Temporal cascade of frontal, motor and muscle processes underlying human action-stopping

Action-stopping is a canonical executive function thought to involve top-down control over the motor system. Here we aimed to validate this stopping system using high temporal resolution methods in humans. We show that, following the requirement to stop, there was an increase of right frontal beta (~13 to 30 Hz) at ~120 ms, likely a proxy of right inferior frontal gyrus; then, at 140 ms, there was a broad skeletomotor suppression, likely reflecting the impact of the subthalamic nucleus on basal ganglia output; then, at ~160 ms, suppression was detected in the muscle, and, finally, the behavioral time of stopping was ~220 ms. This temporal cascade supports a physiological model of action-stopping, and partitions it into subprocesses that are isolable to different nodes and are more precise than the behavioral latency of stopping. Variation in these subprocesses, including at the single-trial level, could better explain individual differences in impulse control.

measured with single-pulse TMS. In studies four and five we turned to the cortical process 85 thought to initiate action-stopping, using the above-mentioned proxy of right frontal beta 31,32 . 86 We measured scalp EEG, derived a right frontal spatial filter in each participant, and then 87 extracted beta bursts 35 in the time period between the Stop signal and SSRT. We tested how the 88 timing of these beta bursts related to CancelTime. 89 90

91
Study 1 (EMG). 10 participants performed the stop-signal task (Fig. 1a). On each trial they 92 initiated a manual response when a Go cue occurred, and then had to try to stop when a Stop 93 signal suddenly appeared on a minority of trials. Depending on the stop signal delay, SSD, 94 participants succeeded or failed to stop, each ~50% of the time). We measured EMG from the 95 responding right index and little fingers (Fig. 1b inset). Behavioral performance was typical, with 96 SSRT (referred to as SSRTBeh) of 216±8 ms, and action-stopping on 51±1 % of Stop trials (Table  97 1). EMG analysis was performed on the trial-by-trial root-mean-squared EMG (EMGRMS; Fig.  98 1b). On 53±6 % of Successful Stop trials (i.e. where no keypress was made) there was a small 99 but detectible EMG response (Partial EMG trials; see Supplementary  We hypothesized that the time when the Partial EMG response starts declining after the 112 Stop signal is a readout of the time when the Stop process is implemented in the muscle 113 (hereafter 'CancelTime'). We observed that, first, CancelTime is much earlier than SSRTBeh (see 114 Study 2 (EMG). We then ran a new sample (n = 32; see Table 1 for behavioral results). Again, 134 we observed partial EMG responses on 49±2 % of Successful Stop trials; where the EMG 135 amplitude was 54±1 % smaller than the amplitude in trials with a keypress (Fig. 2b). Fig. 2c  The rationale was that, at shorter SSDs, the Go process will have been active for a shorter 163 duration, meaning EMG activity will not have increased much before being inhibited, while at 164 longer SSD, the Go process will have been active for a longer duration, meaning EMG activity 165 will have increased much more before being inhibited. Indeed, the amplitude of the Partial EMG 166 responses increased with SSD ( Supplementary Fig. 1c). A 1-way repeated measures ANOVA 167 with amplitude as the dependent variable and the SSD as the independent variable showed 168 significant effect of SSD on amplitude (F(4,24) = 3.7, p = 0.018, " # = 0.4) [also see 23 ]. This 169 suggests that the Partial EMG trials represent inhibited Go responses and not merely a weak Go 170 process (which would presumably not increase across SSDs). 171 To further validate CancelTime, we modelled the behavior using BEESTS (Bayesian 172 Estimation of Ex-gaussian STop-Signal reaction time distributions; see Table 2

Study 3 (TMS).
To further validate CancelTime and relate it to brain processes we turned to a 221 different method -single-pulse TMS over a task-irrelevant muscle representation in the brain. As 222 mentioned above, the reduction of MEPs from task-irrelevant muscles on Successful Stop 223 trials 25-27 , is thought to reflect a basal ganglia-mediated global suppression 29 . Eighteen new 224 participants (see Table 1 for behavioral results) now performed the task with their left hand, 225 while TMS was delivered over the left motor cortex and MEPs were recorded from a task-226 irrelevant, right forearm muscle. MEPs were recorded at different times after the Stop signal on 227 different trials: 100 -180 ms in 20 ms intervals, as well as during the inter-trial interval which 228 served as a baseline. Concurrently, we recorded EMG from the task-relevant left-hand muscles 229 as for studies 1 and 2 above (Fig. 4a). whether the ~15 ms discrepancy in timings could be accounted for by corticospinal conduction 251 delays, we estimated this corticospinal conduction time in a separate phase of the current study muscle (Fig. 4c). This was 23± 0.3 ms. Thus, a decline in muscle activity would be expected to 254 be preceded by a reduction in motor cortical output by ~23 ms, which is very similar to the ~15 255 ms difference between global motor suppression and CancelTime. 256 To further elaborate the temporal relationship between global motor suppression and 257 CancelTime, we performed a trial-by-trial analysis whereby MEP amplitudes were sorted 258 according to the time at which TMS was delivered, relative to the time at which EMG decreased   process at the muscle (Studies 1 and 2, EMG/behavior), which also related tightly with the 286 timing of global motor suppression (Study 3, TMS), we then tested whether this EMG measure 287 was also related to the timing of a prefrontal correlate of action-stopping, specifically the 288 increase of beta power (13-30 Hz) before SSRTBeh at right frontal electrode sites 32,33 . We now 289 measured scalp EEG as well as EMG from the hand, in 11 participants (see Table 1 for 290 behavioral results). We derived beta bursts rather than beta power per se, as bursts have richer 291 features 37 , such as burst timing and duration. across all participants is shown in Fig. 5b inset. For each participant, we estimated beta bursts; 299 First, by filtering the data at the peak beta frequency; and Second, by defining a burst threshold 300 based on the beta amplitude in a baseline period after the Stop signal (500-1000 ms after Stop 301 signal in the Stop trials, and 500-1000 ms after the mean SSD in the Correct Go trials) (see 302 Methods; Supplementary Fig 4). 303 In an exemplar participant, the burst % increased for Successful Stop compared to both 304  Table 1 for behavioral 342 results). As above a right frontal IC was extracted for each participant (average topography Fig.  343 5c inset, see Supplementary Fig 3)  Notably, CancelTime was ~60 ms earlier than SSRTBeh. To better understand this 389 discrepancy, we calculated SSRT based on the EMG response rather than behavior. We saw that 390 SSRTEMG better matched CancelTime than did SSRTBeh. Thus, SSRTBeh could be an over-391 estimation of the duration of the Stop process in the brain. This extra time in SSRTBeh probably 392 reflects a 'ballistic stage' in generation of the button press 40,41 . We suggest that the maximum 393 CancelTime reflects the last point at which a Stop process can intervene to prevent responses. 394 We note that CancelTime (a muscle measurement) is an overestimation of the brain's stopping 395 speed since it does not include the corticospinal conduction time, which we estimated at ~20 ms. 396 Indeed, our TMS results show that global motor suppression, which we take as the time at which 397 motor areas of the brain are suppressed, is ~140 ms (which is ~15 ms less than CancelTime).
One important consequence of our observation that the brain's stopping speed is ~140 ms is that 399 neural events that mediate stopping need to occur before this time. Indeed, we found that right 400 frontal beta activity increased ~120 ms after the Stop signal on Successful Stop trials, and also 401 that, across participants, there was a strong positive relationship between mean BurstTime and 402 mean CancelTime. 403 Taken together, these studies motivate a detailed model of the temporal events of action-404 stopping (Fig. 6). First, we suppose the right frontal beta bursts relate to activity of right inferior 405 frontal gyrus 12,31 , and this happens in ~120 ms, which then leads via basal ganglia 29 to global 406 suppression of the primary motor cortex 25-28,30 at ~140 ms. After a corticospinal conduction 407 delay of ~20 ms, this suppression of motor output is then reflected at ~160 ms as a decline in 408 muscle activity (CancelTime). Finally, SSRTBeh occurs at ~220 ms, after, what we suppose is an 409 electromechanical delay of ~60 ms.  Whereas meta-analysis shows that SSRTBeh is longer for patients (e.g. ADHD, OCD, and 436 substance use disorder) vs. controls [6][7][8][9][10][11] , not all such studies show differences 8,46-48 . We predict 437 that our new single-trial method of CancelTime will be more sensitive than SSRTBeh. 438 Furthermore, future studies can easily estimate within-subject variability in CancelTime, which 439 will likely discriminate patients from controls. Third, our results provide insight into why 440 SSRTBeh might only have a modest relationship with more 'real-world' measures of 441 impulsivity 15-20 . As we show, the SSRTBeh includes not only CancelTime but an extra, and 442 variable, 60 ms ballistic stage. We expect that future studies may show stronger correlations 443 between CancelTime and self-report than that seen between SSRTBeh and self-report (also see 15 ); 444 likewise we predict that right frontal beta burst time might also correlate more tightly with self-445 report measures. More generally, the detailed timing information of frontal beta at ~120 ms, 446 global motor suppression at ~140 ms, and CancelTime at ~160 ms points to subprocesses of 447 action-stopping that provide potential biomarkers that could better explain individual differences 448 in impulse control. 449 In conclusion, we provide a detailed timing model of action-stopping that partitions it 450 into subprocesses that are isolable to different nodes and are surely more precise than the 451 behavioral speed of stopping. At the core of this timing model is a novel method of measuring 452 the speed of stopping from the muscles. This provides a single-trial estimate of stopping speed 453 that could be easily measured with minimal equipment in any lab that studies human participated. Two were excluded from analysis, one for misaligned EEG markers due to a 471 technical issue, while the other lacked a right frontal brain IC, based on our standard method 32,33 . 472 473 Stop-signal task. This was run with MATLAB 2014b (Mathworks, USA) and Psychtoolbox 50 . 474 Each trial began with a white square appearing at the center of the screen for 500±50 ms. Then a 475 right or left white arrow appeared at the center. When the left arrow appeared, participants had to 476 press a key on a vertically oriented keypad using their index finger, while for a right arrow they 477 had to press down on a key on a horizontally oriented keypad with their pinky finger (Fig. 1b  478 inset), as fast and as accurately as possible (Go trials). The stimuli remained on the screen for 1 presented. On 25% of the trials, the arrow turned red after a stop signal delay (SSD), and 481 participants tried to stop the response (Stop trials). The SSD was adjusted using two independent 482 staircases (for right and left directions), where the SSD increased and decreased by 50 ms 483 following a Successful Stop and Failed Stop respectively. Each trial was followed by an inter-484 trial interval (ITI) and the entire duration of each trial including the ITI is 2.5 s (Fig. 1a). 485 following the fixation cue. The data acquisition was triggered from MATLAB using a USB-505 1208FS DAQ card (Measuring Computing, Norton, MA). In all 5 experiments, surface EMG 506 was recorded from both the first dorsal interossei (FDI) and the abductor digiti minimi (ADM) 507 muscles of the hand (Fig. 1b inset). In the TMS experiment, surface EMG was also recorded 508 from the task-irrelevant right extensor carpi radialis (ECR) muscle (Fig. 5a). window of 50 ms. Any EMG activity which was greater than 8 SD of the mean EMG activity in 550 the baseline period (Fixation to Go cue) was marked, on a trial-by-trial basis. Starting from the 551 peak of that EMG activity, the onset was marked at the point where the activity dropped below 552 20% of the peak for 5 consecutive ms. This method of adjusting the threshold based on the peak 553 EMG activity, allowed better onset detection than a fixed threshold, especially when the 554 amplitude of the EMG activity was small. The time when EMG started to decline was 555 determined as the time when, following the peak EMG activity, the activity decreased for 5 556 consecutive ms. Visual inspection of individual trials showed that this method provided a reliable 557 detection of both EMG onsets (see Supplementary Fig. 1a, 1b for EMG onset vs. RT correlation) 558 and decline. Any detected EMG timing which was beyond 1.5 times the inter-quartile range 559 (IQR) of the first and third quartile (Q3) of that particular timing distribution was deemed an 560 outlier. This removed <4% trials. CancelTime was marked as the time of the decline following 561 the Stop signal. For outlier rejection, CancelTimes had a lower cutoff of 50 ms and higher cutoff 562 of Q3+1.5´IQR. This removed <3% trials. 563 As the peak EMG amplitude for the FDI and ADM muscle were quite distinct, before 564 averaging the two EMG activities, we normalized the muscle activity by the peak activity in that 565 particular muscle (VoltageNorm in Fig. 2a, 2b, 3c, Supplementary Fig. 1c TMS over the hand representation of left FDI and measuring MEP from the muscle (Fig. 4c). 577 The earliest MEP onset latency across 10 trials was identified by visual inspection of the EMG 578 traces 52-54 . 579 Trial-by-trial analysis of the time of the CancelTime and time of global motor suppression. To 580 compare the temporal association between the EMG decline and MEP suppression, we 581 performed a trial-by-trial analysis of stop-signal task data only on trials where an EMG burst was 582 detected. We first normalized the time of TMS on a given trial by subtracting the time of EMG 583 decline from the time of the TMS pulse. Hence, negative values mean that TMS was delivered 584 before the EMG decline and positive values mean that TMS was delivered after. We then plotted 585 MEP amplitudes for each of the three response types (Correct Go, Failed Stop, and Successful 586 Stop) against the normalized times binned into 30 ms windows. This analysis meant that for a 587 given individual there were relatively few trials per time bin, and some bins would occasionally 588 contain no data. Therefore, we combined data across all individuals. Prior to this, MEP 589 amplitudes for each individual were normalized to the mean MEP amplitude at the inter-trial 590 interval, to account for inter-individual variability in absolute MEP amplitudes at baseline. We 591 restricted our analysis to time bins that contained at least 50 trials, which resulted in time range -592 90 ms to 60 ms. 593 EEG Preprocessing. We used EEGLAB 55 and custom-made scripts to analyze the data. The data FIR notch filter were applied to remove line noise and its harmonics. EEG data were then re-596 referenced to the average. The continuous data were visually inspected to remove bad channels 597 and noisy stretches. 598 ICA analysis. The noise-rejected data were then subjected to logistic Infomax ICA to isolate 599 independent components (ICs) for each participant separately 38 . We then computed the best-600 fitting single equivalent dipole matched to the scalp projection for each IC using the DIPFIT 601 toolbox in EEGLAB 55,56 . ICs representing non-brain activity related to eye movements, muscle, 602 and other sources were first identified using the frequency spectrum (increased power at high 603 frequencies), scalp maps (activity outside the brain) and the residual variance of the dipole 604 (greater than 15%) and then, subtracted from the data. A putative right frontal IC was then 605 identified from the scalp maps (if not present then we used frontal topography) and the channel 606 data were projected onto the corresponding right frontal IC. The data on Successful Stop trials 607 were then epoched from -1.5 s to 1.5 s aligned to the Stop signal. We estimated the time-608 frequency maps from 4 to 30 Hz, and -100 to 400 ms using Morlet wavelets with 3 cycles at low 609 frequencies linearly increasing by 0.5 at higher frequencies. The IC was selected only if there 610 was a beta power (13 to 30 Hz) increase in the window between the Stop signal and SSRTBeh 611 compared to a time-window prior to the Go cue (-1000 to -500 ms aligned to Stop signal). In 612 each participant, the beta frequency which had the maximum power in this time window was 613 used in the beta bursts computation (Supplementary Fig. 3).
within a period of 500 to 1000 ms (i.e. after the Stop signal in the Stop trials, and after the mean 619 SSD in the Correct Go trials) was pooled across all trials [compared to the ICA analysis here we 620 picked a different time-window to estimate the burst threshold to keep the analysis unbiased. 621 However, picking the same time-window also yielded similar results]. The threshold was set as 622 the median + 1.5 SD of the beta amplitude distribution (Supplementary Fig. 4). Once the burst 623 was detected, the burst width threshold was set as the median + 1 SD. We binary-coded each 624 time point where the beta amplitude crossed the burst width threshold to compute the burst % 625 across trials. For each detected burst, the time of the peak beta amplitude was marked as the 626 BurstTime. 627 628 Statistical analysis. For pairwise comparisons, the data were first checked for normality using 629 Lilliefors test, and if normally distributed a two-tailed t-test (t-statistic) was performed, else a 630 Wilcoxon signed rank test (Z-statistic) was performed. We interpret the effect sizes as small 631 ANOVAs were interpreted as small (partial eta-squared, hp 2 : 0.01-0.06), medium (hp 2 : 0.06-637 0.14), and large (hp 2 > 0.14). For correlational analyses, Pearson's correlation coefficient (r) was In testing the relationship between BurstTime and CancelTime, we performed a 641 permutation test. We sampled BurstTimes randomly from a uniform distribution between 0 and 642 SSRTBeh for a given participant for 3000 iterations. For each iteration, we then computed the 643 correlation (r) between the mean BurstTime and the mean CancelTime across participants. This 644 generated a distribution of r ranging between -1 and 1. The p-value for our analysis was 645 determined as the P(r³rObs|H0) in the permuted data. the group-level distributions. This approach is thought to be more accurate than fitting individual 660 participants and is effective when there is less data per participant 57 . We pooled the subjects 661 across both study 1 and 2 to estimate the individual parameters. The priors were bounded 662 uniform distributions (µGo, µStop: U(0,2); sGo, sStop: U(0,0.5) tGo, tStop: U(0,0.5); pTF: U(0,1)).
The posterior distributions were estimated using the Metropolis-within-Gibbs sampling and we 664 ran multiple chains. We ran the model for 5000 samples with a thinning of 5. The Gelman-Rubin 665 (R ) statistic was used to estimate the convergence of the chain. Chains were considered 666 converged if R < 1.1. 667 668 provided at the above link. 687