Neural dynamics of proactive and reactive cognitive control in medial and lateral prefrontal cortex

Summary

Goal-directed behavior requires adjusting cognitive control, both in preparation for and in reaction to conflict. Theta oscillations and population activity in dorsomedial prefrontal cortex (dmPFC) and dorsolateral PFC (dlPFC) are known to support reactive control. Here, we investigated their role in proactive control using human intracranial electroencephalogram (EEG) recordings during a Stroop task that manipulated conflict expectations. During response selection, conflict processing enhanced dlPFC beta desynchronization, dmPFC theta increases, and high-frequency activity (HFA, which indexes local population activity) in both regions. After responses, conflict suppressed theta and boosted beta rebounds in both regions. Importantly, pre-trial dmPFC theta increased when conflict was anticipated, and within-trial theta, beta, and HFA dynamics were accentuated when conflict was rare. These findings reveal how the balance of reactive and proactive control modulates shared HFA and dissociable theta-beta conflict signals in dmPFC and dlPFC and identifies pre-trial dmPFC theta as a candidate substrate for proactive control.

Graphical abstract

Highlights

• •Pre-trial dmPFC theta increases when conflict is anticipated, reflecting proactive control
• •Conflict enhances population activity in dlPFC and dmPFC early while suppressing theta late
• •dmPFC theta increases and dlPFC beta drops were greater during conflict
• •Proactive control attenuated reactive control signals during conflict processing

Health sciences; Medicine; Medical specialty; Surgery; Neurology; Natural sciences; Biological sciences; Neuroscience; Clinical neuroscience

Introduction

Goal-directed behavior requires exerting cognitive control to resolve conflict between competing options, and control resources are strategically adjusted based on recent experience and future conflict expectations.¹ Extensive research in humans and non-human primates supports the proposal that dorsomedial prefrontal cortex (dmPFC) detects the need for control and recruits dorsolateral prefrontal cortex (dlPFC) to implement cognitive control,^{2,3,4,5,6,7,8,9,10} with theta oscillations and local population activity enhanced in these key regions during conflict processing.^6,7,8,10,11 However, the mechanisms facilitating cognitive control adjustments in preparation for conflict remain unclear.

The dual mechanisms framework of cognitive control posits separate but complementary processes of proactive and reactive control.¹² Proactive control involves preparing cognitive resources for top-down control of behavior, whereas reactive control is triggered in response to control-demanding stimuli, such as conflict between automatic word reading and instructed color naming in the Stroop task.¹³ Reactive control signals in dmPFC and dlPFC are well studied during within-trial conflict processing, where local population activity in the dlPFC and dmPFC increases, as measured using fMRI BOLD signal,^14,15 single unit firing,^5,10,16 and intracranial electroencephalogram (iEEG) high-frequency activity (HFA),^7,8 which is a proxy of local multi-unit activity.^17,18,19 Conflict also increases theta oscillations in dmPFC, but not dlPFC,^6,8 which is proposed to reflect dmPFC’s role in detecting conflict and recruiting control resources.^1,2,20,21,22 Recent work also shows theta oscillations coordinate neural firing in human dmPFC and dlPFC during conflict, and this was also true for beta oscillations, which are proposed to gate information processing by increasing or decreasing to allow reinforcement or modification of neural states, respectively¹⁰ (see previous comprehensive reviews on theta^23,24 and beta^25,26,27 oscillations). Moreover, theta, beta, and population activity signals are linked to adaptations on trials after conflict, which can affect speed and accuracy on subsequent trials (i.e., the Gratton effect^5,8,9,28,29). However, these between-trial adaptations reflect a mixture of reactive control from the previous trial and proactive control preparations.^30,31,32 In sum, theta and population activity in dmPFC and dlPFC are critical control signals elicited by conflict that are implicated in short-timescale adaptations, but few studies have examined their role in proactive control.

Behavioral studies have isolated proactive control by increasing the frequency of difficult trials, which report performance improvements when conflict is more common (e.g., reduced conflict effects on reaction times [RTs]).^{29,33,34,35,36} The dual mechanisms framework hypothesizes that stronger conflict expectations drive anticipatory allocation of cognitive resources,^12,30 and several scalp EEG studies have shown control-demanding trials elicit lower mid-frontal theta power when these trials are more frequent, consistent with reduced need for reactive, within-trial control signals.^29,35,36 One recent study showed pre-trial dmPFC and dlPFC single unit firing predicted conflict-related RTs, suggesting preparatory activity can influence conflict processing.³⁷ However, no study to date has elucidated a candidate mechanism for proactive control adaptations based on task demands, leaving the roles of dmPFC and dlPFC undefined.

Here, we address these issues by utilizing the high spatiotemporal resolution of iEEG recordings from human dmPFC and dlPFC during a Stroop task that manipulated conflict frequency across blocks. We hypothesized that stronger conflict expectations would reduce conflict effects on RTs and reactive neural activity in the dmPFC and dlPFC, and we also predicted that theta oscillations in dmPFC would track these proactive control adaptations. We first re-examined dmPFC and dlPFC activity during within-trial conflict processing, confirming canonical conflict-driven increases in dmPFC theta and HFA in both dmPFC and dlPFC, as well as providing additional insights into a shared post-response theta suppression and dissociable effects of conflict on beta power across regions. Manipulating conflict expectations uncovered an increase in pre-trial dmPFC theta power when conflict was common, revealing a potential neural substrate for proactive cognitive control. These results refine and expand the role of theta, beta, and local population activity in reactive and proactive control and elucidate the dissociable roles of dmPFC and dlPFC during goal-oriented behavior.

Results

Behavioral signatures of proactive and reactive control

iEEG activity in dmPFC and dlPFC was recorded while 23 participants undergoing neurosurgical treatment for epilepsy performed a color-word Stroop task. Conflict expectations were manipulated by varying the percentage of congruent (%Cong) versus incongruent trials across blocks, but all blocks had a constant proportion of neutral trials (“XXXX”), allowing assessment of proactive control on post-neutral trials while controlling for previous trial type and frequency. In this design, 16%Cong blocks (highest number of conflict trials) contained the highest conflict expectation, with 33%Cong blocks being the next highest, and 50%Cong blocks consisting of the lowest conflict expectation. Post-neutral trial analysis was performed when measuring dmPFC theta power between trials as a substrate of proactive control. Apart from this instance, all trials excluding the first trial from each block were included in all analyses unless otherwise specified (e.g., in post-hoc pairwise comparisons in the behavior linear mixed effects model [LMM]). We used LMMs to predict RTs based on task variables, revealing a main effect of current trial type. RTs were slower for Conflict (i.e., incongruent) trials, while congruent and neutral trials were not different and were thus combined into a NoConflict condition for subsequent analyses (see STAR Methods) (n = 23 subjects, 5,888 trials, LMM, t₅₈₆₀ = 35.5, p = 2 × 10⁻¹⁶; Figure 1D). We also found several between-trial adaptation effects to previous conflict, including post-conflict slowing (Figure S1), but our primary focus was on proactive adaptations to %Cong conditions. We observed that NoConflict RTs were faster when conflict was rare (main effect of %Cong, t₅₈₆₀ = −3.76, p = 0.002), and that the effect of conflict on RTs was larger when conflict was more surprising (CurrentConflict:%Cong interaction, t₅₈₆₀ = 5.11, p = 3.37 × 10⁻⁷; Figure 1E). In summary, the slowing effect of CurrentConflict was reduced in low %Cong blocks, indicating proactive control reduces the classic reactive Stroop effect when conflict is expected.

Figure 1.

Task, behavior, and electrode locations

(A) Trial types, %Cong block types, and trial design.

(B) Group-level reconstructions of recording sites for dmPFC and dlPFC. All sites mirrored to left hemisphere for visualization.

(C) RT distributions for Conflict and NoConflict trials.

(D) Stroop effect: RTs are longer on Conflict than NoConflict trials.

(E) Stroop effect (mean RT difference for Conflict-NoConflict) is larger when conflict is rare (50%Cong) than when conflict is common (16%Cong), indicating proactive control adaptation based on conflict expectations. ∗ indicates statistical significance at a = 0.05. Data are represented as mean ± SEM.

Theta, beta, and HFA dynamics in dmPFC and dlPFC during within-trial conflict processing

To first establish the neural dynamics underlying reactive conflict processing within a trial, we used time-resolved LMMs to predict single-trial neural power in theta, beta, and HFA bands (Figure 2). In line with previous conflict studies,^{6,8,9,24,29,38,39} theta power was higher during response preparation in Conflict than NoConflict trials in dmPFC (LMM, all p ≤ 0.022) but not dlPFC (Figure 2C). Similarly, dmPFC and dlPFC both showed HFA increases just before and during the response for Conflict relative to NoConflict trials (LMM, dmPFC: all p = 0, dlPFC: all p ≤ 0.006; Figure 2B). Both regions exhibited pre-response beta suppression, but only dlPFC showed a greater beta decrease on Conflict than NoConflict trials (LMM, all p ≤ 0.008; Figure 2D), reflecting additional processing of conflict. Finally, dlPFC theta power showed a peri/post-response suppression in Conflict relative to NoConflict trials (LMM, p ≤ 0.001; Figure 2C), which also was observed for dmPFC theta after the response (LMM, p ≤ 0.004; Figure 2C). In summary, we observed a cascade of shared HFA but dissociable theta and beta dynamics in dmPFC and dlPFC, where conflict initially increased dmPFC theta (∼300 ms post-stimulus) and dlPFC HFA (∼290 ms post-stimulus), then dmPFC HFA increased (∼420 ms post-stimulus) before a late suppression of theta in dlPFC (∼650 ms post-stimulus/∼450 ms pre-response) then dmPFC (∼24 ms post-response). However, the timing of this sequence of effects is approximate because deriving precise onset latencies from cluster-based statistics is problematic.^40,41 Similarly, stimulus-locked neural signals should be interpreted with caution in later windows where response-locked activity can be confounded by RT differences across conditions and participants, including conflict modulations of beta power desynchronizations (∼526 ms pre-response in dlPFC) and rebounds (∼5 ms pre-response in dmPFC, ∼424 ms post-response in dlPFC).

Figure 2.

Conflict drives similar HFA and diverging theta and beta dynamics in dmPFC and dlPFC

(A) Grand average time-frequency representations for dmPFC (left pair) and dlPFC (right pair) task encoding electrodes for both stimulus aligned (left column of each region’s panel) and response-aligned (right column of each region’s panel) data. Contour lines indicate power values significantly different from baseline (a = 0.05).

(B) Time courses for HFA on Conflict (red) and NoConflict (blue) trials.

(C) Analogous time courses for theta power.

(D) Analogous time courses for beta power. Horizontal black lines indicate significance for CurrentConflict (a = 0.05). Vertical dashed lines indicate group mean RT in stimulus-aligned data and RT in response aligned data. Data are represented as mean ± SEM.

Conflict expectations modulate within-trial conflict processing and increase pre-trial theta power in dmPFC

To determine the neural basis of proactive control, we investigated how within-trial conflict responses were modulated by conflict anticipation. We found that theta and HFA responses in both regions were accentuated when conflict was rare and attenuated when conflict was expected (LMM, CurrentConflict:%Cong interaction, dmPFC: all p ≤ 0.02, dlPFC: all p ≤ 0.023; Figures 3A–3D). However, the nature of these modulations varied across regions and frequency bands. In the theta band, we found an early dissociation where stronger expectations of conflict diminished amplitudes in dmPFC but not dlPFC (CurrentConflict:%Cong interaction, LMM, all p ≤ 0.011; Figures 3C and 3D). In contrast, both regions showed a larger effect of conflict on HFA when conflict was less expected. In the later post-response window, both regions showed a greater suppression of theta in blocks when conflict was rare (CurrentConflict:%Cong interaction, dmPFC: all p ≤ 0.012, dlPFC: all p ≤ 0.016; Figures 3C and 3D), suggesting theta during response monitoring is downregulated throughout the network when conflict resolution was more demanding. Pre-response beta power in dlPFC (but not dmPFC) was suppressed more for conflict trials when it was more surprising (i.e., less proactive control available) (CurrentConflict:%Cong interaction, all p ≤ 0.022; Figures 3E and 3F, see response-locked data in inset).

Figure 3.

Proactive control modulates neural conflict signals

(A) dmPFC HFA in Conflict (red) and NoConflict (blue) trials across 3 block types, with conflict effects increasing from highest to lowest proactive control (left to right: 16%, 33%, and 50% congruent).

(B) Same as in (A) but for dlPFC.

(C and D) Same as (A and B), but for theta power.

(E and F) Same as above, but for beta power. Horizontal black lines indicate significance for the CurrentConflict:%Cong interaction (a = 0.05). Vertical dashed lines indicate group mean RT. Insets show response-locked data (Figure S3 for full response-locked results), which clearly show the modulation by expectation is before the response. Data are represented as mean ± SEM.

Finally, we tested our hypothesis that dmPFC theta facilitates these proactive control effects by examining %Cong block effects on neural activity in the preparatory period from −1 s to stimulus onset. Notably, we avoided confounds of previous trial conflict and frequency by constraining this analysis to trials preceded only by neutral trials (see STAR Methods). This revealed elevated dmPFC theta power approximately 0.5 s before stimulus onset in blocks when conflict was more frequent (LMM, p ≤ 0.023, permutation test; Figure 4). In summary, higher proactive control reduced within-trial conflict dynamics and increased pre-trial dmPFC theta.

Figure 4.

Proactive control increases pre-trial dmPFC theta power

Time courses for stimulus-aligned, post-neutral trial theta power in dmPFC for the 16% Congruent block (green) and 50% Congruent block (purple). Black line indicates significance for the %Cong effect (a = 0.05). Data are represented as mean ± SEM.

Discussion

We used iEEG in human dmPFC and dlPFC to investigate the neural dynamics of proactive and reactive cognitive control by manipulating conflict expectations in a Stroop task. Single-trial modeling of neural power revealed a sequence of within-trial conflict processing dynamics that differentiated dmPFC and dlPFC. In line with previous findings,^{6,7,8,9,24,29,38,39} conflict triggered an increase in dmPFC theta and HFA increases in both regions. Building on these classic effects, we report additional effects of conflict, which increased a pre-response beta desynchronization in dlPFC and a post-response beta rebound in both regions, as well as suppressing theta in both regions after responses. Furthermore, these neural signals were accentuated when conflict was rare and diminished when conflict was common and thus expected, demonstrating how these within-trial dynamics are modulated by the balance between proactive and reactive control. Importantly, we also found that pre-trial theta power in the dmPFC was enhanced during blocks with strong expectation of conflict, revealing a neural signature of proactive control. Collectively, these results characterize how the cascade of neural signals in dmPFC and dlPFC during conflict processing are modulated by proactive and reactive control.

Our findings on the within-trial sequence of dmPFC and dlPFC neural signals underlying conflict processing corroborate and extend prior findings implicating these regions in cognitive control. Increases in HFA in both dmPFC and dlPFC during conflict suggest local neuronal populations within these regions are recruited to resolve competition between responses, aligning with earlier reports of dmPFC and dlPFC involvement in conflict detection and resolution.^{2,6,7,8,10,14} However, the oscillatory dynamics in these regions diverged, providing insight into how dmPFC and dlPFC differentially contribute to conflict processing across different timescales. Consistent with classical findings,^2,5,14,23 dmPFC showed robust increases in theta power during the response preparation epoch of Conflict compared to NoConflict trials, highlighting its role in conflict monitoring and signaling the need for control. In contrast, conflict triggered a strong post-response suppression of dlPFC theta power below baseline. A similar but weaker and later theta suppression after conflict was observed in dmPFC. This finding suggests that theta oscillations may be involved in resetting or downregulating control circuits after a demanding event, which may help minimize control costs via efficient resource allocation. Overall, these findings suggest a temporally structured cascade of conflict processing across dmPFC and dlPFC. According to theories of cognitive control,^1,14,20,21 the contemporaneous rise of dmPFC theta and dlPFC HFA with conflict may reflect their concurrent roles in detection of conflict (dmPFC) and engagement of control processes (dlPFC), whereas the subsequent increase in dmPFC HFA potentially reflects ongoing conflict and performance monitoring. The tapering of these canonical dynamics coincides with the timing of theta suppression in each region. Taken together, these dynamics exemplify how dmPFC and dlPFC handle distinct roles in the cognitive control network.

We also observed dissociable effects of conflict on beta power across dmPFC and dlPFC. The pre-response decrease in beta power was stronger for Conflict than NoConflict trials in dlPFC but not dmPFC, whereas the post-response beta rebound increased on Conflict trials in both regions. Substantial research has linked dlPFC to maintenance and updating of working memory supporting task demands,^14,42,43,44 where the beta rhythm is proposed as a gating mechanism in these circuits with decreases in beta power opening a window for updating.^27,44,45 Accordingly, the greater decrease in dlPFC beta on conflict trials may reflect additional processing demands required to exert control during conflict resolution.^6,20 For example, a previous EEG study showed that lateralization of beta power suppression over motor cortices tracked the evolution of decisions between left and right responses, and that these motor preparatory signals were modulated by current trial conflict and previous trial errors.⁴⁶ Thus, our observation that conflict enhances pre-response beta desynchronization in dlPFC but not dmPFC could be explained by increased top-down control demands for dlPFC to prepare and coordinate color naming and word reading responses in premotor and motor regions. In contrast, previous studies implicate post-response beta increases in regulating the integration of surprising feedback into an internal model of motor commands, where stronger beta increases indicate confidence in maintaining the current motor plans. This suggests that the increased beta rebound in both regions after correct conflict responses may reflect reinforcement of the task rules. This effect was earlier in dmPFC than dlPFC, which aligns with proposals that dmPFC monitors control demands and relays this information to dlPFC.²⁰

After establishing these theta, beta, and HFA dynamics during within-trial conflict processing, we addressed our main objective by examining how these dmPFC and dlPFC signals were modulated by proactive control as conflict expectations build up over time. Behaviorally, participants had a reduced Stroop effect when conflict was more expected, and as hypothesized, these different levels of proactive control revealed a bidirectional influence on neural conflict processing. Canonical increases in dmPFC theta and HFA in dmPFC and dlPFC on Conflict trials were exacerbated for high %Cong blocks but nearly absent in blocks when half of trials had conflict. Similarly, the suppression of pre-response dlPFC beta and post-response dmPFC and dlPFC theta on Conflict trials was larger after surprising conflict and showed a smaller effect after common conflict. Accentuated beta decreases in dlPFC during surprising conflict trials support interactions between reactive control and working memory, possibly reflecting larger working memory updates for flexible reconfiguration when conflict is rare. It remains unclear whether this effect reflects reactive, within-trial adjustments to the current conflict or facilitates updating conflict expectations on future trials. However, our findings underscore a key role for dlPFC beta in enabling goal-oriented behavior over block-level timescales, which align with and extend previous studies showing modulation of beta response dynamics according to errors and difficulty on previous trials.⁴⁶ Overall, this striking demonstration that theta, beta, and HFA conflict signals are exaggerated or attenuated when proactive control is minimal or maximal, respectively, refines the interpretation of these dynamics by linking them more closely to reactive control mechanisms.

Critically, we found that higher conflict expectations increased dmPFC theta power before trial onset, suggesting dmPFC theta may facilitate proactive control by preparing cognitive resources based on expected demands. Notably, this analysis was limited to post-neutral trials, and our task design held the rate of neutral trials constant across blocks, meaning this effect cannot be explained by reactive control adjustments to prior trial frequency or difficulty. Interestingly, a prior EEG study showed that lower mid-frontal theta predicted omission errors,⁴⁷ though they could not determine whether this effect was due to proactive versus reactive control. However, whether preparatory theta increases improve performance remains unclear, as pre-trial dmPFC theta did not predict trial-by-trial RTs. Since our task design did not cue participants to the proportion congruence levels, the proactive control effect in our data may reflect more implicit expectations accumulated across the block. Future studies will be needed to test the behavioral relevance of pre-trial theta using tasks with stronger expectation manipulations. Such experiments need to control for reactive mechanisms, as we also found that post-conflict slowing predicted reduced early dmPFC theta (Figure S2), which may reflect a reduced need to recruit control resources due to prior recruitment on previous Conflict trials. Collectively, these results implicate dmPFC theta in both between-trial and block-level control adaptations.

Conclusions

In summary, our results shed light on how dmPFC and dlPFC process conflict, showing how this sequence of neural signals is modulated by varying levels of proactive control. Examining within-trial effects of conflict showed that canonical theta and HFA increases with conflict were followed by a suppression of theta after the response in both dmPFC and dlPFC, potentially reflecting downregulation in control networks after high demand trials. We also report pre-response beta desynchronization in dlPFC and post-response beta rebound in dmPFC that align with the roles for these regions in exerting and monitoring control, respectively. Moreover, by varying task demands over long timescales, we show that stronger conflict expectations boost dmPFC theta prior to trial onset while also reducing the cascade of within-trial conflict effects. Overall, these findings identify a putative mechanism by which proactive control improves performance by preparing cognitive resources to reduce reliance on within-trial reactive control.

Limitations of the study

The findings presented here should be interpreted with consideration of certain limitations inherent to iEEG studies. First, the patient population comprised individuals undergoing surgical treatment for epilepsy, which can be associated with cognitive impairments that may affect task performance. However, 79% of epilepsy patients have limited or no cognitive impairments,⁴⁸ and negative effects of epileptic activity on cognition are confined to the ∼1 s prior to stimulus onset,⁴⁹ meaning our exclusion of epochs containing epileptiform discharges and artifacts limits the impact of epilepsy-related activity on our results. Additionally, two patients in this study were non-native English speakers with IQ scores above 85, and both these participants showed normal behavior (e.g., RT modulation by conflict). In summary, while our study used standard iEEG methods to obtain insights into neural mechanisms of cognitive control,⁵⁰ but these results should be contextualized within the specific clinical population studied.

Although our sample size is similar to, and in general, larger than previous iEEG studies, our study has statistical power limits due to the rare nature of these data. Thus, these findings should be replicated in future studies to confirm their robustness and generalizability. However, the single-trial reliability of iEEG also addresses the data reliability since significance can be obtained in individual subjects.

Resource availability

Lead contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Anas Uddin Khan (anaskhan@uab.edu).

Materials availability

This study did not generate new unique reagents.

Data and code availability

Preprocessed data have been deposited on Zenodo and are publicly available as of the date of publication at Zenodo Data: https://doi.org/10.5281/zenodo.15795023. All original code has been deposited on GitHub and is publicly available at https://github.com/bentleylab/PRJ_Stroop as of the date of publication.

Acknowledgments

This work was funded by NINDS R01NS021135, NIMH CONTE Center P50MH109429, NIMH F32MH132174, Wellcome Discovery Award (226645/Z/22/Z), NINDS 5K23NS117735, and NSF GRFP.

Author contributions

A.U.K., C.W.H., K.L.A., V.P., and R.T.K. conceptualized the study. C.W.H., J.J.L., R.T.K., and J.N.B. acquired funding. C.W.H., K.L.A., and V.P. were involved in investigation. A.U.K., C.W.H., and K.L.A. developed methodology and performed formal analysis. D.K.S., K.D.L., P.W., J.J.L., R.T.K., and J.N.B. contributed resources. C.W.H., R.T.K., and J.N.B. provided supervision. A.U.K. and C.W.H. wrote the original manuscript. A.U.K., C.W.H., K.L.A., V.P., D.K.S., K.D.L., P.W., J.J.L., R.T.K., and J.N.B. reviewed and edited the manuscript.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Software and algorithms
FieldTrip Toolbox	Oostenveld et al.51	http://www.fieldtriptoolbox.org/
SPM12	Penny et al., 2011	https://www.fil.ion.ucl.ac.uk/spm/software/spm12/
MATLAB 2023b	MathWorks Inc.	RRID:SCR_001622
PsychoPy	Peirce et al.52	https://www.psychopy.org/
Freesurfer 5.3.0	Dale et al.53	https://surfer.nmr.mgh.harvard.edu/
Zenodo Data	https://doi.org/10.5281/zenodo.15795023

Experimental model and study participant details

Data was collected from 31 surgical epilepsy patients (mean ± SD = 35.7 ± 12.4 years old; 8 women; see Table S1). However, patients were excluded for technical problems during recordings (n = 4), poor data quality (n = 2), or lack of dmPFC or dlPFC coverage (n = 2). The 23 patients available for analyses were implanted with stereo-electroencephalography (SEEG; n = 21) probes and/or subdural electrocorticography (ECoG; n = 3) grids/strips, which were determined solely based on clinical decisions in the best interest of the patient without regard to research. Patients were studied at the University of California, Berkeley, University of California, Irvine (UCI), and California Pacific Medical Center. All patients had normal IQ and spoke English as a primary language except two, which were both fluent in English with IQ above 85. Race and ethnicity demographics for individual participants in this study from UCI are not available, but the general population recruited from UCI over multiple studies funded by NINDS R01NS021135 was 52% White, 20% Asian, 16% Black or African American, 4% American Indian or Alaskan Native, and 8% unknown or not reported, with 45% are Hispanic or Latino/Latina, 50% are not Hispanic or Latino/Latina, and 5% are unknown or not reported. The participants recruited from California Pacific Medical Center in this study were both White, with one Hispanic or Latino/Latina and one not.

Patients performed a color-word Stroop task written in PsychoPy⁵² (v1.82.01) with congruent, incongruent, and neutral stimuli consisting of the words “BLUE”, “RED”, “GREEN”, and “XXXX” (neutral) displayed in blue, red, or green ink on a gray computer screen (Figure 1A). Color-word stimuli were presented for 1.5 s followed by inter-trial intervals randomly sampled from a uniform distribution between 1.05 and 1.8 s, except for one patient who had inter-trial intervals ranging from 2 to 2.3 s (see Table S1). Patients were asked to name the color of the ink, and verbal RTs were obtained from the microphone signal by manually marking the first deflection in the audio waveform above baseline noise for vocalizations naming the color. The task was organized into 9 blocks of 36 trials, except one patient who had 12 blocks of 24 trials (see Table S1). One patient performed the task twice. The proportion of congruent trials was manipulated across blocks such that they were either mostly congruent (50% congruent, 33.3% neutral, and 16.7% incongruent), equal proportions (33.3% congruent, 33.3% neutral, and 33.3% incongruent), or mostly incongruent (16.7% congruent, 33.3% neutral, and 50% incongruent; Figure 1A). Note that the consistent frequency of neutral trials across block types provides an unbiased diagnostic condition of block-level changes in conflict expectations. Block types were randomized and occurred in equal numbers.

Method details

Data were recorded at either the University of California (UC) Irvine Medical Center (n = 28), USA or California Pacific Medical Center (n = 3), USA. Patients at Irvine were implanted with SEEG electrodes with 5 mm spacing and/or ECoG grids with 1 cm spacing, and patients at CPMC were implanted with ECoG strips with 1 cm spacing. At both sites, electrophysiology and analog photodiode event channels were recorded using a 256-channel Nihon Kohden Neurofax EEG-1200 recording system and sampled at 500 (n = 2), 1000 (n = 2), or 5000 Hz (n = 19). For three datasets from UC Irvine, a separate Neuralynx ATLAS recording system was used to record the analog photodiode channels (n = 2 at 4000 Hz and n = 1 at 8000 Hz) and a subset of iEEG channels (n = 1 at 4000 Hz and n = 2 at 8000 Hz). Photodiode events were then aligned to the iEEG data acquired in parallel via the Nihon Kohden clinical amplifier based on cross-correlation of shared iEEG channels.

Pre-operative T1 MRI and post-implantation CT scans were collected as part of standard clinical care, and recording sites were reconstructed in patient space by aligning scans via rigid-body co-registration as described in Stolk et al.⁵⁴ Anatomical locations of electrodes were determined by manual inspection in native patient space under supervision of a neurologist. Electrode positions were then warped to a standard MNI 152 template brain using volume-based registration in SPM 12 as implemented in Fieldtrip.⁵⁴ Group-level electrode positions are plotted in MNI coordinates relative to the cortical surface of the fsaverage brain template from FreeSurfer,⁵³ with right hemisphere electrodes mirrored to visualize all electrodes on the left hemisphere.

Data cleaning, preprocessing, and analyses were conducted using the Fieldtrip toolbox⁵¹ and custom MATLAB code. Raw iEEG traces were manually inspected by a neurologist for epileptiform discharges along and artifacts (e.g., machine noise, signal drift, amplifier saturation, etc.). Data in regions or epochs with epileptiform or artifactual activity were excluded from further analyses. Preprocessing involved applying a 0.5-250 Hz anti-aliasing Butterworth filter and notch filtering for line noise at 60 Hz and at the next 4 harmonics (120, 180, 240, and 300 Hz) using a 2 Hz bandwidth Butterworth filter, resampling to 500 Hz, and re-referencing using adjacent bipolar montages for SEEG and common average schemes for ECOG grids/strips. Continuous data were then visually re-inspected for quality. Trials were rejected for task interruptions and behavioral outliers (RTs missing, <0.3 s, >2.0 s, or >3 standard deviations from the patient mean), including errors and partial errors, which were too infrequent to analyze. Finally, trials were segmented from -0.25 to 2.5 seconds relative to stimulus onset and rejected for excessive variance in the preprocessed time series or the differentiated preprocessed time series. Exclusion criteria for trial variance were based on patient-specific thresholds of trial-level standard deviations ranging from 5 to 10 standard deviations. Between 0 and 13 trials per patient were rejected for excessive variance (mean ± S.D.: 4.0 ± 3.2 trials). In total, this process resulted in 128-605 trials per patient (mean ± S.D.: 282.5 ± 83.8) available for analyses. Table S1 includes basic demographic information and per-subject information on channels and trials excluded within our ROIs.

We used the FieldTrip toolbox⁵¹ and custom MATLAB (Mathworks) scripts to convolve trial data (S-locked: -0.5 to 1.25 s; R-locked: -0.75 to 0.75 s) with 51 logarithmically spaced complex Morlet wavelets (number of cycles = 6) ranging from 2 to 152 Hz. We included a 1.5 s buffer period on both sides of the signal, which was discarded afterwards. We obtained a continuous estimate of instantaneous power by squaring the magnitude of the continuous wavelet transform and log transformed the result in preparation for linear modeling. We then normalized the log-transformed power values by randomly sampling values from the baseline period (-0.5 to -0.2s relative to stimulus) of all trials 1000 times and used the mean and standard deviation of this surrogate distribution to z-score within each electrode and frequency separately. For the case of investigating the preparatory period, we normalized the power values to a bootstrapped distribution sampled from the entire task recording. HFA was treated as a separate aggregate band (70 - 150 Hz). It was created by bandpass filtering the time series in 8 10-Hz bins (70 - 80, 80 - 90, etc.) and extracting the amplitude of the Hilbert transform within each bin. The HFA was normalized in the same manner as described above, separately within each bin, then averaged together to produce one composite band. Theta (4 – 8 Hz) and beta (12 – 30 Hz) power was extracted by averaging single-trial data within the respective bands and downsampling to 40 Hz prior to statistical analysis to improve computational efficiency. HFA was downsampled to 100 Hz.

Quantification and statistical analysis

For all of our analyses, we used linear mixed-effects models (LMM). Neutral and Congruent trials were combined into “NoConflict” due to the lack of behavioral and neural differences between these conditions. RTs were log transformed prior to fitting. We began with the omnibus model predicting RTs (Equation 1) and excluded terms that were not significant at α = 0.05 to reduce the model complexity.

$$RT\sim CurrentConflict+PreviousConflict+\%Cong+CurrentConflict:PreviousConflict+CurrentConflict:\%Cong+PreviousConflict:\%Cong+CurrentConflict:PreviousConflict:\%Cong+\left(1|Subject\right)$$ (Equation 1)

Here, CurrentConflict has 2 levels (Conflict and NoConflict) and represents the fixed effect of the current trial type, with NoConflict as the reference. PreviousConflict is the analogous factor for the trial type before the current trial. %Cong (percent congruent) is either 16%, 33%, or 50% and was centered at 33% in the model. Colons (“:”) indicate interactions and the “(1 | Subject)” term represents a random intercept for each subject. P-values were obtained using Wald t tests. The final model after the top-down model selection procedure is given below (Equation 2).

$$RT\sim CurrentConflict+PreviousConflict+\%Cong+CurrentConflict:PreviousConflict+CurrentConflict:\%Cong+\left(1|Subject\right)$$ (Equation 2)

For neural data, we used a well-established information theoretic approach^{54,55,56,57,58,59} to select for task encoding electrodes for subsequent analyses. Task encoding was defined as an electrode in which trial type (con, neu, inc) explained a significant amount of variance in the HFA using a one-way analysis of variance (ANOVA). Electrodes that exhibited a significant main effect of trial type for more than 100 ms were kept for further analysis. This process yielded 93 task encoding electrodes for dmPFC and 121 for dlPFC. LMMs were performed at every time point. We corrected for multiple comparisons using permutation tests on the Wald t time series by shuffling the trial labels within each subject 1000 times before fitting the LMMs to the data and applying threshold-free cluster enhancement (TFCE) with E = 2/3 and H = 2.^60,61 TFCE scores in the top 5% of the surrogate distributions were considered statistically significant. P values for permutation tests are reported as the largest p value among significant clusters for each effect.

Published: August 14, 2025