Abstract
Executive control of movement enables inhibiting impulsive responses critical for successful navigation of the environment. Circuits mediating stop commands involve prefrontal and basal ganglia structures with fMRI evidence demonstrating increased activity during response inhibition in the dorsolateral prefrontal cortex (dlPFC)—often ascribed to maintaining task attentional demands. Using direct intraoperative cortical recordings in male and female human subjects, we investigated oscillatory dynamics along the rostral-caudal axis of dlPFC during a modified Go/No-go task, probing components of both proactive and reactive motor control. We assessed whether cognitive control is topographically organized along this axis and observed that low-frequency power increased prominently in mid-rostral dlPFC when inhibiting and delaying responses. These findings provide evidence for a key role for mid-rostral dlPFC low-frequency oscillations in sculpting motor control.
- cognitive control
- dorsolateral prefrontal cortex
- electrocorticography
- local field potentials
- response inhibition
Significance Statement
This work offers insights into the neural mechanisms underlying executive control of movement. By employing intraoperative cortical recordings, our study uncovers the specific role of low-frequency oscillations in the mid-rostral dlPFC during response inhibition. We demonstrate a topographical organization of low-frequency power along the rostral-caudal axis of the dlPFC. This finding supports existing work suggesting the PFC hierarchy may be rooted in cognitive demand and extends it by showing this hierarchy can be represented by low-frequency oscillations. Finally, understanding the spatial and temporal dynamics of inhibitory control may enable more effective neuromodulation therapies in the future, such as those aimed at Parkinson disease patients to address impulsivity.
Introduction
Evaluating choices and quickly selecting the optimal response is crucial for survival. For example, at an intersection, the light turns green but opposing cars are still crossing. Avoiding the urge to release the brakes prevents a collision, illustrating the rapid, automatic processes underlying cognitive control (Miller and Cohen, 2001). Cognitive control is thought to be implemented both proactively and reactively (Braver, 2012). Proactive control represents top-down control processes, based on current goals, that enable stopping of inappropriate response tendencies, whereas reactive control is the external signal-driven stopping of a response (Aron, 2011; Braver, 2012; Talanow et al., 2020).
The prefrontal cortex (PFC) is a central hub for mediating such processes (Miller and Cohen, 2001). Electrophysiological and neuroimaging studies have delineated two key subdivisions, the medial PFC (mPFC) and ventrolateral/dorsolateral PFC (vlPFC/dlPFC) in cognitive control. Current theories propose that mPFC is engaged in monitoring external stimuli and evaluating when and to what extent control is needed and lPFC implements this control (MacDonald et al., 2000; Smith et al., 2019). Neural activity in mPFC/cingulate increases when control is needed (Botvinick et al., 1999; Menon et al., 2001; Ridderinkhof et al., 2004; Cavanagh et al., 2011, 2012; Cohen and Cavanagh, 2011; Horga et al., 2011; Cohen and Donner, 2013; Cavanagh and Frank, 2014; Zavala et al., 2018; Smith et al., 2019). The role of vlPFC in motor control is well established (Levy and Wagner, 2011), whereas dlPFC has been implicated in more diverse cognitive functions such as inhibiting prepotent responses (Garavan et al., 1999; Menon et al., 2001; Rubia et al., 2001a,b; Nakata et al., 2008; Chikazoe et al., 2009a; Chikazoe, 2010), modulation of task-related attentional demands (MacDonald et al., 2000), maintenance of working memory (Barbey et al., 2013; Baumert et al., 2020), goal-directed conflict resolution (Shenhav et al., 2013; Smith et al., 2019), and error processing (Menon et al., 2001; Cavanagh et al., 2009). Its specific roles during response inhibition, however, remain understudied. The dlPFC displays a graded pattern of connectivity with other brain networks (Jung et al., 2022). Growing evidence documents that cognitive control in the PFC is organized hierarchically with greater engagement of rostral areas in more abstract (Koechlin et al., 2003; Badre and D’Esposito, 2007, 2009; Badre, 2008; Badre et al., 2009; Azuar et al., 2014) or difficult (Crittenden and Duncan, 2014) cognitive functions.
Theta-frequency oscillations in the PFC are linked to cognitive control (Cohen and Cavanagh, 2011; Cavanagh et al., 2012; Cavanagh and Frank, 2014; Zavala et al., 2018) with other frequencies performing distinct, complimentary functions (Zavala et al., 2018; Prochnow et al., 2022a,b; Wendiggensen et al., 2023). For example, inter-regional cross-frequency coupling is modulated by task abstraction (Voytek et al., 2015). Further, lPFC beta has been associated with attention selection (Dubey et al., 2023), managing working memory (Schmidt et al., 2019), and signaling the status quo of cognitive and sensorimotor states (Engel and Fries, 2010).
Here, we utilized a modified Go/No-go task. While No-go trials in this design can involve cognitive features that are not inhibition [working memory and goal maintenance (Redick et al., 2011), for example], Go/No-go tasks with a small No-go:Go ratio are proposed to create a prepotent tendency to respond and make the No-go trials reflective of inhibition (Chikazoe et al., 2009a; Wessel, 2018; Young et al., 2018). Our design used a 0.33 No-go:Go ratio with a 1,000 ms maximum intertrial interval, which are similar to parameters known to induce prepotent motor activity (Wessel, 2018).
We investigated dlPFC oscillatory activity changes during response inhibition and whether it exhibits a distinct spatial topography. We recorded signals from subdural electrocorticography (ECoG) strips placed over dlPFC during a Go/No-go task in 13 Parkinson's disease (PD) patients undergoing deep brain stimulation surgery (DBS). We hypothesized that No-go trials would increase low-frequency (LF) power in dlPFC and that the inhibition-related spectral response would display a rostro-caudal spatial distribution.
Materials and Methods
Research participants
We recruited 10 male and 3 female neurosurgical patients undergoing DBS electrode implantation for motor symptoms related to PD. The subthalamic nucleus (STN) or globus pallidus interna (GPi) was chosen as the surgical target based on multidisciplinary consensus recommendations by neurologists, neurosurgeons, neuropsychiatrists, and nurses. Study activities were performed in accordance with the University of Alabama at Birmingham Institutional Review Board (IRB)-approved protocols and the study protocol is registered at ClinicalTrials.gov (NCT #04735458). All patients considered for DBS surgery underwent preoperative motor testing and neuropsychological evaluation followed by multidisciplinary conference review to determine surgical candidacy. Patients were approached for study participation with exception of the following exclusion criteria: age younger than 18, inability to provide full and informed study consent, unable to participate in study-related activities, history of prior ischemic/hemorrhagic stroke, subdural hemorrhage, or seizure.
Surgical procedure
All participants underwent 3T MR imaging (Magnetom PRISMA, Siemens Healthcare) prior to surgery as part of routine workup. Surgery was performed in the off-medication, awake state, at least 12 h after the last medication dose. Once in the operating room, DBS surgery proceeded as routinely performed, with the addition of placement of a 6-contact subdural strip electrode (Ad-tech Medical) prior to DBS electrode implantation. The strip electrode was targeted to the middle frontal gyrus using the preoperative MRI, anterior to the premotor area. We passed the strip along the longitudinal trajectory of the gyrus, as detailed in our previous work (Bentley et al., 2020). The DBS electrode and subdural strip were localized intraoperatively by merging a 3D fluoroscopic image to the preoperative MRI. After the completion of research recording, the subdural strip was removed, and the DBS electrode was secured using routine surgical techniques.
Behavioral task
We implemented the Go/No-go task to study inhibitory control. Participants performed practice sessions prior to experimental sessions to minimize potential contributions from task learning. Go/No-go trials were pseudorandomized, with a total of ∼107 trials with a ratio of 2:1 Go versus No-go that creates a propensity to respond to any stimulus (Young et al., 2018). The task was presented using E-Prime software (E-Prime 3.0, Psychology Software Tools) on Steris VividImage 4K intraoperative monitors (Steris IMS) positioned ∼18 inches from patients. As patients were positioned semireclined for surgery, the distances varied within several inches to ensure monitor visibility.
Each experimental session began with task instructions followed by a central crosshair (cue). Sessions consisted of three blocks: rest, “Low Control Demand” (LCD; Go trials only), and “High Control Demand” (HCD; 2:1 ratio of Go:No-go trials), which are presented in consistent order. Stimuli were letters of the alphabet with “X” designated as the No-go stimulus and all other letters as “Go” stimuli. During the rest block, instructions were given to “relax and look at the cross in the middle of the screen.” For the LCD block, participants were told that there will not be any “Xs”, and during the HCD block, instructions included the prompt that there will be “Xs.” Response periods were limited to a maximum of 1,000 ms (ranging from 750 to 1,000 ms based on a 80% accuracy cutoff in a practice block prior to starting the experiment) followed by a feedback period of 250 ms, which consisted of a green color change for accurate responses and a red color change for inaccurate responses. For this study, feedback periods were not included in the current analyses. The intertrial interval was randomized to range between 500 and 1,000 ms. Responses were made with the thumb or index finger using a single-button actuator of the hand ipsilateral to the side of surgery to minimize neural activity related to motor responses that are typically prominent in the contralateral hemisphere. As participants had PD with varying degree of tremor, we ensured that the actuator was successfully held and button presses performed accurately prior to the experimental recordings.
Electrophysiologic recordings and signal processing
We recorded neural activity at 10 kHz from the prefrontal ECoG strip and DBS electrode using an actiCHamp active channel amplifier (BrainVision). We also placed an EMG electrode on the ipsilateral hand's first dorsal interosseus muscle for simultaneous recording of muscle activity corresponding to responses. All recordings were imported into MATLAB (MathWorks) and processed using custom scripts. Each recording channel was visually inspected for artifacts, and those containing excessive noise or large amplitude/sharp artifacts were excluded from further analysis. We then bandpass-filtered the ECoG recordings between 0.5 and 500 Hz using a second-order Butterworth filter and downsampled to 1 kHz. We subsequently re-referenced monopolar ECoG electrodes to adjacent contacts creating five virtual bipolar electrodes. Recordings were separated into 1,500 ms trial epochs beginning 500 ms prior to cue onset to 1,000 ms after cue onset. EMG signals were bandpass filtered between 50 and 100 Hz using a second-order Butterworth filter, downsampled to 1 kHz, rectified, and smoothed using a 100 ms moving average.
LFP power
We convolved 67 logarithmically spaced (10 voices/octave) complex Morlet wavelets (wave number 6) ranging from 2 to 194 Hz with 1.5 s of trial data (−500 to 1,000 ms relative to cue onset), including a 1,500 ms buffer on either side to account for edge effects. We squared the magnitude of the continuous wavelet transform to obtain an estimate of instantaneous power. We then z-scored the LFP power for each time–frequency point and contact to the baseline period of all trials (−500 to −200 ms relative to cue onset) for that frequency and contact.
Multivariate pattern analysis
We performed multivariate pattern analysis (MVPA) to determine whether LF and beta oscillations interact during cognitive control. We used the MVPA-Light toolbox (Treder, 2020) to train a classifier to discriminate between Go and No-go trials. We used a two-class L1-support vector machine (SVM) with five-fold cross-validation (Wendiggensen et al., 2023). This was done by (1) training and testing within the LF and beta time courses and (2) training on the LF time course and testing on the beta. We investigated which time points the classifier could discriminate between conditions and visualized the results using the area under the ROC curve (AUC). We followed this with cluster-based permutation tests where the test statistic was created using Wilcoxon tests with α = 0.05 to determine which time points AUC values differed from the chance level of 0.5. A total of 1,000 iterations were used, and the cluster statistic was computed by summing the Wilcoxon statistic mass of suprathreshold clusters.
Postoperative electrode localization
We performed offline analysis to localize ECoG contacts using the preoperative T1 MRI scan and an intraoperative CT. First, we generated a 3D cortical surface reconstruction using FreeSurfer (Fischl, 2012) and then imported the MRI DICOMs with cortical reconstruction into Brainstorm (Tadel et al., 2011), where the MNI normalization was computed for each subject (Ashburner and Friston, 2005). We then merged the preoperative MRI with the intraoperative CT. Following coregistration, we confirmed electrodes were positioned within dlPFC using the Desikan–Killiany (Desikan et al., 2006) atlas generated by FreeSurfer cortical segmentation. Only contacts that were within the atlas-defined dlPFC were included in our analyses.
Statistical analysis
Only successful Go and No-go responses were included in our main analyses. We used an exclusion criterion for subjects based on whether they had a significantly different response rate between the Go and No-go conditions by comparing the proportion of Go trials in which a response was made to the proportion of No-Go trials in which a response was made using a null hypothesis of equal proportions (Pearson's chi-squared test). This ensured that only subjects who were attentive, comprehended task rules, and correctly engaged with the task were included in our analyses.
To identify time–frequency windows for subsequent statistical analyses, we first sought to identify time–frequency regions where No-go power was different from Go power. We first computed the mean difference in normalized power at each time–frequency point between correct Go and No-go trials within each contact, and then we averaged across contacts and subjects. To determine whether this empirical difference was significant, we used a nonparametric permutation test with a null hypothesis that the difference in power at each time–frequency point was 0. This involved permuting the conditions (Go vs No-Go) prior to computing the mean difference within each contact for each subject and averaging across contacts and subjects. We repeated this procedure 1,000 times and z-scored the true difference value for each time–frequency point using the mean and standard deviation of the null hypothesis distribution and thresholded each pixel at p = 0.05. To correct for multiple comparisons, we used a cluster mass-based approach (Maris and Oostenveld, 2007). Briefly, this involved summing the z-score mass of suprathreshold clusters and keeping the largest in each iteration. We then compared the suprathreshold clusters from the real data with this distribution and classified clusters as significant if they were in the top 5%. In this way we were able to determine the frequencies and time points that exhibited significant differences between the conditions.
After finding a difference in LF power between the conditions, we tested whether these differences might be restricted to subregions within dlPFC. We first classified electrode contacts as mid-rostral or caudal by determining the midpoint of the atlas-defined dlPFC (middle frontal gyrus) in the y-dimension in MNI space for each subject. We then determined each contact's coordinates in MNI space and used the y-dimension coordinate to represent distance from the motor cortex, using the midpoint to assign labels of mid-rostral or caudal to each contact. From our exploratory spectral analysis, we found the most prominent response in the LF band occurred in the 2–6 Hz region between +200 and +700 ms relative to the cue. For beta, this was 12–30 Hz between +500 and +1,000 ms. We averaged across this frequency band to test for differences in LF power in the two subregions. We subsequently computed the true mean difference across subjects between the two conditions. To test whether the differences were significant, we used the same nonparametric permutation procedure described above. This process was done separately for the mid-rostral and caudal contacts.
Comparisons between the beta time courses of the various combinations of conditions for investigating whether it tracked motor activity were done using the same nonparametric permutation testing described above. This was also the case for comparisons between error and correct No-go trials in the LF and beta ranges and for comparing post-Go and post-No-go No-go trials LF time course.
For behavior–physiology relationships, we used linear mixed effects modeling (LMM) to determine whether there was a statistical relationship between LF power and Go trial reaction time. We extracted the mean LF power from the window described above and averaged across contacts and then modeled single trial reaction time as a function of LF power. We represented LF power as a fixed effect and incorporated a random intercept for subjects (R function lmer). We determined statistical significance of the fixed effect coefficient for LF power by computing an F statistic using Satterthwaite's approximation for denominator degrees of freedom (Luke, 2017; R function LmerTest). This process was repeated for beta power and reaction time as well.
To determine whether there was a statistical relationship between LF power and distance from the motor cortex on the anterior-posterior dlPFC axis, we again used an LMM. We used the above defined time–frequency window to extract average LF power for each trial and electrode. We then tested for the influence of distance on the difference in power between the two conditions by modeling power on condition, distance, and the interaction between the two while allowing for random slopes with subjects as the random intercept. We used the same approach described above to evaluate the significance of the fixed effect interaction coefficient. The LMMs for Power versus Block Type and Reaction Time versus Block Type were evaluated the same way. All statistical analysis was performed in R/MATLAB (4.2.1/2021b).
Results
We obtained unilateral ECoG recordings from either left or right dlPFC from 13 PD patients (10 males; mean age = 66.6 ± 3.4 years; 9 right, 4 left) while performing the task (Fig. 1A). The task consisted of a baseline period at rest, an LCD block of only Go trials, and an HCD block of Go and No-go trials with a Go:No-go ratio of 2:1. The average HCD Go reaction time (RT) across subjects was 543.9 ± 21.8 ms (Fig. 1B) and the average commission error rate (incorrect No-go trials) was 20.9 ± 3.2%. The trial numbers for each condition across all participants were HCD Go (701), LCD Go (298), and No-go (351). All participants had a higher response rate in the Go versus No-go conditions (p = 0.035; Pearson's chi-squared test; Fig. 1C), indicating adequate task understanding; therefore, no subjects were excluded based on behavioral criteria. To analyze oscillatory activity, we first excluded data from electrodes exhibiting excessive noise or other artifacts (3/78 electrodes, 3.9%; Fig. 1D). Given the lack of consensus boundaries for anatomic subregions of dlPFC, we implemented a boundary at the midpoint of the middle frontal gyrus and classified contacts as mid-rostral or caudal to that location (Fig. 1E). We calculated the length of the middle frontal gyrus and computed its mean to determine the midpoint within each participant (mean MNI y-coordinate = 47.17 ± 1.89 across subjects). We classified electrodes with MNI y-coordinates anterior to this midpoint as mid-rostral and posterior coordinates as caudal.
Go/No-go task structure and behavioral metrics. A, Top panel, Successful Go trial during which subjects are shown non-“X” letters and respond within 1,000 ms with a button press. Correct responses are indicated by green color change during the feedback period. Bottom panel, The No-go condition is illustrated in which subjects are shown the letter “X” and with prior instruction to withhold their response. An example of an incorrect response with the color of the letter changing to red is shown. Trial times ranged from 750 to 1,000 ms, titrated to 80% accuracy during a practice block. The intertrial interval ranged from 500 to 1,000 ms, with a subsequent feedback period of 250 ms. B, Histogram illustrating distribution of Go reaction times across subjects, with mean time of 543.9 ± 21.8 ms. C, Response rate rates for Go and No-go conditions across subjects showing significantly more responses during Go trials, indicating successful task understanding and completion (Pearson's chi-squared test, all 13 subjects responded more in Go than No-go, p = 0.035). D, Three different views of ECoG electrode coverage: right hemisphere (left), center view (middle), and left hemisphere (right). E, Locations of bipolar virtual electrodes in dlPFC reflected to right hemisphere for visual clarity, with mid-rostral electrodes labeled red-orange and caudal electrodes labeled as blue. White, translucent electrodes are those excluded either due to excess noise/artifacts or not being within the dlPFC as determined by the Desikan–Killiany atlas. The MNI y-coordinate used to classify mid-rostral versus caudal electrodes across subjects was 47.17 ± 1.89.
We then investigated whether spectral power differed between Go and No-go conditions along the rostral-caudal extent of the middle frontal gyrus. In the HCD block, we found increased LF power (2–6 Hz; p = 0.001; permutation test; Fig. 2A) and decreased beta power (12–30 Hz; p = 0.016; permutation test; Fig. 2A) during the post-cue period of No-go trials compared with Go trials, coinciding with mean response times during the Go trials. LF power during Go trials peaked at 572 ± 58 ms correlating with the EMG peak at 572 ± 37 ms, and beta power during Go trials peaked at 689 ± 77 ms. Similarly, No-go LF power peaked at 524 ± 61 ms and EMG confirmed absence of a motor response. Due to the possibility of No-go trials inducing LF oscillatory changes related to task switching, we compared the LF time course of No-go trials that followed Go trials with those that followed No-go trials and did not find a significant difference between these two conditions. Thus, it is unlikely that the LF power increases we see in No-go trials is due to task switching. To determine whether power differed between broadly defined mid-rostral and caudal regions, we repeated the analysis in electrodes grouped by region. We found no difference in LF power between conditions in the caudal contacts (p > 0.05; permutation test; Fig. 2B, left panel) but observed enhanced No-go LF power in the mid-rostral contacts compared with Go trials (p = 0; permutation test; Fig. 2B, right panel). We tested whether LF power depends on anatomic location employing a linear mixed effects model (LMM) examining the interaction between each electrode's MNI y-coordinate and (No-go vs Go) physiology. We found that LF power increased as a function of distance from the central sulcus along the gyrus y-axis (LMM; F(1,138) = 17.76; p = 4.5 × 10−5; Fig. 2C,D, Movie 1). Because we found condition-specific differences in beta power as well, we assessed whether these oscillations may also be distributed along a rostral-caudal gradient. We did not observe a significant gradient in this case, suggestive of a more diffuse cortical response in the beta range.
Response Inhibition increases low-frequency power in mid-rostral but not caudal dlPFC. A, Mean cue-aligned, z-scored power spectra in Go and No-go conditions with No-go–Go difference spectrogram. Dashed lines indicate the mean EMG signal across subjects for each condition, indicating successful withheld responses during No-go trials. Statistical significance indicated by black contour lines in the difference spectrogram; low-frequency (p = 0.001, permutation test, cluster corrected) and beta band power (p = 0.016, permutation test, cluster corrected) showed significance condition-specific differences. B, Left, Same as in A but narrowband low-frequency (2–6 Hz) power in caudal dlPFC electrodes showing no significant difference between conditions (p > 0.05, permutation test, cluster corrected). B, Right, The same as in B, left, but mid-rostral dlPFC electrodes showing a significant difference in low-frequency power between Go and No-go conditions (p = 0, permutation test, cluster corrected) indicated by the horizontal solid black bar and star. Vertical black dashed lines indicate the mean Go reaction time over all subjects included (528.7 ms for mid-rostral group and 540.1 ms for caudal group). C, Box plots (box length, IQR; whiskers, 1.5 × IQR; and transecting lines, medians) of distance quintiles versus low-frequency power predicted values from the LMM with regression lines for the Go and No-go conditions (n = 57 electrodes, 701 Go trials, and 351 No-go trials). D, Mean low-frequency power averaged over 200–700 ms mapped onto each electrode projected onto left hemisphere for visual clarity with cortical vertices taking on color values corresponding to weighted average of neighboring electrodes for Go trials (left), NoGo trials (middle), and difference between NoGo and Go (right).
We investigated potential links between beta power and motor activity by first comparing error trials to correct trials within conditions and did not find significant differences in the beta time courses either in the Go or No-go conditions (Fig. 3A,B). Next, we tested whether changes in condition could explain beta power when motor responses were held constant. This involved (1) comparing Go and No-go trials where a response was present in both (correct Go and error No-go) and (2) comparing Go and No-go trials where a response was absent in both (error Go and correct No-go). We found a significant difference in the former (p = 0.007; permutation test; Fig. 3C) but not the latter (p > 0.05; permutation test; Fig. 3D). Three out of the four tests did not support a motor-related phenomenon. The comparison between Go errors and correct No-go trials, conditions with distinct rules but identical response profiles (no response), however, could indicate a movement-related signal. However, in the original comparison between correct Go and No-go trials, the beta increase was observed in a condition that contained movement, contradicting the known phenomenon of cortical beta being involved with movement suppression. Motor signals are also typically recorded caudal (i.e., pre-motor/M1), whereas our recordings were considerably more rostral (dlPFC). Thus, while we cannot definitively rule out dlPFC beta being a motor signal, several lines of evidence support an alternative mechanism. We also analyzed the No-go errors to determine whether errors induced activity distinct from response inhibition but did not observe any differences in the LF or beta time courses.
Late-trial (postresponse) beta does not strictly track motor activity. A, 12–30 Hz power averaged across subjects (n = 12) for error (red) versus correct (blue) NoGo trials showing no significant difference (permutation test, cluster corrected, p > 0.05). B, Same as in A but for the Go condition (n = 10) with no significant difference (Permutation test, cluster corrected, p > 0.05). C, Comparing across conditions where a response was made showing significant beta suppression in error NoGo trials (red) compared with correct Go trials (blue; permutation test, cluster corrected, p = 0.007). D, Same is in C but for trials without a response showing no significant difference (permutation test, cluster corrected, p > 0.05).
Expecting the need to stop movements corresponds with increased reaction times (RTs), a mechanism thought to be a control strategy to increase inhibitory efficiency (Chikazoe et al., 2009b; Jahfari et al., 2010; Swann et al., 2013). We quantified this by comparing RTs between HCD Go trials and LCD Go trials and found longer reaction times in the HCD block compared with the LCD block (LMM; F(1,985) = 88.24; p = 2.2 × 10−16; Fig. 4A). We hypothesized that if this difference in RTs was supported by low-frequency oscillations (LFOs), HCD Go trials should have greater LF power than LCD Go trials. Indeed, we found a significant difference in LF power between Go trials of the two blocks (LMM; F(1,987) = 5.105; p = 0.02; Fig. 4B). We also reasoned that LF power might correlate with response delay and found a significant relationship between reaction time and power such that longer reaction times were associated with greater power (LMM; F(1,988) = 18.47; p = 1.9 × 10−5; Fig. 4C). We repeated this analysis for beta power did not find a relationship with response delay (LMM; F(1,899) = 0.1362; p > 0.05; Fig. 4D).
dlPFC low-frequency power is a substrate for proactive control. A, Box plots (box length, IQR; whiskers, IQR × 1.5; and transecting lines, medians) of response time versus block type across all 13 subjects (LMM; F(1,985) = 88.24; p = 2.2 × 10−16; n = 701 cognitive Go trials and 298 motor Go trials). B, Same as in A but low-frequency (2–6 Hz) power versus block type. C, Box plots (box length, IQR; whiskers, IQR × 1.5; and transecting lines, medians) of predicted reaction times from the LMM versus low-frequency power quintiles (LMM; F(1,987) = 5.105; p = 0.02; n = 999 Go trials). D, Box plots (box length, IQR; whiskers, IQR × 1.5; and transecting lines, medians) of predicted beta power from the LMM versus reaction time quintiles (LMM; F(1,899) = 0.1362; p > 0.05; n = 999 Go trials).
Since we observed temporally distinct changes in LF and beta power, we investigated whether the two frequencies might interact during cognitive control. As a validation step, we first employed MVPA on the LF power time courses from each channel for all subjects (train and test within LF band). The MVPA discriminated between Go and No-go trials significantly above chance level between 500 and 625 ms post-cue (AUCmean = 0.584; p = 0.036; permutation test; Fig. 5A). We then repeated within-frequency training and testing using the beta band. The MVPA was again able to distinguish between trial types above chance level between 625 and 700 ms (AUCmean = 0.541; p = 0.040; permutation test; Fig. 5B). Finally, we investigated whether LF activity could predict beta activity by training on LF time courses and testing on beta. We did not find any time points with significantly above chance level discrimination between trial types (Fig. 5C).
dlPFC low-frequency and beta oscillations do not interact during cognitive control. A, MVPA within low-frequency power (train and test) showing significantly above chance condition discrimination between 500 and 625 ms (AUCmean = 0.584; p = 0.036; permutation test, cluster corrected). B, Same as in A, but for beta power, showing significantly above chance condition discrimination between 625 and 700 ms (AUCmean = 0.541; p = 0.040; permutation test, cluster corrected). C, MVPA trained on low-frequency power and tested on beta power showing no significant condition discrimination at any time point (permutation test, cluster corrected, p > 0.05).
Animation of Figure 2D mean z-scored low-frequency power difference from 500 ms before cue onset to 1,000 ms after cue onset. Note the early prominence in mid-rostral dlPFC. [View online]
Discussion
Functional imaging studies show that dlPFC coactivates with the action stopping circuit during tasks involving response inhibition (Garavan et al., 1999; Menon et al., 2001; Nakata et al., 2008; Chikazoe, 2010), but the BOLD signal lacks the temporal resolution for tracking fast fluctuations in LF activity. Our results indicate that dlPFC LF power increases when inhibiting a prepotent response and peaks shortly before the response time. Thus, dlPFC LF activity may serve an antikinetic role in action stopping similar to mPFC theta (Zavala et al., 2018). Some evidence indicates cognitive control exhibits frequency specificity within the delta-theta range: response inhibition appears to be driven mostly by delta (∼2–4 Hz) power (Zavala et al., 2018; Kaiser et al., 2019), whereas theta (∼4–8 Hz) may be a more general control substrate (Cavanagh and Frank, 2014). Our data support this framework and extend its relevance to the lateral PFC as the 2–4 Hz frequency range contained the bulk of the inhibition-related activity in our task, with some involvement up to 6 Hz. Collectively, these findings also suggest that the dorsolateral and medial PFC may coordinate stopping behaviors via LFOs. Such medial-lateral interactions have been observed during conflict monitoring (Hanslmayr et al., 2008; Cohen and Cavanagh, 2011), and LFOs are known to coordinate the timing of neuronal firing across the PFC (Smith et al., 2019). Our findings add to this body of work by (1) demonstrating the role of dlPFC LFOs in response inhibition, without the explicit presence of conflict, and (2) providing a potential mechanism by which specific neuronal populations are preferentially recruited in a task-specific manner through focal increases in LF power.
dlPFC/lPFC activity does not change during error commission as indexed both by EEG theta power (Cavanagh et al., 2009) and the BOLD signal (Menon et al., 2001). However, dlPFC is involved in error processing as lPFC and mPFC synchronize during errors (Cavanagh et al., 2009), and dlPFC lesions attenuate the cingulate-associated error related negativity (Gehring and Knight, 2000). Our results were therefore consistent with previous reports demonstrating the lack of an error signal in dlPFC.
Cognitive control occurs at multiple timescales and is often grouped into proactive and reactive categories (Braver, 2012). Proactive control may involve the mPFC, in which theta fluctuates depending on expectations of control demands (Messel et al., 2021). Trials requiring increased cognitive control exhibit greater theta power, especially when the relative proportion of these trials is smaller (Dippel et al., 2016, 2017; Chinn et al., 2018). This suggests that mPFC theta coordinates the allocation of control resources, adapting to changing demands. dlPFC activity has been linked to increasing top-down control (MacDonald et al., 2000) and cue-induced preparations to inhibit a response (Swann et al., 2013). In this study, we show longer RTs in the HCD versus LCD block. This may be an adaptive process to prepare for the need to inhibit a response. Due to the relationship between RTs and LF power, we propose dlPFC LFOs may represent an ongoing modulation of inhibitory tone based on task demands.
One important question is whether dlPFC is functionally homogeneous. Structural and functional connectivity studies suggest that this is not the case, as resting-state connectivity between dlPFC and other brain regions and canonical networks exhibits a gradient along the dlPFC rostro-caudal axis (Jung et al., 2022), and similar findings have been observed during executive control (Cieslik et al., 2013). A growing body of work indicates that lateral PFC is hierarchically organized, with evidence suggesting that the organizing framework may be either (1) level of rule abstraction (Badre and D’Esposito, 2007, 2009; Badre, 2008; Badre et al., 2009) or (2) task difficulty (Crittenden and Duncan, 2014), with increasingly rostral activity associated with greater abstraction and difficulty, respectively. Our data support this hypothesis, as we found LF activity associated with a difficult task (inhibiting a prepotent response) occurring prominently in the mid-rostral subregion of dlPFC. Granular establishment of dlPFC structure and function has important implications such as honing stimulation targets for neuropsychiatric disease.
We also observed differences in beta power in the postresponse period between conditions, although, without a gradient. Interestingly, a similar study performed in PD patients with recordings over the mPFC found no such differences between Go and No-go trials (Zavala et al., 2018). Beta oscillations have numerous proposed functions: from clearing out working memory to maintaining cognitive or motor sets (Engel and Fries, 2010; Schmidt et al., 2019). In the Go/No-go task, participants navigate trials and establish a motor set (respond to stimuli) due to the increased proportion of trials that require a response (Go) compared with those that require withholding a response (No-go). This is intentionally built into the task to create a prepotent tendency to respond, making inhibiting a response difficult. While observing beta changes between conditions with different motor profiles might suggest a simple movement mechanism, our results do not support this interpretation. We found no effect due to the presence of the motor response on the beta activity within conditions. Yet, beta differences were noted between conditions when only evaluating trials that contained a response. Furthermore, our recording sites were not in the typical motor areas, and cortical beta is usually suppressed when making movements, which is inconsistent with our data. One alternative explanation is that beta power increasing after Go trials in our study may reflect dlPFC maintenance of the current motor set with beta suppression following No-go trials signaling a departure from the “status quo” (Engel and Fries, 2010). Thus, dlPFC beta power may function as a mechanism for planning movements in response to pending stimuli.
Several reports have shown that theta and alpha band activity can share information and interact in various cognitive functions (Prochnow et al., 2022a; Pscherer et al., 2023; Wendiggensen et al., 2023). We observed temporally distinct patterns of activity in No-go trials compared with Go and investigated whether these two frequency bands interacted. Our original framework proposes LF activity in the dlPFC as a motor inhibition signal and beta activity as a mechanism for navigating motor sets. The interaction of these two frequencies would potentially suggest a more complex system. Based on the lack of interplay, we propose LF and beta activity in the dlPFC reflect distinct, parallel cognitive processes enabling goal-directed behavior.
Inhibitory control constructs are often studied within Stop Signal and Go/No-go paradigms. The Go/No-go task is traditionally used to investigate response inhibition by creating a prepotent response tendency, making the No-go trials indicative of reactive inhibition; although No-go cues can involve other cognitive processes like working memory and goal maintenance (Redick et al., 2011). Reactive control in this context is driven by external cues, requiring participants to inhibit their response when presented with a No-go stimulus. Conversely, proactive control involves prestimulus anticipation and preparation to inhibit responses. In the Go/No-go task, it has been investigated by adding prestimulus cues informing the participant about the nature of the upcoming stimulus (Benedetti et al., 2020; Talanow et al., 2020). Here, we demonstrate a proactive process arises even in the absence of such cues. Participants learned to predict future needs to stop based on the preblock instructions and their experience navigating the block supporting proactive inhibitory control in the absence of a “stop” stimulus (No-go). On No-go trials, the LF response may have both proactive and reactive inhibitory control components. This is difficult to disentangle as there is overlap between their neural underpinnings (Aron, 2011). Stop trials in the Stop Signal task might better indicate reactive inhibition compared with No-go trials in a typical Go/No-go task, and these two tasks have differing neurophysiologic profiles (Raud et al., 2020). However, the parameters in our task are similar to those known to induce prepotent motor activity (Wessel, 2018), and reactive inhibition is required to overcome it. The Go/No-go task can be viewed as a hierarchical task composed of two simpler ones (the Go condition being the “respond-to” task and the No-go condition being the “do-not-respond-to” task). Thus, one possibility is that differences between correct Go and No-go trials may be due to task switching; however, our analyses do not support this interpretation.
Regarding our anatomic analyses, consensus is lacking on how to define boundaries within the dlPFC. We designated rostral/caudal labels using the mid-point of the rostral MFG. While this is an arbitrary definition of the rostro-caudal axis, it attempts to quantify differences along the dlPFC in a reproducible manner and provides useful information about heterogeneity in prefrontal activity. Our study was performed intraoperatively which has high spatiotemporal resolution but also creates some limitations such as time and patient fatigue. Experimental factors are also challenging to control, although we took steps to maximize consistency.
Footnotes
This work was supported by NIH NINDS (5K23NS117735).
The authors declare no competing financial interests.
- Correspondence should be addressed to J. Nicole Bentley at nbentl{at}uab.edu.
