, (C) C3, and (D) C4. The mean stop-signal reaction time (SSRT) across participants is marked by vertical gray lines. Significant differences between conditions are denoted by colored bars below each graph (p < 0.05, FDR corrected), excluding the first 100 ms after SSD to avoid early perceptual responses. Shadings surrounding the graphs or the SSRT indicate standard error of the mean (SEM). See also Figures S4 and S5. Beta burst rate patterns at the frontal and motor areas The grand averaged beta burst rates across participants in the CG, FS, and SS trials at individual sites are shown in Figure 2. At the right frontal site (F8), beta burst rates decreased in CG, while persisting in SS and FS (Figure 2A). For both left and right movements, the timing at which beta burst rates became higher in SS than in CG was earlier than that between FS and CG. Specifically, for left movements, beta burst rates were higher in SS than in CG in time windows at 100–125 ms (q = 0.047–0.031, Cohen’s d = 0.40–0.44) and 200–225 ms (q = 0.014–0.018, Cohen’s d = 0.48–0.53), while beta burst rates in FS exceeded those in CG at 125 ms (q = 0.014, Cohen’s d = 0.54) and 275–325 ms (q ≤ 0.018, Cohen’s d = 0.51–0.60). For right movements, beta burst rates were higher in SS than in CG at 125–175 ms (q ≤ 0.022, Cohen’s d = 0.47–0.66) and 225–250 ms (q ≤ 0.011, Cohen’s d = 0.53–0.55), while being at 150–250 ms (q ≤ 0.022, Cohen’s d = 0.45–0.64) between FS and CG. No significant difference between SS and FS was found at this site. At the frontocentral site (Fz), beta burst rates decreased in CG but not in SS and FS, similar to the pattern observed at F8 (Figure 2B). However, beta burst rates in SS and FS exhibited greater temporal fluctuation at Fz than at F8. For both left and right movements, again, higher beta burst rates in SS than in CG were observed earlier than those between FS and CG. For left movements, beta burst rates were higher in SS than in CG at 200–375 ms (q ≤ 0.035, Cohen’s d = 0.40–0.59), whereas higher in FS than in CG at 250–375 ms (q ≤ 0.036, Cohen’s d = 0.41–0.64). For right movements, beta burst rates were higher in SS than in CG at 125–225 ms (q ≤ 0.037, Cohen’s d = 0.37–0.71) and 275–425 ms (q ≤ 0.022, Cohen’s d = 0.36–0.67), while the higher beta burst rates in FS compared to CG were shown at 150–200 ms (q ≤ 0.045, Cohen’s d = 0.36–0.38) and 275–425 ms (q ≤ 0.024, Cohen’s d = 0.39–0.85). A significantly higher rate in SS than in FS was also observed at 200 ms (q = 0.03, Cohen’s d = 0.39) for right movements. The results at F8 and Fz revealed that the onset timing of beta burst increases, which reflects the onset of the stop process, indicates the success of movement cancellation. Figures 2C and 2D show the temporal dynamics of beta burst rates at bilateral sensorimotor sites (C3 and C4), time locked to individual SSDs. Consistent with patterns observed at frontal sites, beta burst rates declined during movement in CG trials. Higher beta burst rates were observed in SS and FS than in CG at both ipsilateral and contralateral sites (q < 0.05). Only in SS, beta burst rates transiently increased around SSRT, whereas no such increase was observed in FS. Statistical analysis revealed a significant difference between SS and FS only at contralateral sites near SSRT: For the right movements, a significant difference was found at C3 at 225 ms (q = 0.036, Cohen’s d = 0.35). For left movements, differences were observed at C4 at 125 ms (q = 0.007, Cohen’s d = 0.45) and 225–250 ms (q ≤ 0.011, Cohen’s d = 0.44–0.46). These results suggest that increased beta burst rates in the contralateral MC may serve as a neural signature distinguishing successful from failed movement cancellation. Earlier decoding onset for successful movement cancellation We performed time-resolved single-trial decoding of SSD-locked multichannel beta burst rates into stop or go responses. For left movements, decoding accuracy first significantly exceeded the chance level (50%) at 225 ms in SS and at 325 ms in FS (p < 0.05, false discover rate [FDR] corrected) (Figure 3A). For the first significant cluster, SS showed a mean decoding accuracy of 51.3% (95% confidence interval [CI] [50.5, 52.0], Cohen’s d = 0.51 relative to chance), whereas FS reached 51.0% (95% CI [50.1, 52.0], Cohen’s d = 0.33 relative to chance). Similarly, for right movements, significant decoding accuracy was achieved starting at 275 ms in SS and at 300 ms in FS (p < 0.05). The first significant cluster of SS showed a mean decoding accuracy of 51.5% (95% CI [50.7, 52.4], Cohen’s d = 0.58 relative to chance), while FS reached 51.3% (95% CI [50.2, 52.3], Cohen’s d = 0.38 relative to chance). Although decoding accuracy was only modest, indicating minimal discriminative information,38 these results imply that beta burst patterns distinguishing stop from go responses emerged earlier in successful than in failed stops, consistent with the subject-level findings earlier.Figure 3Temporal decoding of beta burst rates into stop and go trials(A) Subject-averaged temporal decoding accuracy. Shaded areas indicate SEM. Time points with decoding accuracy significantly above chance (horizontal gray dotted line) are marked with dots below the traces (non-parametric permutation test, n = 1,000, FDR corrected). See also Figures S6 and S7.(B) Distributions of decoding onset times for SS and FS across participants, showing earlier decoding onsets for SS (∗∗∗p < 0.001).Multivariate beta burst rates across the entire EEG channels at each time points were decoded to infer stop or go trials via binary classification using linear SVM. Temporal decoding of beta burst rates into stop and go trials (A) Subject-averaged temporal decoding accuracy. Shaded areas indicate SEM. Time points with decoding accuracy significantly above chance (horizontal gray dotted line) are marked with dots below the traces (non-parametric permutation test, n = 1,000, FDR corrected). See also Figures S6 and S7. (B) Distributions of decoding onset times for SS and FS across participants, showing earlier decoding onsets for SS (∗∗∗p < 0.001). Multivariate beta burst rates across the entire EEG channels at each time points were decoded to infer stop or go trials via binary classification using linear SVM. To examine the emergence onset of decodable beta burst patterns in more detail, we conducted a bootstrap resampling (1,000 iterations) in each participant to statistically compare the onset between SS and FS (Figure 3B). For left movements, the median onset in SS was 225 ms, significantly earlier than 250 ms in FS (Z = −8.1, p < 0.001, r = 0.32, Wilcoxon signed-rank test). For right movements, although the median onset values were identical for SS and FS (275 ms), a Wilcoxon signed-rank test revealed a significant difference (Z = −11.5, p < 0.001, r = 0.42), indicating that the paired differences were biased toward earlier SS decoding. Distinct low-dimensional neural representations We examined low-dimensional representations of multichannel beta burst rates in movement cancellation via principal-component analysis (PCA). We projected the full 31-channel data onto a 3D space constructed by the first three principal components (PCs) and inspected the evolution of 3D trajectories (Figure 4A). Initially, the trajectories for SS, FS, and CG overlapped during the period from fixation to early post-go cue. Then, they became distinctive among SS, FS, and CG during an SSD-SSRT interval. In this interval, SS trajectories quickly diverged from CG trajectories, whereas FS trajectories remained closer to CG trajectories and deviated later after a certain delay. After SSRT, CG and FS trajectories converged toward similar endpoints, whereas SS trajectories evolved on a distinct path.Figure 4Low-dimensional trajectories of beta burst rates during the stop-signal task(A) Subject-averaged low-dimensional trajectories of beta burst rates from entire channels. Lower images show the full trial (−300 to 1,000 ms from the go cue), with the SSD-SSRT interval highlighted in bold. Upper images zoom into the SSD-aligned window (0–500 ms).(B) Temporal evolution of plane angles of SS and FS trajectories relative to go, quantified over consecutive 50 ms windows. The gray-shaded time window indicates the example plane illustrated on the right.(C and D) Same analyses as (A) and (B) but limited to seven frontocentral channels. Low-dimensional trajectories of beta burst rates during the stop-signal task (A) Subject-averaged low-dimensional trajectories of beta burst rates from entire channels. Lower images show the full trial (−300 to 1,000 ms from the go cue), with the SSD-SSRT interval highlighted in bold. Upper images zoom into the SSD-aligned window (0–500 ms). (B) Temporal evolution of plane angles of SS and FS trajectories relative to go, quantified over consecutive 50 ms windows. The gray-shaded time window indicates the example plane illustrated on the right. (C and D) Same analyses as (A) and (B) but limited to seven frontocentral channels. We also quantified trajectory differences by computing 3D-plane angles between SS/FS and CG trajectories, after 100 ms post-SSD to exclude early perceptual responses to the cue (Figure 4B). Notably, SS-CG angle approached 90°, indicating an orthogonal relationship shortly after the stop signal. This orthogonal divergence emerged earlier in SS-CG than FS-CG. To further verify whether the low-dimensional representations of movement cancellation were robust to channel selection, we repeated the PCA using beta burst rates from a subset of frontal and motor channels (F8, FC1, FC2, Fz, C3, Cz, and C4) that were selected based on their proximity to the channels used in univariate analysis. The resulting trajectories were similar: SS diverged from CG mainly during the SSD-SSRT interval, whereas FS showed a delayed deviation after SSD and evolved in a similar path to CG after SSRT (Figure 4C). The plane angle also revealed that SS-CG trajectories became orthogonal earlier than FS-CG trajectories, highlighting distinct neural dynamics underlying successful inhibition (Figure 4D). Sequential organization of beta bursts over frontal-motor areas To assess the sequential flow of beta burst rates across frontal-motor sites (F8, Fz, C3: left motor, and C4: right motor), beta burst rates in FS and SS were corrected relative to CG (Figure 5A). This correction removed shared movement-related activity, isolating inhibition-related changes for direct comparison between stop conditions. Overall, these CG-corrected beta burst rates were greater than zero for most time points and channels, as expected. Significant differences between CG-corrected FS and CG-corrected SS (paired t test, p < 0.05) appeared sequentially across regions; for right movements, the median time points exhibiting significant differences appeared at 150 ms in F8, then at 200 ms in Fz, and finally at 225 ms in C3; for left movements, the median time points appeared at 225 ms in F8 and at 238 ms in C4, excluding Fz (Figure 5B).Figure 5Sequential increases in beta burst rates during successful stopping(A) Beta burst rate (BBR) differences between stop (SS/FS) and go trials. Shaded areas represent SEM, and gray vertical lines denote the SSRT. Time points at which differences between SS-CG were significantly larger than those between FS-CG are marked with star (∗) above the traces (paired t test, p < 0.05).(B) Median latencies of significant time points from (A) across channels for left and right movements. The data at motor channels (C3 and C4) are presented only for contralateral sites. Horizontal black lines indicate mean SSRT (gray shading, SEM).(C) Direct differences in beta burst rates between successful and failed stops (SS − FS) are plotted across channels, with shaded areas indicating SEM.(D) Topographical illustration of beta burst rate difference between SS and FS. Sequential increases in beta burst rates during successful stopping (A) Beta burst rate (BBR) differences between stop (SS/FS) and go trials. Shaded areas represent SEM, and gray vertical lines denote the SSRT. Time points at which differences between SS-CG were significantly larger than those between FS-CG are marked with star (∗) above the traces (paired t test, p < 0.05). (B) Median latencies of significant time points from (A) across channels for left and right movements. The data at motor channels (C3 and C4) are presented only for contralateral sites. Horizontal black lines indicate mean SSRT (gray shading, SEM). (C) Direct differences in beta burst rates between successful and failed stops (SS − FS) are plotted across channels, with shaded areas indicating SEM. (D) Topographical illustration of beta burst rate difference between SS and FS. To visualize these contrasts, Figures 5C and 5D illustrates SS-FS differences in beta burst rates across three sites (F8, Fz, and contralateral motor site [C3 or C4]). The peaks in the traces appeared sequentially from F8 through Fz to C3, particularly in right movements, consistent with the median time points derived earlier. These results indicate a sequential propagation of beta burst increases from frontal to motor sites. Comparison between beta bursts and beta amplitudes for the representation of movement cancellation To assess whether beta burst rates better represented movement cancellation than beta amplitudes, we analyzed trial-averaged beta power changes (ERD/S) (Figure 6). At frontal sites (F8 and Fz; Figures 6A and 6B), beta ERD/S showed no timing difference between SS and FS, unlike beta burst rates, which increased earlier in SS (see Figures 2A and 2B). Specifically, at F8, ERS occurred in both SS and FS after SSRT, with significant SS > CG differences emerging no earlier than FS > CG (Figure 6A). Time-frequency maps were also similar between SS and FS (Figure 6E). At Fz, ERS also occurred in both SS and FS, while no temporal difference was observed. Rather, ERS was significantly greater in FS than in SS after the SSRT for both movement directions (Figures 6B and 6F).Figure 6Temporal profiles of beta power during the stop-signal task(A–D) Time-varying subject-averaged beta power changes in the correct-go (CG), failed stop (FS), and successful stop (SS) trials at F8, Fz, C3, and C4. Gray vertical lines indicate SSRT. Significant differences between CG, FS, and SS trials are indicated by colored bars beneath the traces (p < 0.05, FDR corrected). Shadings surrounding the graphs or the SSRT indicate SEM.(E–H) Time-frequency map of beta ERD/S for each of CG, FS, and SS trials, shown separately for left and right trials. Time range is identical to that in (A)–(D).EEG channels presented here include (A and E) F8, (B and F) Fz, (C and G) C3, and (D and H) C4. Temporal profiles of beta power during the stop-signal task (A–D) Time-varying subject-averaged beta power changes in the correct-go (CG), failed stop (FS), and successful stop (SS) trials at F8, Fz, C3, and C4. Gray vertical lines indicate SSRT. Significant differences between CG, FS, and SS trials are indicated by colored bars beneath the traces (p < 0.05, FDR corrected). Shadings surrounding the graphs or the SSRT indicate SEM. (E–H) Time-frequency map of beta ERD/S for each of CG, FS, and SS trials, shown separately for left and right trials. Time range is identical to that in (A)–(D). EEG channels presented here include (A and E) F8, (B and F) Fz, (C and G) C3, and (D and H) C4. At central sites, beta ERD/S also differed from beta burst rates. Although both showed similar patterns, the significant increase around SSRT at contralateral sites evident in beta burst rates (see Figure 2) was absent in beta ERD/S (Figures 6C and 6D). At C4, a significant ERS difference between SS and CG emerged only after SSRT for left movements (Figure 6D), contrasting with the SSRT-aligned beta burst rate peak in SS. Time-frequency maps at C3 and C4 further showed contralateral beta ERD in CG (Figures 6G and 6H). In FS, this ERD pattern resembled CG more than SS, consistent with movement execution. These results indicate that beta burst rates, rather than beta power, more precisely reflect movement cancellation. Discussion Our primary question was whether the beta burst temporal dynamics differentiate successful from failed movement cancellation, expecting earlier frontal increases during successful compared to failed stops. We further examined whether this increase propagated from the FC to the contralateral MC. Finally, we compared the temporal dynamics of beta burst rates with conventional beta oscillatory amplitudes to evaluate whether beta bursts provide a more temporally sensitive marker of movement cancellation. Early increase in beta burst rate as a SS signature We conducted multifaceted analyses of beta burst dynamics, from single-channel to multi-channel levels, all converging on a key finding: earlier beta burst dynamics characterized successful movement cancellation. Consistent with the hypothesis, our results showed that frontal beta bursts increased earlier after SSD in successful than in failed stops (Figure 2). Temporal decoding results also highlighted the role of temporal beta burst patterns in inhibitory control28 (Figure 3) and corroborated this at the single-trial scale, while decoding onset appeared relatively later than in the univariate beta bust analyses, which may partly reflect the integration of distributed activity across all channels rather than restriction to predefined electrodes. The investigation of low-dimensional manifolds demonstrated that neural trajectories of stop diverged from go trajectories earlier in successful stops than in failed ones, forming an orthogonal geometry between go and stop trajectories (Figure 4). These orthogonal trajectories may conceptually parallel the well-known potent-null subspace organization observed in motor cortical neuronal population activity,39 where neuronal states occupy a null space to maintain preparatory activity before transitioning to a potent space for movement execution. Accordingly, the earlier orthogonal divergence in successful stops may reflect a rapid transition of beta burst activity into a null-like inhibitory subspace after the stop signal, functionally decoupling the motor system from downstream output. Such an interpretation suggests that successful inhibition may depend on timely engagement of neural dimensions to suppress motor commands before execution. Past ERP studies have similarly shown that the frontocentral P300 appears earlier in successful stops than in failed ones.10,35,36,40 However, unlike the broader cognitive processes reflected by the P300, beta bursts provide a temporally precise and mechanistically specific marker of movement cancellation. Although the SSD-SSRT interval provides a useful summary measure,27,32,41 successful stopping depended more on beta bursts within finer temporal windows. Evaluating beta burst rates in a 100-ms window preceding SSRT offered a more sensitive probe of movement cancellation than analyses using the broader SSD-SSRT interval (Figure S4). This suggests that broad SSD-SSRT measures may obscure focal dynamics. In line with prior studies,28,42 fine beta burst dynamics serve as a sensitive marker of movement cancellation. Neural evidence for the behavioral race model Previous studies reported that beta burst rates were higher in successful than in failed stops within the SSD-SSRT interval,26,27,32 implying that failed inhibition may result from weaker neural control signals. Our results, however, rule out this possibility, showing that success depends on the timing rather than the magnitude of beta burst rate increases (in fact, the height was similar between SS and FS in Figure 2). Moreover, earlier burst onset in successful stops would naturally elevate the aggregate rate within the SSD-SSRT interval, consistent with prior findings.26,27,32 In the race model, earlier stop-signal presentation increases stopping probability, yet this framework does not explain what determines success or failure at a certain fixed SSD.1,7 Extended race models, such as the interactive race or trigger-failure, account for this variability through stochastic drift rate or inhibition strength.3,4,33,43 Consequently, these still explain behavioral variability (SS or FS) with the assumption of fixed onset of the stop process after the stop signal. In contrast, the threshold model links beta bursts near SSRT to transient increases of thresholds.42 Similarly, the pause-then-cancel (PTC) model relates beta burst onset to the transition from pause to cancel phases.34,44 However, no behavioral model explicitly predicts that the variability of stop process onset after the stop signal, which is reflected by the variability of beta burst onset, would determine behavioral variability between successful and failed stops. Our findings suggest that the beta burst onset timing, rather than overall inhibitory strength alone, is critical for successful inhibition. Conceptually, such variability in timing could be instantiated in the existing race models by jittering an onset parameter governing when the stop-related accumulator begins. While our results emphasize the importance of onset timing, a trial-by-trial variability in the drift rate may also capture fluctuations in parallel cognitive processes such as attentional state or preparatory control that can influence stopping outcomes. Incorporating these mechanisms may allow behavioral models to better align with emerging neural evidence. Potential source of stop process onset variability What determines the variability in stop process onset timing remains unclear. The PTC model34,44,45 posits two sequential stages: a rapid, salience-driven pause process via the hyperdirect pathway (right inferior frontal gyrus [rIFG]-STN)34,46 and a later cancel process via the indirect pathway (pre-SMA-striatum-GPi) that terminates movement drive.47,48 Within this framework, beta bursts may reflect different aspects of the stopping-related process depending on their timing and cortical region.44 In the present study, we observed beta burst increases and neural trajectory divergence before or shortly after (∼100 ms) the stop signal in some trials, preceding the canonical window of reactive inhibition. These early effects may align with prior studies reporting that proactive inhibition manifests as slowed go RTs,27,49 which might include other parallel cognitive processes such as attentional engagement.25,45,50 Moreover, because beta burst activity decreases during go trials and increases during stop trials, the observed beta burst changes may reflect the combined outcome of interactions between go and stop processes rather than a pure index of either process alone. Given this perspective, variabilities in proactive inhibition and go process strength could shape subsequent stopping-related dynamics. Nevertheless, the consistent association between earlier onset and SS underscores the importance of understanding the mechanisms that give rise to this temporal variability. Future work should therefore examine links between frontal beta burst onset and proactive control, go-stop interactions, and attentional neural markers such as alpha oscillations or P300 amplitude.51,52,53,54 In addition, variability in the initiation of the stop process, as proposed in the trigger failure model, may obscure neural distinctions between successful and failed stopping.43,55 Integrating neurophysiological markers with computational models of inhibitory control will be essential to disentangle these possibilities and to more precisely characterize the neural determinants of movement cancellation. Spatial downstream of the beta burst rate increase We expected beta burst increases to follow a spatiotemporal progression from FC (F8 and Fz) to contralateral MC (C3 and C4). Consistent with this, we observed a transient increase in beta burst rates in the contralateral MC around SSRT during successful but not failed stops, suggesting that this brief rise would be critical for successful inhibition (Figures 2C, 2D, and 5). In Figure 5, we observed frontal activity peaking ∼200 ms after the stop signal and preceding MC bursts by ∼25 ms. This finding is consistent with prior reports showing maximal beta burst activity for successful stops within the 175–225 ms window and prefrontal-to-motor delays of 20–40 ms.12,27,32,41 Within the FC, the relative roles of rIFG and pre-SMA remain debated21: some studies suggested that rIFG is the key functional region,13,16 whereas others propose that stopping is initiated by pre-SMA, with rIFG supporting context-specific task rule encoding or attentional control.15,47 A magnetoencephalography (MEG) study has directly examined their causal relationship and found that rIFG beta oscillatory activity precedes pre-SMA activity.56 In the present study, beta burst activity over right frontal electrodes tended to emerge earlier than that over frontocentral electrodes. This temporal pattern is broadly consistent with prior findings and extends them by demonstrating that similar ordering is evident at the level of transient burst dynamics. Moreover, the right frontal channel showed sustained high beta burst rates in SS compared with CG, consistent with its role in preventing movement initiation,13,21,57 whereas the frontocentral channel exhibited a transient increase likely driving the stop process.15,47,57 While EEG’s limited spatial resolution precludes these circuit-level conclusions, these distinct temporal dynamics between frontal sites suggest that spatiotemporal patterns of beta bursts may serve as a critical basis for movement cancellation. Comparison with beta oscillation power Finally, burst-based measures more clearly distinguished successful from failed stops than oscillatory power (Figures 2 and 6), indicating that beta bursts capture temporally precise inhibitory signals often obscured in power estimates. Beta amplitudes were higher in failed than successful stops at the frontocentral channel, Fz. This difference may stem from two factors. First, the baseline period included early post-go cue activity, during which proactive inhibition likely elevated beta activity in successful stops, reducing apparent post-signal amplitudes after baseline correction. Early motor beta bursts have been linked to tonic inhibition that facilitates stopping and typically follow frontocentral bursts.27,49 Second, failed stops recruited broader beta frequency range, extending into higher beta (Figures 6E–6H). Previous studies have shown that high beta (>20 Hz) is associated with attentional engagement, whereas low beta (≤20 Hz) is more directly related to movement.58,59,60,61 Our data showed that low-beta activity predominated in successful stops, while both low- and high-beta activities co-occurred in failed stops. The pronounced frontal high-beta activity near SSRT in failed stops (Figures 6E and 6F) may reflect compensatory attentional responses for the error when stop process is initiated too late. Moreover, at contralateral motor sites, high-beta power decreased more strongly during the SSD-SSRT period in failed stops, indicating reduced focal attention (Figures 6G and 6H). Together, these findings suggest that attentional deficits are reflected in high-beta dynamics. Asymmetric inhibitory control between dominant and non-dominant hands In the temporal dynamics of beta burst rates, the distinction between SS and CG trials was more pronounced for dominant- (right) than non-dominant (left) hand movements, showing more distinct onset of increases in beta burst rates (Figure 2), more focal windowing at motor cortical sites (Figure S4), more clearly defined fronto-motor propagation pattern (Figure 5), and clearer differentiation in the low-dimensional trajectories (Figure 4). This lateralized pattern suggests that inhibitory control may be affected by hand dominance, with the dominant hemisphere engaging inhibitory processes more efficiently. Prior large-scale evidence has shown that handedness is linked to functional lateralization in motor and associative cortices, rather than structural asymmetries.62 Similarly, our result may indicate that the dominant hand engages inhibitory-related neural activity earlier and within a narrower temporal window. In contrast, some multivariate analysis showed relatively earlier onset for left-hand trials, suggesting that distributed patterns across all channels may capture sustained or spatially broader inhibitory dynamics that are not fully reflected in focal univariate burst measures. For example, in our data, non-dominant movements involve a more sustained inhibitory state according to proactive beta burst increases in the contralateral motor channel (Figures 2 and S4). However, these asymmetrical findings should be interpreted cautiously and cannot yet be generalized beyond right-handed individuals. Implications Our findings establish temporal beta burst dynamics as the neural signature of inhibitory control of movement cancellation. By revealing that their early onset and spatiotemporal organization determine successful movement cancellation, this work suggests the need for refining current behavioral models of movement cancellation and highlighting the importance of timing in behavioral inhibition. In addition, the temporal specificity of beta bursts may also be clinically relevant in conditions with impaired inhibitory control, such as Parkinson’s disease, attention-deficit hyperactivity disorder (ADHD), and impulse-control disorders,63,64,65 where abnormal beta burst dynamics could provide a biomarker for diagnosis for translational applications. Limitations of the study Several limitations of the present study should be acknowledged. First, the limited spatial resolution of scalp EEG constrains the dissociation of neighboring frontal regions such as rIFG and pre-SMA. Although our results suggest sequential progression from right frontal to frontocentral and then to motor area, this interpretation requires caution. In the multivariate analyses, PCA using frontocentral channels revealed clearer condition separation, but temporal decoding did not yield an earlier decoding onset for successful stops when restricted to these channels. Hence, both the interpretation of sequential progression and the identification of optimal channel sets should be validated using methods with higher spatial precision, such as MEG, intracranial recordings, or functional magnetic resonance Imaging (fMRI). Second, our decoding and temporal analyses are correlational. In single-trial temporal decoding, the accuracy showed barely exceeding chance level, reflecting the inherent difficulty of classifying sparse and transient burst events within short time windows, and therefore should be interpreted cautiously. Nevertheless, to our knowledge, no EEG study has performed single-trial temporal decoding of beta bursts in inhibitory control. Most comparable inhibitory control studies have instead decoded ERPs and reported the single-trial decoding accuracy in the range of area under the curve (AUC) 0.5–0.72.45,66 Compared to these studies, our decoding accuracy was in the range of AUC 0.52–0.54. This gap in decoding accuracy might be due to methodological and feature-related differences. Specifically, beta burst events yield sparser feature spaces than ERP-based signals, making decoding more challenging. Moreover, prior studies utilized a larger number of channels (e.g., 63 channels45) or performed single-trial classification without time-resolved temporal decoding.66 However, despite the limited discriminative strength of our results, the consistent temporal ordering observed across analyses suggests that decoding captures meaningful aspects of burst timing dynamics and supports the feasibility of beta burst decoding at the single-trial level. Moreover, the identification of beta burst onset timing was defined relative to CG trials so that beta burst timing alone may be insufficient to determine stopping outcomes, considering stop trials in isolation. Thus, the causal role of beta bursts in implementing inhibition remains to be established. Combining non-invasive recordings with perturbation methods (e.g., transcranial magnetic stimulation) or invasive approaches in patients, alongside refined behavioral modeling, will be crucial to clarify whether and how beta bursts directly influence successful movement cancellation. Furthermore, although beta bursts provide a temporally precise neural signature associated with movement cancellation, the underlying cognitive mechanisms they reflect remain incompletely understood,26,67 and their trial-level specificity continues to be debated.68 Finally, further work is needed to establish the generalizability and the scope of the measurements. While we focused on beta burst rate to characterize temporal dynamics and compared it with conventional beta power, some studies have employed other burst metrics, such as burst volume,29,30,31 which integrates burst timing and power amplitude to distinguish stopping outcomes. For example, burst volume has been shown to differentiate successful versus failed stopping,29 ignoring versus stopping to an additional signal,30 and stopping versus changing to an additional signal.31 Incorporating multiple burst metrics in future studies may provide a more comprehensive characterization of the neural dynamics underlying inhibitory control. Furthermore, although we identified hand-dominance effects in contralateral MC, all participants were right handed. Future studies should examine whether these temporal burst dynamics generalize across handedness, clinical populations, and different inhibitory contexts. Resource availability Lead contact Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Sung-Phil Kim (spkim@unist.ac.kr). Materials availability This study has not generated any new reagents. Data and code availability •All data reported in this paper will be shared by the lead contact upon request.•This paper does not report any original code.•Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. •All data reported in this paper will be shared by the lead contact upon request. All data reported in this paper will be shared by the lead contact upon request. •This paper does not report any original code. This paper does not report any original code. •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Acknowledgments This work was supported by the Alchemist Project (Fully Implantable On-body Neuro-AI Interface for Enhancing Physical Performance) funded By the 10.13039/501100003052Ministry of Trade, Industry and Energy (MOTIE, Korea) (no. RS-2025-02634714). This work was also supported by the 10.13039/501100003725National Research Foundation of Korea (NRF) grant funded by the Korea government (10.13039/501100014188MSIT) (no. RS-2023-00302489). We also thank P.G. Bissett for helpful conversation about this work. Author contributions Conceptualization, E.O. and S.-P.K.; methodology, E.O., Y.S.C., and S.-P.K.; software and formal analysis, E.O.; visualization, E.O.; investigation, E.O. and Y.S.C.; data curation, E.O.; writing – original draft, E.O.; writing – review and editing, E.O., Y.S.C., and S.-P.K.; funding acquisition, S.-P.K.; resources, S.-P.K.; supervision, S.-P.K. Declaration of interests The authors declare no competing interests. STAR★Methods Key resources table REAGENT or RESOURCESOURCEIDENTIFIERSoftware and algorithmsMATLAB 2023aMathWorks; https://www.mathworks.com/products/matlab.htmlRRID: SCR_001622 REAGENT or RESOURCESOURCEIDENTIFIERSoftware and algorithmsMATLAB 2023aMathWorks; https://www.mathworks.com/products/matlab.htmlRRID: SCR_001622 Experimental model and study participant details Participants Forty-seven healthy adults (23 females, mean age: 23.0±3.38 years old, all right-handed) were recruited for this study. All participants had normal or corrected-to-normal vision and no history of neurological diseases. Informed written consent was obtained from all participants prior to the experiment, and all procedures were approved by the Institutional Review Board of Ulsan National Institute of Science and Technology (UNISTIRB-22-88-C), in compliance with the Declaration of Helsinki. During data processing, the data of three participants were excluded. One participant was excluded because of frequent coughing during the experiment, which induced significant noise in the EEG data. The other two participants were excluded due to poor signal quality during recording, caused by a detached ground electrode and electrode abnormalities detected in the channel of interest (Fz) during the motor execution session. Consequently, the data from forty-four subjects (22 females, mean age: 23.07±3.40 years old, all right-handed) were used in the data analysis. Method details Experimental procedure The experimental procedure of the study was originally designed to investigate the temporal dynamics of beta burst rates during both motor execution (ME) and motor imagery (MI), although the experimental data during motor execution were only used in this study. Upon arrival, participants completed the Movement Imagery Questionnaire–Revised, Second Version (MIQ-RS).69 After the survey, they first practiced the SST and then performed the SST under two conditions with the ME session always preceding the MI session. During these tasks, participants executed responses or stops, either with (ME) or without (MI) actual movements, while their EEG data were recorded. Since the present study focused exclusively on analyzing EEG data from the ME task, the descriptions of the procedures related to MI are omitted hereafter. Task Participants performed the visually cued SST consisting of both go and stop trials (Figure 1A). Each trial began with a 1,500 ms fixation cross, followed by the presentation of a white leftward or rightward arrow as a go cue. Participants were instructed to respond to the arrow’s direction as quickly and accurately as possible using their left or right thumbs, pressing response buttons assigned to the “3” key (Left) or the “9” key (Right) on a numeric keyboard. The go cue remained on the screen for 1,000 ms or until a response was made. In stop trials, the go cue appeared first, but after a brief delay, the white arrow turned red to indicate a stop signal. Participants were instructed to withhold their response immediately after they saw the stop signal. An inter-trial interval was set to be 1,500 ms. The monitor refresh rate was 60 Hz, and the viewing distance was roughly 70 cm. The time interval between the go gue and stop signal, known as SSD, varied over trials. It was initialized as 250 ms and was dynamically adjusted on a trial basis based on participants’ performance in the previous stop trial to maintain the probability of responding on stop trials approximately as 0.5 (P(respond | stop signal presented) ≈0.5). The maintenance of the probability was achieved through a staircase procedure as follows8,37: if participants failed to stop in the previous stop trial, SSD decreased by 33 ms; if they successfully stopped, SSD increased by 33 ms. The lower bound of SSD was set to 40 ms to prevent the stop signal from appearing too soon, possibly before participants could perceive the go signal, which would interfere with proper initiation of the go process. The task consisted of 10 blocks of 80 trials each, resulting in a total of 800 trials (600 go trials and 200 stop trials, with stop trials comprising 25% of the total). The order of go and stop trials in a block was randomized. Behavioral analysis A number of behavioral measures were employed in this study, including mean SSD, SSRT, P(respond|stop-signal), mean Failed-stop RT, mean Go RT, Go omission rate, and Go error rate. Mean SSD and SSRT were measured to mark critical time points for the temporal dynamics of beta bursts. P(respond|stop-signal) was measured to ensure that SSD was appropriately adjusted throughout stop trials and also to estimate SSRT (see below). Mean Failed-stop RT and mean Go RT were measured to verify that RT measurements satisfied the condition of the race model. Go omission rate and Go error rate were measured to identify any participants who might not have performed the task properly. The probability of responding in stop trials, P(respond|stop-signal), was calculated as the ratio of the number of responded stop trials to the total number of stop trials (Figures S1A and 2B). The Go omission rate was calculated as the ratio of the number of go trials in which no response was made until 1,000 ms after the go cue to the total number of go trials. The Go error rate was calculated as the ratio of the number of go trials with incorrect directional responses to the total number of go trials. Mean SSD was the mean of SSD values from all stop trials. Mean Failed-stop RT and mean Go RT were obtained by averaging RT values across failed stop trials and correct go trials, respectively. According to the race model, mean Failed-stop RT should be shorter than mean Go RT (Failed-stop RT < Go RT) because failed stops occur when the go process reaches the response threshold before the stop process ends. Thus, Failed-stop RTs reflect the faster portion of the Go RT distribution, typically falling on the left tail of the Go RT distribution.1,8 We verified this by statistically evaluating all participants’ mean Failed-stop RT and mean Go RT data using a paired t-test (Figure S1B). To estimate stop-signal reaction time (SSRT), two representative methods are commonly used. The mean method calculates SSRT by subtracting the mean stop-signal delay (SSD) from the mean Go RT. The integration method identifies the time point on an individual’s integrated Go RT distribution that corresponds to their probability of responding on stop trials (i.e., P(respond|stop-signal)).8 In this study, we employed the integration method with omission replacement, as recommended by the previous study.8 Specifically, when we constructed the Go RT distribution, we used all correct and omitted go trials while replacing the RT of omitted go trials with the maximum RT of correct go trials. This modified integration method mitigates the influence of the right tail of the ex-Gaussian distribution on Go RT, making the calculation more reliable and less biased. After constructing the Go RT distribution for each participant, the n-th Go RT was determined by multiplying the number of Go RTs (i.e., 600 in this study) by individual P(respond|stop-signal). That is, we identified the n-th percentile of Go RTs where n = P(respond|stop-signal) x 100 % (Figure S2A). The SSRT was then calculated as the n-th Go RT - mean SSD (Figure S2C). EEG recordings EEG data were acquired from 31-channel wet electrodes using an EEG amplifier (actiChamp plus GmbH, Brain Products, Germany) at a sampling rate of 500 Hz. The placement of those 31-channel electrodes followed the international 10/20 system (FP1, FPz, FP2, F7, F3, Fz, F4, F8, FC9, FC5, FC1, FC2, FC6, FC10, T7, C3, Cz, C4, T8, CP5, CP1, CP2, CP6, P7, P3, Pz, P4, P8, O1, Oz, and O2). The ground and reference electrodes were placed on the left and right mastoids, respectively. During the experiment, electrode impedance was maintained below 10 kΩ. Trigger events were recorded at the time of fixation, go and stop signals, as well as participants’ responses. EEG data preprocessing EEG signals were preprocessed using the EEGLAB toolbox version 14.1.1 in MATLAB.70 EEG signals were filtered with finite impulse response (FIR) filters, applying a high-pass cutoff of 0.3 Hz and a low-pass cutoff of 30 Hz. Bad channels were identified and rejected based on joint probability between channels, with a threshold set to 5.70 When bad channels were removed, spherical interpolation was applied to reconstruct the signals of the removed channel. Independent Component Analysis (ICA) was performed to identify and remove independent components associated with artifacts, including eye movements, cardiac activity, and electrode noise.71 Additionally, current source density (CSD) transformation was applied using an m-constant of 4, reflecting the flexibility of the spherical splines, to reduce volume conduction effects in scalp EEG data for the analysis of spatial patterns.72,73 The continuous EEG data were segmented into epochs from 1,000 ms before to 1,000 ms after the onset of the go signal. The final EEG data matrix after preprocessing had dimensions of 31 channels x 1,000 time points x 800 trials for each participant. Time-frequency transformation of EEG EEG signals within the epoch at each channel were convolved with a complex Morlet wavelet with 7 cycles for evenly spaced frequencies across 1-30 Hz with a frequency resolution of f/7 Hz. We ensured that the time-frequency transformation included the beta frequency band (15–29 Hz) that is generally used for beta burst detection.25,27,29 The Morlet wavelet was given by:(Equation 1)w(t,f)=Aexp(−t22σt2)exp(2iπft)where A=1σt2π is the normalization factor, f is the center frequency of the wavelet, and σt=cycles2πf at time t. To avoid edge artifacts, the convolution output was trimmed by removing the half-length of the wavelet from both sides. The squared magnitude of the convolved value was extracted to calculate the power for each time and frequency. This transformation of EEG signals yielded a 2D time-frequency power map for each channel and each trial (1,000 time points x 30 frequencies x 31 channels x 800 trials). Beta burst detection Beta bursts were detected based on the methods described in the previous EEG beta bursts studies.25,27,29 For each trial and each channel, the time-frequency power map was smoothed using a 2D Gaussian filter with a standard deviation of 2. Local maxima on the smoothed power map were subsequently identified using MATLAB’s imregional() function. A beta burst was deemed to occur at time t if a local maximum was identified at t and it exceeded a beta burst threshold, defined as 6 times the median of power values calculated at the same frequency as the center of the local maximum in a given trial (Figure 1B). Temporal dynamics of beta burst rates The temporal dynamics of beta bursts were examined based on time-varying beta burst rates. A time-varying beta burst rate was estimated by counting the number of beta bursts within a 50 ms sliding window with 50% overlap for each trial and dividing the count by the window length (Figure 1C). This overlapping windowing allowed a high temporal resolution for estimating the time-varying beta burst rate. Also, we set the window size 50ms because using narrower windows (e.g., 25 ms) resulted in noisy and unstable rate estimates, whereas substantially wider windows reduced temporal precision (see Figure S5). Since this study focused on response inhibition, the beta burst rate was time-locked to SSD determined in each trial. However, we also compared the beta burst rates between stop and go trials, which necessitated the alignment of the beta burst rates of go trials with those of stop trials. Since no SSD was available in go trials, for a given go trial, we utilized SSD from the latest stop trial and assigned it as a putative SSD, to which the best burst rate of that go trial was time-locked. When there were no preceding stop trials for a given go trial, the initial SSD (250 ms) was used instead. Comparison of temporal dynamics of beta burst rates among behaviors For the temporal dynamics analysis, the beta burst rate was averaged across trials for each participant and behavior: Successful Stop (SS), Failed Stop (FS), and Correct Go (CG). To account for individual differences in the number of trials across conditions, the averaged beta burst rates were z-scored separately within each behavior for each participant. The mean and standard deviation for z-scoring were calculated using the average beta burst rate within 500 ms window preceding SSD (stop-cue). These z-scored beta burst rates were then used to compute the grand average beta burst rates and standard errors over participants for each behavior. To assess across-condition differences in beta burst rates, non-parametric permutation tests were conducted at each time point within a period of 0 to 450 ms after the stop-signal onset. For each of the three pair-wise comparisons (FS vs. CG, SS vs. CG, SS vs. FS), a permutation test with 1,000 iterations was performed. In each iteration, subject-wise condition labels were randomly shuffled to generate a null distribution of mean differences. Observed differences were compared to this null distribution to compute two-tailed p-values. To control multiple comparisons across time points, p-values were corrected using the Benjamini-Hochberg false discovery rate (FDR) procedure,74 and statistically significant time points (p < 0.05) were identified for each comparison. Cohen’s d, effect size, and 95% confidence interval for the null distributions were also estimated via percentiles. Difference data in the initial 100 ms were excluded to account for potential cue-evoked perceptual responses.75 Comparison of beta burst rates across different time windows To further characterize the temporal dynamics of beta burst activity, the analysis window was divided into three predefined time intervals: SSD-SSRT, pre-SSRT (100 ms before SSRT), and post-SSRT (100 ms after SSRT). Time-average beta burst rates were calculated within each window for both SS and FS trials. A paired t-test was conducted to statistically compare the time-average beta burst rates between SS and FS within each window, separately for left and right movement trials. Beta oscillation power The Short-time Fourier transform (STFT) was applied to the preprocessed EEG data at each channel. A 100 ms of window with 75% overlap (25 ms step size) was used for spectrogram() MATLAB function from Signal Processing Toolbox. The frequency resolution was set to 1.95 Hz by specifying 256 points for the Discrete Fourier Transform (DFT). This procedure resulted in [frequency × time] power-value matrix of size R14×121 per trial. For each trial, power maps were time-aligned to SSD, consistent with the beta burst rate alignment. After SSD-locking, power values were converted to relative power (% change) by subtracting and dividing by the mean power during the baseline period (-500 ms to 0 ms from SSD), and multiplying by 100 at each frequency. The baseline-corrected ERSPs were then averaged across the beta band (15-29 Hz) and across trials for each condition. To assess condition-wise differences, the same non-parametric statistical procedure described for the beta burst rate analysis was applied. Also, to examine how the beta power relates to the beta burst rate, we computed Pearson correlation coefficients (CC) using two complementary approaches: (i) across participants at each time point (time-resolved correlation), and (ii) across time points within subject (within-subject correlation). These analyses were performed for each key channel (F8, Fz, C3 and C4) and condition (CG, FS, SS). The time-resolved correlation analysis revealed positive CC at some time points, but overall mean CC was near zero (mean r = 0.06±0.03). The within-subject correlation analysis showed that the grand-average CC across subjects was near zero (mean r = -0.05± 0.04). Together, these results indicate that while beta power and beta burst rates are partially related, they capture distinct aspects of beta dynamics. Temporal decoding of beta burst rates To examine how multivariate beta burst rate patterns encode behaviors in movement cancellation, we performed within-subject trial-level classification using support vector machines (SVM) with linear kernel on SSD-locked beta burst rates. At each time point, a linear SVM classifier with ridge regularization was trained to discriminate between pairs of behaviors: FS vs. CG or SS vs. CG. The decoding procedure was conducted as follows: For each subject and for each movement direction (left/right), beta burst rate data were partitioned into 10 folds for stratified cross-validation. Since the number of CG trials typically exceeded that of stop trials (e.g., 300 CG trials vs. 50 FS or SS trials in each movement direction), class weights were applied during training to balance the model. Specifically, weights were computed inversely proportional to class frequencies, such that each class contributed equally to the loss function (i.e., 0.5 divided by the number of trials per class). This follows the balanced class weighting scheme implemented in scikit-learn.76 Training and testing data at each channel were z-scored using the mean and standard deviation of the training data, computed across all time points and trials within each condition and fold, to account for channel-wise imbalance and to ensure interpretability of the Haufe-transformed spatial patterns.77,78 Decoding was performed at every time point independently by training the classifier on the z-scored beta burst rates and evaluating its performance on the test set. Balanced accuracy was calculated as the average of true positive rate (sensitivity) and true negative rate (specificity) to account for unequal class sizes. To assess the group-level temporal decoding performance, the mean and standard error of balanced accuracy were computed across subjects at each time point. Statistical significance was evaluated using a non-parametric permutation test with 1,000 random shuffling of class labels. For each permutation, class labels were randomly reassigned, and the decoding procedure was repeated to generate a null distribution of decoding accuracy. Time points at which the observed accuracy exceeded the 95th percentile of the null distribution were considered significantly above chance level (50%), with statistical significance assessed at p < 0.05 and corrected for multiple comparisons using the Benjamini–Hochberg FDR method. Moreover, Haufe-transformed weight maps were computed from SVM model coefficients and averaged across folds and subjects to visualize condition-specific spatial patterns over time. The value was visualized as a topographical map to assess the importance of each channel in terms of its contribution to the decoding performance between stops and go. To assess whether decoding of SS trials emerged earlier than that of FS trials, we performed a bootstrap-based onset analysis at the group level. At each of 1,000 bootstrap iterations, subjects were resampled with replacement, and the decoding accuracy was averaged across the selected subjects for each trial type (FS and SS) and movement direction (left/right). For statistical evaluation, the corresponding null distribution was constructed by averaging the subject-specific null distributions (obtained from individual permutation tests) from the same selected subjects. At each time point, a p-value was calculated by comparing the observed group-level accuracy against this bootstrapped null distribution. Decoding onset was defined as the first time point (with exclusion of the first 100 ms interval for perceptual responses to the cue) within a cluster of at least two consecutive time bins showing significant decoding (p < 0.05, uncorrected). This process yielded distributions of decoding onset times across bootstrap samples for FS and SS. The bootstrapped decoding onset distributions were compared between FS and SS using Wilcoxon signed-rank tests. Low-dimensional representations of beta burst rates To visualize the temporal dynamics of beta burst rates in a lower-dimensional latent space, we applied principal component analysis (PCA) to beta burst rates. Prior to PCA, beta burst rates were sampled every 12.5 ms and z-scored at each channel. Then, multi-channel beta burst rates was concatenated across time points, behaviors, and participants into a single matrix of size [T×C×S]×N, where T is the number of time points, C is the number of behaviors, S is the number of participants, and N is the number of EEG channels. For instance, when all EEG channels were used, T = 159, C = 3, S = 44 and N = 31. This approach allowed us to derive a common principal component (PC) axis across participants. Following PCA, the top three principal components (PCs) were selected, and the original high-dimensional multi-channel beta burst rates were projected onto these PCs, yielding [T×C×S]×3 matrix. The projected data were then averaged across participants (S) in each behavior, yielding T×3 matrix per behavior. In other words, the projection generated a 3D trajectory of low-dimensional representations of multi-channel beta burst rates (over T). To visualize the trajectory smoothly, a Gaussian filter (σ = 4) was applied. The trajectory was time-locked to either the go cue to visualize overall dynamics or the SSD to visualize focused dynamics after SSD. To characterize the geometric orientation of 3D trajectories, we estimated a time-varying plane using singular value decomposition (SVD) with 4 consecutive time point, so 50 ms window was used with 1 step (12.5 ms). A back-causal binning approach was used such that the orientation at each time point reflected the plane defined by the current and three preceding time points. For each time window, we first computed the centroid of the points and centered the data accordingly. SVD was then applied to the centered data matrix, and right singular vector was used to draw planes.(Equation 2)X=UΣVT Since the third right singular vector is the direction of minimal variance, it was used as the normal vector, while the first and second right singular vectors defined the local axes within the plane. To quantify differences in the geometric orientation of neural trajectories across conditions, we computed the angle between the planes of different behaviors over time. At each window, pairwise angles between the normal vectors (CG–FS or CG–SS) were calculated using the cosine similarity formula:(Equation 3)θ=cos−1(ni→·nj→‖ni→‖‖nj→‖) To ensure interpretability within a 0°–90° range, angles greater than 90° were reflected across the 90° axis. The resulting angle time series were smoothed using a Gaussian filter (σ = 1) to reduce high-frequency fluctuations. We then visualized the temporal evolution of angular differences between SS/FS and CG, highlighting the time windows of interest in the stop signal and SSRT period. Note that time points earlier than 100 ms were excluded to minimize the influence of early visual processing, consistent with the rationale used in the beta burst rate analyses. Spatiotemporal pattern of beta burst dynamics To analyze spatiotemporal patterns of beta bursts over the brain regions, we identified time points at which beta burst activity of inhibitory control was observed. Here we putatively determined the occurrence of beta burst activity of inhibitory control if: 1) beta burst rates in stop trials were higher than those in CG trials, and 2) beta burst rates in SS trials were higher than those in FS trials. To directly compare the temporal dynamics of FS and SS trials, we first re-referenced each by subtracting the beta burst rate of CG trials, treating the CG condition as a reference. These differences of beta burst rates (FS–CG and SS–CG) were then statistically tested using a non-parametric permutation test (1,000 iterations) within the time window of 150–450 ms, during which beta burst activity was typically elevated. These comparisons were not intended as confirmatory tests, but rather to identify representative time windows for qualitative illustration of sequential dynamics instead of simply selecting peaks. For each channel and movement direction (left/right trials), the median of the significant timepoints was computed to visualize the spatiotemporal sequence of beta burst modulation across the channels of interest (F8, Fz, C3, and C4 contralateral to the response side). To visualize this channel-wise latency intuitively, we also drew the time-resolved beta burst rate difference (SS – FS) across channels. Quantification and statistical analysis All statistical analyses were performed using MATLAB (MathWorks). In all analyses, the unit of statistical inference was the participant. Data are reported as mean ± standard error of the mean (SEM) unless otherwise stated. The significance level was set at α = 0.05. Statistical details for each analysis, including the statistical tests used, sample sizes, and significance levels, are also reported in the corresponding figure legends and results section. Behavioral measures, including mean Go RT and failed stop RT, were compared using paired t-test across participants (n = 44). For time-resolved analyses of beta burst rates and beta power, non-parametric permutation tests (1,000 iterations) were conducted at each time point to assess condition differences (e.g., SS vs. FS). To control for multiple comparisons across time points, p-values were corrected using the Benjamini–Hochberg FDR method74 with a significance threshold of q < 0.05. Effect sizes (Cohen’s d) and 95% confidence intervals were also computed where applicable. For predefined time-window analyses (SSD-SSRT, pre-SSRT, and post-SSRT), paired t-tests were used to compare beta burst rates between conditions. For temporal decoding analyses, classification performance (balanced accuracy) was evaluated using stratified 10-fold cross-validation within each participant. Statistical significance of decoding accuracy was assessed using permutation test (1,000 iterations) by randomly shuffling class labels to generate null distributions. Time points exceeding the 95th percentile of the null distribution were considered significant (p < 0.05), with FDR correction applied to time points. Decoding onset times were estimated using bootstrap resampling (1,000 iterations), and onset differences between conditions were tested using Wilcoxon signed-rank tests.