We administered a reactive behavioral switching paradigm to two groups of post-surgery neuropsychiatric patients while intracranial electroencephalographic activity (iEEG) was recorded, and to one group of healthy controls. The first group consisted of three patients with drug-resistant partial epilepsy implanted with linear depth macroelectrodes in the dorso-medial prefrontal cortex (dmPFC). The second group consisted of four patients with severe and drug-resistant OCD implanted with deep brain stimulation macroelectrodes in the subthalamic nucleus (STN). The third group consisted of ten healthy controls (from whom we collected only behavioral recordings). We report demographic and clinical data in Table S1, 2.Each trial started with two colored squares (yellow/pink) randomly displayed on each side of a central white square (Fig. 1). After 500 ms, the white central square turned either pink or yellow (ongoing rule) to prompt the patient to press a button indicating on which side of the screen the cue matching the central square color was displayed (left/right bimanual response). This stimulus-response association rule pseudo-randomly changed every 2 to 6 trials. Overall, the task consisted of two to four sessions composed of 50 switch trials each (138 ± 49 switch trials per subject).Fig. 1: Behavioral task and results.a Task-switching paradigm. Patients had to indicate on which side the stimulus (STIM) matching the central square color (RULE) was. This response was followed by a feedback cue (FB) indicating whether it was correct and/or whether it fell within the allowed response time window. b Behavioral performances of patients with OCD (upper row; n = 4), epilepsy (EPI: middle row; n = 3) and of healthy controls (HC: bottom row; n = 10). Evolution of hit rates and reaction times averaged according to trial relative position to switch trials (left and central panels) or as a function of trial type (red: switch; black: non-switch) and accuracy (hit or error). Error-bars indicate SEM between patients. Two-tailed paired t-tests showed that patients (OCD and EPI; n = 7) and healthy controls made more errors on switch trials (patients: p = 1.09 × 10−4; HC: p = 5.73 × 10−5), were longer during correct switch trials (p = 9.66 × 10−5; HC: p = 7.61 × 10−5) and their reaction times were also significantly faster during incorrect switch trials relative to incorrect non-switch trials (p = 2.02 × 10−2; HC: 2.29 × 10−2). *p < 0.05. See also Table S3 for statistics across trials of each of the sevent patients. S: switch (sw) trial; S-2 and S + 2: non-switch (nsw) trial occurring two trials before (S-2) or after (S + 2) a switch.Patients were instructed to select the correct response as quickly and as accurately as they could. Hence, the task required to form a preliminary response by preemptively applying the previous rule as soon as the current stimulus pair was displayed. In switch trials (20% of trials), patients had to override this preliminary response to instantiate and apply the new rule, resulting in the selection of the alternative action. To promote speeded decisions, in line with the original studies in monkeys3,4, we implemented a response time window for non-switch trials (excluding switch trials, which typically have longer durations). An incorrect feedback was given during non-switch trials whenever the response time exceeded a predefined limit. The limit was adjusted based on participants’ performance after each switch trial: it increased by a 40 ms increment after incorrect switch trials and decreased by the same increment after correct switch trials, ensuring adaptive stringency. Note that this procedure did not censor observed RT distributions and maintained a consistent level of time pressure throughout the task as participants quickly reached a (fixed) floor value (300 ms).Behavioral performance: the cost of switchingGiven the striking similarity between OCD and epileptic patients’ behavioral patterns (see Fig. 1), in the following, we report behavioral statistics pooled across the two groups of patients. It is noteworthy that these behavioral patterns closely mirrored those observed in healthy controls (Fig. 1b), while we also found consistent findings from ANOVAs followed by post-hoc tests performed separately for each group and even for each patient within the group (dmPFC group: Fig. S1; STN group: Fig. S2; see also Table S3 for statistics). First, we performed an analysis of variance (ANOVA) on mean reaction times (RTs), which demonstrated a significant interaction between trial type (switch vs. non-switch) and accuracy (hit vs. error) (F(1,24) = 15.12; p = 6.98 × 10−4; Fig. 1b; repeated measures ANOVA). Post-hoc comparisons confirmed that RTs on correct trials were significantly longer on switch trials in comparison to non-switch trials (99 ± 11 ms; t(6) = 9.14; p = 9.66 × 10−5; paired two-tailed Student’s t-test). RTs were also significantly faster during incorrect switch trials relative to incorrect non-switch trials, supporting the idea that errors during switch trials resulted from a failure to override the preliminary response to apply the new rule (−86 ms ± 28 ms; t(6) = −3.07; p = 2.02 × 10−2; Fig. 1b; paired two-tailed Student’s t-test). Unsurprisingly, patients also made significantly more errors on switch trials in comparison to non-switch trials (switch cost(error)= 26 ± 3%; t(6) = 8.09; p = 1.09 × 10−4; paired two-tailed Student’s t-test).Behavioral switching: Neural activity in the dmPFC-STN networkTo investigate whether neural activity in the dmPFC-STN network reflected behavioral switching, we first contrasted the power estimated in the time-frequency domain between correct switch (Sw/Hit) and correct non-switch (NoSw/Hit) trials across all recording sites within each brain region. dmPFC and STN LFP recordings were analyzed separately since these data came from distinct groups of patients (dmPFC: n = 33; STN: n = 24 recording sites). There was a significant increase in theta activity at rule onset when switching, both in the dmPFC (Sw/Hit vs NoSw/Hit, [−156 ms; +664 ms], sum(t(32)) = 232.7, pc = 1.33 × 10−4, Fig. 2a, b) and in the STN ([4 ms; +1088 ms], sum(t(23)) = 653.5, pc = 1.00 × 10−4, Fig. 3a, b), as well as an increase in high-gamma activity (dmPFC: [+98 ms; +625 ms], sum(t(32)) = 488.4, pc = 1.67 × 10−5, Fig. 2a; STN: [-523 ms; +189 ms], sum(t(23)) = 327.8, pc = 5.58 × 10−4, Fig. 3a). These switch-related changes in theta and high-gamma activity replicated when time-locking iEEG activity at response, which further revealed an increase in dmPFC’s beta (15–30 Hz) activity (Sw/Hit vs NoSw/Hit; [−631 ms; 229 ms], sum(t(23)) = 743.6; pc = 1.67 × 10−5, Fig. 2a and Fig. S3).Fig. 2: dmPFC neural activity during behavioral switching.a Time-frequency analysis of neural activity time locked on rule (left panels) or on response (right panels) onset across all dmPFC contact-pairs (n = 33). Warm (cold) colors indicate significant increases (decreases) of power (pc< 0.05, FWE cluster-corrected). b Anatomical location of recording sites in the dmPFC plotted on a 3D reference brain. Each colored dot represents a contact-pair displaying a significant modulation in the theta (green) and/or high-gamma (red) during task-switching. c Time course of theta power (5–10 Hz) averaged across correct (red) switch, incorrect switch (orange) or correct non-switch (black) trials (n = 14 dmPFC contact-pairs). Bold traces indicate average activity and shaded areas correspond to SEM across contacts. The black horizontal bar at the bottom indicates time points for which the statistical contrast between incorrect and correct switch trials was significant (pc< 0.05). Orange (red) vertical arrows at the bottom indicate average peak latencies for error (hit) switch trials (averaged across contact-pairs; horizontal lines indicate SEM). Vertical grey shaded rectangles correspond to 95% confidence intervals of RT (left panels) or rule onset (right panels). d Neuro-psychometric curves depicting the relationship between hit rate and averaged theta power (across n = 14 dmPFC contact-pairs) binned into deciles separately during switch and non-switch trials. Error-bars correspond to SEM across contact-pairs. e Time course of high-gamma power (60–200 Hz; n = 8 dmPFC contact-pairs). Bold traces indicate average activity and shaded areas correspond to SEM across contacts. f Neuro-psychometric curves depicting the relationship between hit rate and averaged high-gamma power (across n = 8 dmPFC contact-pairs) binned into deciles separately during switch and non-switch trials. Error-bars correspond to SEM across contact-pairs. dmPFC: dorsomedial prefrontal cortex; Sw: switch; NoSw: non-switch.Fig. 3: STN neural activity during behavioral switching.a Time-frequency analysis of neural activity time locked on rule (left panels) or on response (right panels) onset (contrasts across all STN contact-pairs; n = 24; pc < 0.05 FWE cluster-corrected). b Anatomical location of STN recording locations with higher behavioral switch related theta increase (blue dots) displaying switch-related theta activity. Each contact-pair is plotted on a 3D reconstruction of STN sensorimotor (in green), associative (in purple) and limbic (in brown) territories. c Time course of theta power (5–10 Hz) averaged across correct (red) switch, incorrect switch (orange) or correct non-switch (black) trials (n = 8 STN contact-pairs). Bold traces indicate average activity and shaded areas correspond to SEM across contacts. The black horizontal rectangle at the bottom indicates time points for which the statistical contrast between incorrect and correct switch trials was significant (pc< 0.05). Orange (red) vertical arrows at the bottom indicate average peak latencies for error (hit) switch trials (average across contact-pairs; horizontal lines indicate SEM). Vertical grey shaded rectangles correspond to 95% confidence intervals of RT (left) or rule onset (right). d Neuro-psychometric curves depicting the relationship between hit rate and averaged theta power (across n = 8 STN contact-pairs) binned into deciles separately during switch and non-switch trials. Error-bars correspond to SEM across contact-pairs. e Time course of high-gamma power (60-200 Hz; n = 8 STN contact-pairs). Bold traces indicate average activity and shaded areas correspond to SEM across contacts. f Neuro-psychometric curves depicting the relationship between hit rate and averaged high-gamma power (across n = 8 STN contact-pairs) binned into deciles separately during switch and non-switch trials. Error-bars correspond to SEM across contact-pairs. STN subthalamic nucleus. Sw switch; NoSw: non-switch.Next, we focused our analyses on contacts exhibiting a differential activity when there was a change of stimulus-response mapping or not (i.e., contrasting switch vs. non-switch trials independently from accuracy) to investigate the precise time course of theta (5–10 Hz) or gamma (60–200 Hz) band activities in the dmPFC (Fig. 2) and in the STN (Fig. 3). In the dmPFC, we found that the timing of theta and gamma activities differed between switch hit and switch error trials. Theta activity peaked at response when patients failed to switch (0 ± 0.32 ms), but peaked significantly earlier when the switch was successful (Sw/Hit vs. Sw/Err, −279 ± 60 ms; t(13) = −4.6, p = 4.54 × 10−4 paired two-tailed Student’s t-test). Consistent with previous neuronal recordings in monkey dmPFC3, we observed a similar pattern in the high-gamma band where activity peaked before the response for correct switches (−129 ± 26 ms) and after the response when it was missed (+90 ± 40 ms; Sw/Hit vs. Sw/Err, −219 ± 42 ms; t(8) = −5.23; p = 7.93 × 10−4; paired two-tailed Student’s t-test, see Fig. 2e). This high-gamma activity peak in the dmPFC reliably followed the peak of theta activity (Sw/Hittheta vs. Sw/Hithigh_gamma, −149 ± 44 ms; t(21) = −2.5; p = 2.02 × 10−2 unpaired two-tailed Student’s t-test). Furthermore, both theta and gamma activities were also significantly higher for incorrect than for correct switch trials (Theta band: Fig. 2c: Sw/Hit vs. Sw/Err, [−299 ms, +522 ms] relative to response, sum(t(13)) = 85.1, pc = 1.67 × 10−5; Gamma band: Fig. 2E, Sw/Hit vs. Sw/Err, [−6 ms, +366 ms] relative to response, sum(t(8)) = 57.7, pc = 1.67 × 10−5). Taken together, these results demonstrate that successful behavioral switching depends on a structured pattern of increased theta/high-gamma activity in the dmPFC from rule onset to response and precisely controlled in both time and amplitude. Consistent with this hypothesis, we found a significant amplitude-amplitude coupling between high gamma and theta activities in the dmPFC during correct switch trial (Fig. S4; [0–500 ms] time window post-rule onset; t(13) = 4.96; p = 2.61 × 10−4; one sample two-tailed Student’s t-test). This coupling was no longer significant during switch error trials (t(13) = 2.16; p = 5.04 × 10−2 > 0.05). Comparing directly the coupling strength for switch hit vs. switch error confirmed a higher theta-gamma coupling for correct switches (t(26) = 2.27; p = 3.14 × 10−2; paired two-tailed Student’s t-test).In the STN of OCD patients, similar to what we observed in the dmPFC of epileptic patients, theta activity was significantly higher for incorrect than for correct switch trials (Sw/Hit vs. Sw/Err, see Fig. 3c, [−0.04 s, 0.39 s] relative to response; sum(t(7)) = 24.1, pc = 1.91 × 10−3). Moreover, theta activity in the STN peaked at response (Sw/Hit: −0.076 ± 0.042 s) when OCD patients successfully switched, but peaked after the response when they failed (Sw/Err: +0.172 ± 0.051 ms; Sw/Hit vs. Sw/Err, −219 ± 39 ms; t(7)= −5.37; p = 1.00× 10−3; paired two-tailed Student’s t-test), with a timing (relative to response) reminiscent of the dmPFC’s high gamma activity observed in the group of epileptic patients (Sw/Hit vs. Sw/Err, −219 ± 39 ms; t(8)= −5.23; p = 7.93 × 10−4; paired two-tailed Student’s t-test). Finally, there was no difference in STN high gamma activity prior to/at the response between correct and incorrect switches, which reflected instead the upcoming motor response (Ipsilateral vs. Contralateral, Fig. S5). Instead, we found that theta increase for successful behavioral switches was larger in the STN ipsilateral to the newly selected response than in the contralateral STN, but not for unsuccessful switches (see Fig. S6, t(7) = 2.15; p = 0.03), suggesting a direct role for STN theta oscillations in rule-switch execution. Recognizing that dmPFC and STN recordings were obtained from distinct patient groups, thereby precluding dmPFC-STN connectivity analyses, we thought that it remained interesting to compare theta dynamics across these patient groups and structures. Our analysis revealed a consistent temporal lag in rule-switch-related theta increase in the STN compared to the dmPFC (dmPFC-STN theta onsets relative to the rule: dmPFC: −0.419 ± 0.03 s; STN: −0.234 ± 0.08 s; t(20) = −2.57, p = 1.81 × 10−2, see Fig. S7). Moreover, the theta increase peaked significantly later in the STN than in the dmPFC (peak latencies in dmPFC: +0.199 ± 0.074 s vs. +0.575 ± 0.052 s in the STN, t(20) = −3.54, p = 2.08 × 10−3). Overall, this dmPFC-STN theta/high-gamma dynamic suggests that the dmPFC might drive STN neural activity when external cues trigger a rule-switch and casts the STN as an executive structure downstream the dmPFC along the hyperdirect pathway.Consistent with this view, we found that neural activity in the dmPFC further encoded the anticipation of upcoming rule switches: dmPFC’s beta activity at stimuli onset increased in proportion with the probability of rule changes (dmPFC: [−846 ms; −534 ms], sum(t(32)) = 59.99, pc = 9.33 × 10−4, Fig. S8), 500 ms before the onset of the cue indicating that the rule had changed. Indeed, in our task, the probability of switch increased with the number of previous consecutive non-switch trials. Patients implicitly used this information to anticipate switches so that switch costs were large if a switch occurred after 2–3 non-switch trials (switch cost(2nsw) = −43 ± 13%, switch cost(3nsw) = −37 ± 7%). This effect disappeared if a switch occurred after 5-6 non-switch trials (switch cost(5nsw) = −17 ± 14%, switch cost(6nsw) = 3.4 ± 13%, see Fig. S8).It is noteworthy that our main results were statistically robust and reproduced when using mixed-effect analyses to assess differences in theta activity between correct switch and non-switch trials, both in the dmPFC (t(5455) = 4.72, p = 2.4 × 10−6; see also Fig. S9) and in the STN (t(6244) = 3.69, p = 0.00023; see also Fig. S10), as well as between hit and error switch trials (dmPFC: t(1284) = 4.26, p = 2.2 × 10−5; STN: t(1332) = 2.33, p = 0.019). Finally, we tested whether baseline had an influence on subsequent theta dynamic in dmPFC and STN by comparing the time course of theta power across conditions separately for high vs. low level of theta power prior rule-onset (Fig. S11; from -250ms to 0 ms; median split). This revealed that neither dmPFC (Fig. S11A–C) nor STN (Fig. S11D–F) baseline theta power significantly modulated the post-rule theta peak amplitude across conditions.Trial-by-trial fluctuations in theta/high-gamma bands predict successful behavioral switching in dmPFC and in the STNTo explore the functional role of dmPFC-STN theta/high-gamma dynamic, we then investigated whether spontaneous trial-by-trial fluctuations in neural activity related to actual fluctuations in behavioral performance. To do so, we tested brain-behavior correlations between normalized theta or high-gamma activity and performances during switch and non-switch trials separately for each brain region and for each trial type (dmPFC: Fig. 2d–f; STN: Fig. 3d–f).During switch trials, there was a negative correlation between theta activity trial-by-trial fluctuations and performances, both in the dmPFC and the STN (dmPFC: Fig. 2d; βhit~switch = −0.33 ± 0.06; t(13) = −5.67, p = 7.65 × 10−5; STN: Fig. 3d; βhit~switch = −0.13 ± 0.03; t(7) = −4.24, p = 3.82 × 10−3). However, during non-switch trials, there was no influence of STN theta activity trial-by-trial fluctuations on performances, while a negative correlation between theta activity and performances was still found in the dmPFC (theta: βhit~nonswitch = −0.14 ± 0.02; t(13) = −6.19, p = 3.29 × 10−5 < 0.01), suggesting a functional decoupling of STN and dmPFC theta activities during non-switch trials. Interestingly, we also found significant positive correlations between performances and pre-response high-gamma activity in the dmPFC during switch trials (Fig. 2f; βhit~switch = 0.18 ± 0.05; t(7) = 3.32, p = 1.06 × 10−2), whereas this correlation was negative during non-switch trials (Fig. 2f; βhit~nonswitch = −0.09 ± 0.01; t(7) = −5.66, p = 4.75 × 10−4 < 0.01). Thus, during non-switch trials, there was a negative correlation (i.e., like what was observed in its theta band and in pre-response beta, see Fig. S12), whereas this correlation became strongly positive during switch trials, suggesting that dmPFC pre-response high gamma activity reflects a neural process selective to behavioral-switching.Theta activity is negatively associated with the starting point during behavioral switching in the dmPFC and in the STNFinally, to further understand how these trial-by-trial fluctuations in neural activity influenced behavior, we used a hierarchical drift-diffusion model (HDDM30) to disentangle behavioral variability arising from switching between rules vs. selecting a new response. At the behavioral level, drift diffusion models (DDMs) are commonly used to model the dynamics of action selection as an accumulation of evidence over time until a certain threshold is reached and a response is triggered31 (see Fig. 4a). DDMs have four main parameters26: the time of non-decision t, the initial level of evidence z, a drift rate v (the accumulation speed of the information relevant to action selection) and thresholds a, which controls the speed-accuracy tradeoff of the selection process and accounts well for its type I errors9,28. In our task, we expected to find a shift of the starting point for switch trials, (1) as correct switches were associated with slower RTs and failed switches with faster RTs (compared to correct non-switch responses; see Fig. 1b), (2) as dmPFC encoded a priori the likelihood of upcoming switches and (3) as early behavioral studies suggested that, during task switching, reconfiguring the set of rules seemed to occur before the selection of a new response26,27. To test this hypothesis, we first fitted a hierarchical drift diffusion model to RT distributions of correct and incorrect responses from all patients31 (i.e., combining the behavioral data of OCD and epileptic patients), and tested which model parameter best accounted for observed switch costs (M1 model space, also included all possible parameter pairs, see Methods).Fig. 4: Behavioral switching modeling across epileptic and OCD participants.a Drift diffusion model. Response execution is preceded by an accumulation of evidence increasing sequentially in favor of one of the two options until a boundary is reached. The decision processes start at stimulus onset and at an initial level depending on the subject’s prior beliefs. The initial level turns out to be different per trial type. b Behavioral model comparison. Relative value of the deviance information criterion (DIC) per model when considering the decision threshold model a as a reference. c Behavioral model posteriors. Probability density of the initial level of evidence values for switch and non-switch trials. The initial level of evidence is lower during switch trials in accordance with subjects’ belief of an upcoming non-switch trial. The statistical difference between switch and non-switch posteriors was significant (significance from posterior probabilities: p < 0.001). d Spontaneous trial-by-trial activities included in the neural HDDM (hierarchical drift diffusion model) model space. Neural model comparison for dorsomedial prefrontal cortex (dmPFC; e) and subthalamic nucleus (STN; g). Relative value of the DIC per model with a model based on a normally distributed noise as a reference. Neural model posteriors for dmPFC (f) and STN (h). Probability density for the effect of theta power on the initial level of evidence z. A negative (resp. positive) regression coefficient means z decreases (resp. increases) when theta power increases. Here, the regression coefficient is negative for non-switch trials and strongly negative for switch trials. The statistical difference between switch and non-switch regression coefficient was significant (significance from posterior probabilities: p < 0.001). ***p < 0.001.We found that switch costs were best captured by a shift of the starting point (z parameter) toward the lower (erroneous) model boundary (Fig. 4b). Moreover, the sampled posterior distribution of z parameter indicated that HDDM starting point was significantly lower during switch trials (zsw = 0.23 ± 0.06) compared to non-switch trials (znsw = 0.44 ± 0.06; p < 0.001; Fig. 4c). We found identical results when testing which model best accounted for healthy subjects’ behavior (see Fig. S13). Posterior predictive checks confirmed that this HDDM model reproduced all key behavioral patterns observed after stimulus-response mapping changes (see Fig. 1b), for each individual in each group (dmPFC Epileptic group; see Fig. S14; STN OCD group; see Fig. S15).Next, we tested whether dmPFC and STN neural activity further modulated that initial level evidence (z). We reasoned that using band-specific neural fluctuations to adjust trial-by-trial HDDM’s starting point would significantly improve our model fit only if it reflected the switches between rules. Hence, we built a second neural HDDM model space consisting of HDDMs with starting point modulations by all the switch-related bands identified in our previous analysis of STN and dmPFC neural activity, as well as linear and multiplicative interactions with these frequency bands (M2 model space, see Methods, Fig. 4d).In the dmPFC, the neural HDDM including theta fluctuations as a modulator of the starting point outperformed other single-band models and was not significantly outperformed by more complex neural HDDMs including an interaction term between theta and high-gamma neural activities (see Fig. 4e). Consistent with our other findings in the STN, the neural HDDM model including theta fluctuations as a modulator of the starting point significantly outperformed all the other neural HDDM tested (Fig. 4f). Figure 4f–h show the posterior distribution of the regression coefficient between HDDM starting point and residual trial-by-trial power in the theta band. The significant shift of the posterior distribution toward negative correlation coefficients between the starting point and theta power residuals (dmPFC: Fig. 4f; STN: Fig. 4h, both p < 0.01) is consistent with our previous observation that, although theta activity in the dmPFC-STN network increases on average during switch trials, too high an increase becomes detrimental to performances. Interestingly, this negative correlation between the starting point and residual trial-by-trial theta power fluctuations during switch trials was also observed as early as 250 ms before rule onset in both the dMPFC and in the STN (see Fig. S16), at a moment when there was no difference on average between correct and erroneous switch. Thus, these results are consistent with the role of a prior modulation, which is usually associated with the starting point parameter.Some key brain regions, known to exhibit rule-switch activity are not mentioned in this study because an inherent limitation to intracranial EEG recordings is their limited spatial sampling. Note however that the intracranial recordings from the epileptic group allowed us to explore brain activity beyond the hyperdirect dmPFC-STN circuitry. For example, we found that broadband gamma (60–200 Hz) and beta (15–35 Hz) power in the anterior insular cortex was higher during correct switch trials than during non-switch trials (Fig. S17). The main differences with the dmPFC were that (1) there was no modulation of theta band activity in aIns and (2) the increase in both beta and broadband gamma amplitude during incorrect behavioral switches was only observed after the response, thus lying out of our temporal window of interest (in this study, we focused on the exploration of the cognitive processes related to switch between rules).
