Abstract
Background Compared to healthy individuals, patients with obsessive-compulsive disorder (OCD) are found to rely more on model-free decision-making strategies which may underlie symptom expression. It is, however, unclear whether these behavioural differences are represented in neural alterations of model-free and model-based decision-making when tested simultaneously.
Methods We investigated the neural signatures of 22 OCD patients and 22 matched controls who completed a two-step Markov decision-making task during functional MRI scanning. We used hierarchical Bayesian modelling and Bayesian statistics to examine model-based and model-free decision-making behaviours. Parametric regressors were employed for model-free and model-based reward prediction errors to inform neural reward presentation, which we analysed using a Bayesian Multilevel Modeling (BML) approach. Associations between significant activations and symptoms as well as cognitive scores were explored using Bayesian linear regression.
Results While controls received significantly more rewards and were significantly less stochastic compared to patients, both groups similarly relied on model-free decision-making strategies. Importantly, our group comparison of neural reward prediction error responses showed greater activation for model-based reward prediction error in the medial prefrontal cortex (mPFC) in OCD patients compared to controls, but no differences for model-free reward prediction error processing. Increased mPFC activity was associated with lower depressive symptoms but not with other clinical or cognitive scores.
Conclusion These findings support the notion that OCD is associated with a bias towards habitual behaviour, which may be expressed through altered mPFC activity underlying goal-directed behaviour. Tackling the hyperactivity on the mPFC may provide a new target for optimised treatment and interventions.
Introduction
One of the key impairments of obsessive-compulsive disorder (OCD) is the inability to make effective and advantageous decisions (1,2). Decision-making allows successful interaction with the environment through planning, monitoring, feedback integration and feedback-based updating. This process can be categorized into two distinct systems – model-free and model-based decision-making, which operate in parallel to choose the most suitable action (3–5). Model-free decision-making is reinforcement-learning based, and therefore automatic, guided by past successful choices, and inflexible to changes (6–8). Model-free decision-making may be described as habit learning, which requires direct responses to environmental stimuli without following a goal or explicit value. Model-based decisions, on the other hand, require an understanding of action-outcome contingencies, and with this an understanding of a given task. They are, therefore, deliberate, flexible, and adaptive, aiming for optimal outcomes (6,9,10), which relates to goal-directed behaviour. They may be described as a form of instrumental conditioning, where outcomes are induced by explicit actions. Distinct neural systems represent these goal-directed and model-based actions (6,11). A recent meta-analysis (7) revealed overlapping but distinct regions, showing that while the goal-directed system has been linked to activity in the ventral striatum, the medial prefrontal cortex (mPFC) and the orbitofrontal cortex (OFC), the habit system has been associated with activity in ventral striatum as well, but also in the globus pallidus and caudate head.
Individuals with OCD are usually characterized by a relatively high inflexibility and tendency towards sticking to habits and routines. Thus, it does not come as a surprise that they often exhibit an imbalance in these two systems, favouring rigid and repetitive decisions (1,10,12), which lead to inflexibility and resistance to change (10,13–15), potentially due to an overreliance on model-free decision-making. Additionally, compulsions in OCD are also shown to arise from this core difficulty of regulating actions and thoughts in a goal-directed manner (16–19). Neuroimaging studies looking at model-based and model-free decision-making separately, suggest that this shift towards habitual decisions in OCD is associated with alterations in the cortico-striato-thalamo-cortical (CSTC) circuits. Activity within these loops, involving – amongst others – OFC,, anterior cingulate cortex (ACC) and caudate nucleus, which are responsible for governing habits and automated behaviours (7,20–22), have been proposed to become hyperactive or unresponsive to inhibitory signals in OCD, potentially establishing the neuronal basis for the incapacity to inhibit obsessive or compulsive thoughts or behaviours (23,24). Additionally, reduced grey matter volume and decreased activity in the medial prefrontal cortex (mPFC), a region responsible for constructing, updating and maintaining an internal model of the environment to guide goal-directed behaviour (25–27), has been linked to over-reliance on habitual control in OCD patients (25,28). Furthermore, disruption of activity in various regions of the basal ganglia, such as, subthalamic nucleus, nucleus accumbens, has been found to be associated with impaired behavioural inhibition, resulting in heightened impulsive choices (29,30).
Few studies investigate model-based and model-free decision-making simultaneously in OCD (8,31). The two-step Markov decision-making task (9) distinguishes between model-free and model-based learning behaviourally, and when combined with fMRI, revealed that the striatal reward prediction error (RPE) signal reflected both model-free and model-based decisions in healthy individuals (9). In a behavioural set-up of this two-step task, Voon and colleagues (32) found that individuals with different compulsive behaviours (binge eating disorders, methamphetamine addiction, and OCD patients), resorted to more model-free decisions. A follow-up behavioural study comparing the influence of reward and loss outcomes on decisions in this task demonstrated that, following rewarded trials, OCD patients leaned more towards model-free decisions, while the reverse was observed in loss trials (8). Notably, patients displayed a heightened tendency to maintain the same first-level choice in consecutive trials, irrespective of win or loss, compared to healthy controls. Furthermore, the severity of compulsion scores correlated with habitual learning in rewarded trials whereas obsession scores correlated with increased choice switching following loss trials. In a very recent study, Kim and colleagues (31), for the first time, explored neurocomputational alteration during model-based and model-free decision-making in OCD patients using the two-step task and applying a computational model which was designed to account for an uncertainty-based arbitration process of model-based and model-free strategies. They found, first, that patients showed hypoactivity in the inferior frontal gyrus when tracking the reliability of goal-directed behaviour; and second, that, patients exhibited weaker right ipsilateral ventrolateral prefronto-putamen coupling to reach specific goals, which was correlated with more severe compulsivity. More evidence, however, is needed to understand the neural mechanisms of altered model-based and model-free decision-making and their relevance for clinical symptoms in OCD.
Therefore, the aim of this study was to explore the neural mechanisms of model-based and model-free decision-making in OCD patients using the two-step task with fMRI, incorporating hierarchical Bayesian modelling and Bayesian Multilevel Modeling (BML) which performs Bayesian analyses based on pre-specified regions of interest. All other statistical analyses were also performed using Bayesian statistical methods, following a timely debate of misinterpretations of p-values and the replication crisis (33). Bayesian statistical approaches incorporate prior information into the analysis allowing for regularisation of models and preventing overfitting, which increases the precision of the findings, making them more generalisable and reliable (34,35). Based on the literature, our regions of interest included the caudate, putamen, nucleus accumbens, mPFC and ACC for both model-free and model-based RPE processing. Timeseries of the model-free and model-based decision-making were used to explore which strategy was reflected in each region and whether this differed across groups. We hypothesized that the OCD patients would exhibit predominantly model-free behaviour with caudate, putamen, nucleus accumbens and ACC activations reflecting more model-free RPE, while the healthy controls would display a mixture of model-free and model-based behaviour, with the mPFC and the nucleus accumbens reflecting more model-based RPE. Furthermore, we expected that neural alterations would link to symptom strength. We also expected that neural alterations in model-based prediction-errors would be associated with stronger compulsivity.
Methods
Participants
Twenty-five adults with OCD and 23 healthy controls were recruited for this study out of which one healthy control and three OCD patients had to be excluded from the final analysis due to corrupted imaging data. The final data for the analysis, therefore, consisted of 22 OCD patients and 22 healthy controls. Demographic information, clinical and cognitive scores are presented in Table 1. Patients met Diagnostic and Statistical Manual of Mental Disorders (DSM)-IV criteria for OCD (36). Exclusion criteria for both groups were a history of severe head injuries, seizures, neurological diseases, schizophrenia, autism, substance and alcohol abuse/dependency, mental retardation, severe medical conditions, and pregnancy. 16 patients were under medication while the other six were medication-naïve or had stopped medication at least one week before scanning. The study was approved by the Technical University of Munich ethics committee, and all participants gave informed consent after receiving a complete description of the study.
Task description
All the participants performed the two-step task by Daw and colleagues (9) inside an fMRI scanner, which is a reinforcement learning task to measure the relative degree of model-free against model-based decision-making. A detailed description of the task has been described elsewhere (37,38), in the supplementary material, and in Figure 1.
On each trial, a first stage choice between two images (here, A or B) leads probabilistically to a second stage choice (here, images in set ‘a’ or images in set ‘b’). Importantly, each first-stage stimulus is more strongly (70% vs 30%) associated with a particular second-stage state throughout the experiment, imposing a task structure that can be exploited in a model-based choice. All stimuli in stage 2 are associated with a probabilistic reward changing slowly and independently based on Gaussian random walks, forcing subjects to continuously learn and explore the second stage choices.
Analysis
Behavioural Analysis
Performance analysis
To measure performance in the task, reward percentages for both groups were calculated as the ratio of number of rewarded trials to the total number of trials per subject and compared using robust Bayesian ANOVA, followed by a Bayes factors (BF10) analysis, quantifying the more likely hypothesis based on the observed data. Evidence for alternative hypothesis over null hypothesis was identified if BF10 > 1 (BF10 [1-3]: anecdotal evidence; BF10 [3-10]: moderate evidence; BF10 [10-30]: strong evidence; BF10 [30-100]: very strong evidence; BF10 > 100: extreme evidence).
Analysis of stay probability
To identify differences between patients and controls towards a decision bias, we assessed the probability for staying with the same stage 1 choice in the subsequent trial which is deterministic of model-free and model-based choices. See supplementary material for details. Stay probabilities for ideal model-free and model-based performances were also simulated along and compared visually with actual stay probability in the task.
We used Bayesian regression analysis, assuming normal distribution, to investigate whether the stay probability varied based on participant groups (healthy controls, OCD patients), previous trial transition type (common or uncommon), previous trial outcome (rewarded or unrewarded), as well as their interactions. As we had multiple observations per person, we incorporated a group level intercept to account for the resulting dependency in the data. A substantial effect in the interaction between transition type and outcome would reflect model-based decisions, and lack of a substantial effect of interaction would indicate model-free decisions that is purely guided by reward without considering the transition structure. The equation for the model that we fit is as follows:
stay probability ∼ previous trial reward*previous trial transition type*participants + (1|Subject)
Computational modelling of task behaviour
Hierarchical Bayesian modelling was employed on the behavioural data (39) to understand the underlying mechanisms of decision-making in OCD patients compared to healthy controls and inform the imaging analysis. We fitted four models, each with 4 chains, 1000 burn-in samples and 3000 samples, after which model selection for further analysis was determined based on model-convergence and lowest leave one out information criterion (LOOIC) values. The 6-parameter model was found to be the winning model and was used for further analysis. See supplementary material for description of models, fitting procedures, and model comparisons. The 6-parameter model estimates the following parameters:
Learning rate (‘alpha 1 and alpha 2’): showing the efficiency of learning over the trials in stage 1 and stage 2, takes values from 0 to 1. The higher the score, the better the subjects understand and perform in the task.
Inverse-temperature/choice stochasticity (‘beta1 and beta2’): referring to the proportions of stochastic/random choices made during the task in stage 1 and stage 2, with beta=0 for completely random responding and beta=∞ for stochastically choosing the highest value option. Perseverance (‘pi’): Perseverance determines how strongly the subject(s) stick to their decisions. Higher the score on the perseverance scale, lower the chance of switching to a different image.
Model-weights (‘w’): referring to the degree of model-based influence on choices, takes values from 0 to 1, ‘0’ indicating more model-free decisions and ‘1’ being more model-based.
Trial-by-trial model-regressors were extracted from the winning computational model for both model-free and model-based prediction errors to be used as parametric modulators in the imaging analysis ((9), supplementary material).
To investigate group differences for the model parameters, we employed Bayesian robust one-way ANOVAs with each parameter as the dependent variable and group as the predictor variable, followed by a Bayes factors (BF10) analysis.
fMRI Analysis
Details of image acquisition and pre-processing are described in the supplementary material. After preprocessing, subject-specific design matrices were defined using general linear modelling as implemented in SPM12 (40). The fMRI analysis was centred around analysing the time-series of model-free and model-based RPEs. Four conditions were entered to calculate the average BOLD responses across all trials at each of the four time points in the task (i.e., onset first stage images, button-press first stage, onset second stage images, and button-press with reward presentation on the second stage). The computationally derived reward prediction error signals (mf_RPE and mfb_RPE) were entered as parametric regressors in the fourth condition, i.e. the button-press second stage with reward presentation. For the second-level models, the contrasts of interest defined were ‘model-free reward prediction error’ and ‘model-based reward prediction error’. The per-subject estimate of the model-based effect (‘w’ obtained from the computational model) was also included as a second-level covariate for the model-based regressor to test the correspondence between behavioural and neural estimates of the model-based effect. These contrasts were taken to a second-level one-sample t-test across all controls and patients.
The second level analysis was performed on specific regions of interest that have been implicated in model-free and model-based RL in past studies – caudate, putamen, nucleus accumbens, medial prefrontal cortex (mPFC) and anterior cingulate cortex (ACC) (Figure 3A). Anatomical masks of the caudate, putamen and nucleus accumbens were derived from the Harvard-Oxford Subcortical Atlas and created and combined using FSL (41); anatomic masks for the ACC was created combining Brodmann areas 24, 32, 33 as implemented in the WFU PickAtlas Toolbox in SPM, and mPFC was created by constructing a sphere of 5mm radius around the mPFC coordinates (x=-4, y=60, z=14) identified by Daw and colleagues (9) for model-free and model-based decisions. We used a sphere to define the mPFC to precisely target the specific coordinates since the mPFC is a functionally heterogeneous area and atlas-based regions can be quite large, potentially encompassing areas not specifically relevant to our study’s hypotheses. Parameter estimates for these ROIs were extracted from the second level contrast images using MarsBaR in SPM and entered into the Region-Based Analysis Program through BML (42), in AFNI (43).
As the BML is a recent neuroimaging technique, a concise overview is presented here (for a deeper understanding, refer to Chen and colleagues (42)). BML involves estimating the likelihood of a hypothesis based on observed data, represented as a probability density function termed the posterior distribution (abbreviated as P+). This distribution is derived by combining empirical data with a model and prior expectations. BML offers the advantage of consolidating data across subjects and Regions of Interest (ROIs), facilitating the integration of shared information among these regions. The more the posterior distribution deviates from 0, the stronger the evidence for an effect. The evidence for effects with a probability (P+) of <.10 or >.90 as categorized as weak, <.05 or >.95 as moderate, and <.025 or >.975 as strong. A probability (P+) of >.10 or <.90 is considered no evidence. Effects with a P+ of <.10 or >.90 are discussed in the main text, while comprehensive effects are presented in the probability distribution plots.
Associations between neural responses and clinical and cognitive scores
To examine the relationships between clinical and cognitive symptoms, and brain activations in patients that were significantly different from those in controls, we conducted Bayesian linear regressions for each of the symptom (i.e., Y-BOCS obsession and compulsion, OCI-R scores, HAMD) and cognitive scores (i.e., TMT-A, TMT-B, TMT-BA) as the outcome variable and significant brain activations as predictors covarying for age of illness onset. The response variables were assumed to follow a Gaussian distribution for the residuals, and priors for the regression coefficients were specified using a normal distribution with a mean of 0 and standard deviation of 10. The posterior distributions were sampled using the No-U-Turn Sampler with 4 chains, each containing 2000 iterations, including a 1000-iteration warmup phase. In Bayesian regression analysis, an effect was identified by inspecting the credibility intervals. If 0 is not contained in the credibility interval, a significant effect was detected.
General statistical implementations
All statistical analyses were performed using R Statistical Software (version 4.3.1) (44); hierarchical Bayesian modelling was performed using hBayesDM (39); Bayesian linear regression and ANOVA analysis was carried out using the brms package (version 2.20.4) (45); imaging analysis was performed using Statistical Parametric Mapping software, SPM12 (40); masks for region of interest analysis was made using FMRIB Software Library (FSL) (version 6.0.5.2) (41); and BML analysis for the regions of interest was performed using the Region Based Analysis (RBA) Program (42) in the Analysis of Functional NeuroImages (AFNI) Software (version 23.2.12) (43).
Results
Performance differences across groups
The robust Bayesian ANOVA revealed a substantial group difference in reward percentage for healthy controls (Mean: 60.3, SD: 3.43) and OCD patients (Mean: 54.6, SD: 3.5) (B=5.61, EE=1, 95% CI [3.64, 7.57]) with Bayes Factor analysis revealing extreme evidence (BF10=4779.379), indicating that controls performed better than OCD patients.
Stay probability differences across groups
Results of the Bayesian regression analysis revealed no main effect for group, but a substantial main effect of reward (rewarded vs. unrewarded: B=0.14, EE=0.03, 95% CI [0.07, 0.21]), indicating that all participants stayed on the same first level choice in the next trial if the previous trial was rewarded, suggesting a model-free behaviour. We found no evidence for interaction effects indicating that the subjects followed a model-free decision-making approach. Stay probabilities for ideal model-free and model-based behaviour are displayed in Figure 2a and actual stay probabilities in the current task are displayed in Figure 2b.
a. Ideal Model-free and Ideal Model-based decision-making behaviour: Model-free RL predicts that a first-stage choice yielding a reward is likely to be repeated on the upcoming trial, regardless of a common or an uncommon transition, Model-based RL predicts that an uncommon transition should affect the value of the next first stage option, leading to a predicted interaction between reward and transition probability. b. Actual stay proportions, averaged across subjects, displaying a characteristic of using both model-free and model-based decision-making strategies.
Modelling parameters differences across groups
The robust Bayesian ANOVA revealed group differences in stochasticity at stage 1 ‘beta 1’ (healthy controls vs. OCD patients: B=1.57, EE=0.7, 95% CI [0.21, 2.94]). The Bayes Factor analysis for stochasticity at stage 1 ‘beta 1’ scores indicated anecdotal evidence (BF10=2.65), suggesting that patients were more random in their choices compared to controls. None of the other parameters revealed substantial effects.
Differences in the neural signature of model-free and model-based prediction errors
Group comparisons between healthy controls and OCD patients in the model-based RPE signal revealed that the OCD patients had a higher activation in the mPFC compared to healthy controls (strong evidence; P+=0.02). We did not find any group differences in the model-free RPE signal in patients and controls. Parameter estimates for group comparisons in model-based and model-free RPE signal are displayed in Figure 3B.
Within group effects for model-based RPE revealed activations in the mPFC (moderate evidence; P+=0.96) in healthy controls; and activations in the mPFC (strong evidence; P+=1), right (strong evidence; P+=0.99) and left (strong evidence; P+=0.98) caudate, ACC (strong evidence; P+=0.98) and right putamen (weak evidence; P+=0.93) in the OCD patients. Within group effects of model-free RPE revealed activations in the mPFC (strong evidence; Controls: P+=1, OCD: P+=1), right accumbens (strong evidence; Controls: P+=1, OCD: P+=1), left accumbens (strong evidence; Controls: P+=1, OCD: P+=1), left putamen (strong evidence; Controls: P+=1, OCD: P+=1), right putamen (strong evidence; Controls: P+=1, OCD: P+=1), ACC (moderate evidence; Controls: P+=0.97, OCD: P+=0.97) and right caudate (weak evidence in Controls; P+=0.94, moderate evidence in OCD; P+=0.97) for both healthy controls and OCD patients. Parameter estimates for group comparisons in model-based and model-free RPE signal for healthy controls and OCD patients are displayed in Figures 3C and 3D respectively.
ROIs for model-free and model-based RPE; Masks were derived from the Harvard-Oxford Subcortical Atlas using FSL. B. ROI analysis showing posterior distributions for OCD patients compared with healthy controls (left: Model-free RPE; right: Model-based RPE). Values >.90 represent higher activation in healthy controls, and <.10 represent lower activation in the healthy controls. C. ROI analysis showing posterior distributions within each group for model-free RPE (left: Healthy controls; right: OCD patients). D. ROI analysis showing posterior distributions within each group for model-based RPE (left: Healthy controls; right: OCD patients). Note: Posterior distributions: strong evidence: (P+ > 0.975 or <0.025), moderate evidence (P+ > 0.95 or <0.05), weak evidence (P+ >0.90 or <0.10) of an effect. Boxes outlined in red highlight regions with substantial effects in each of these plots.
Bayesian Linear Regressions results
First, we examined the relationships between model-based RPE signal in the mPFC, OCD symptoms and cognitive scores in OCD patients (Table 2). The Bayesian regression revealed that mPFC activity significantly and negatively predicted HAMD (B=-0.15, EE=0.08, 95% CI [-0.31, -0.00]), controlling for age of onset. No other regression analyses revealed significant results.
Discussion
We investigated neural correlates of model-free and model-based prediction errors in OCD patients versus healthy controls using an fMRI-version of the two-step Markov decision task and computational modelling. Behaviourally, the model parameters revealed that the patients were more random and received fewer rewards than healthy individuals, and relied more strongly on model-free strategies compared to model-based, which confirms previous findings (8,38,46). Importantly, we found higher activation in the mPFC in OCD compared to controls for model-based RPE processing, but no differences for model-free RPE processing. These findings support the notion that OCD is associated with a bias towards habitual behaviour, which may be expressed through altered mPFC activity underlying goal-directed behaviour (11,32,47).
Activations in the mPFC for model-based RPE align with literature highlighting the region’s role in updating and maintaining an internal model of the environment to guide goal-directed behaviour (25–27). Supporting our findings, previous studies have observed relative hyperactivation of the mPFC, including the supplementary motor area in OCD patients during task performances involving inhibitory control and conflict processing compared to healthy controls (10,48). In a study combining volume-localised proton magnetic resonance spectroscopy and fMRI in an inhibitory control task, Yücel and colleagues (48) argued that while having primary relevance to the pathogenesis of the disorder, increased activation in the mPFC reflects a compensatory mechanism to achieve adequate task performance, potentially due to biochemical abnormalities in the dorsal ACC. Together these results suggest that hyperactivity in the mPFC might be a potential neural correlate of altered goal-directed behaviour in OCD. Additionally, it has been proposed that the mPFC plays a central role in resolving conflicts between the brain’s habitual and goal-directed system (49,50), either by deliberately overriding the habitual system when required (49), or by mediating the two systems to naturally interact (50). This supports our findings as both groups, when analysed separately, show model-based and model-free prediction error processing in the mPFC.
OCD patients showing higher activations for model-based RPE in the mPFC compared to controls, and also showing overlapping region activations during both model-free and model-based decisions, further suggests that the patients may over-recruit regions responsible for habitual decisions when they are unable to resolve conflict between habitual and goal-directed strategies. This may support previous findings of a disrupted arbitration mechanism in OCD patients (28,51) and might potentially be responsible for the dominance of habitual behaviours in OCD. In line with our findings, past studies have also found structural and functional alterations within the fronto-striatal circuits, particularly the mPFC, putamen and caudate responsible for the imbalance between model-free and model-based decision-making in OCD (28,52). Additionally, the mPFC is a crucial part of the dorsal cognitive CSTC circuit which contributes to habit formation and compulsivity in OCD (53). Since the nucleus accumbens and the left putamen were activated only during model-free RPE, our results suggest activity in these regions as potential neural correlates for model-free decisions in OCD.
Furthermore, our results revealed that the same brain regions reflected model-free RPE in the controls and patients. These regions included activations in the bilateral nucleus accumbens and putamen, the ACC, the mPFC, and the right caudate nucleus, which are in accordance with our hypothesis and the literature (7,20,23,24,46,54–57), confirming the contributions of these regions to model-free RPE processing. Left caudate was not associated with model-free RPE in our study, as also reported by Guida and colleagues (58), indicating fewer and less strong activations in the left caudate for habitual activities in everyday-life.
Moreover, in OCD patients mPFC RPE signalling negatively predicted depressive symptoms measured using HAMD. However, we did not find any other symptom associations. Given the hyperactivity of the mPFC in patients it is surprising that depressive symptoms are declining with increasing activity. Although mPFC signalling alterations during reward and reward prediction error processing have been reported in patients with major depressive disorder (59–61), future studies should further explore these associations.
Limitations
This study has several limitations. First, our moderate sample size did not allow to explore subgroups of patients with specific symptoms. Subsequent studies should aim to replicate the results in a larger sample allowing for a differentiation in habitual and goal-directed decision-making between several symptom subtypes (e.g., washers versus checkers). Second, we were unable to elicit model-based decision-making in healthy controls, potentially due to a combination of task instruction, reward presentation and task duration (62,63). Future studies should replicate our results with more trials, potentially in a longitudinal framework to provide a more thorough understanding. Third, our analysis included both medicated and unmedicated patients. Future studies should aim for larger sample sizes to verify results in separate analyses for medicated and unmedicated individuals.
Conclusion
Taken together, this study revealed an overreliance on model-free behaviour and a hyperactivity in the mPFC in the OCD patients compared to controls in response to model-based prediction error processing, suggesting a disruption of the arbitration system responsible for resolving conflict between habitual and goal-directed decisions. Thus, the mPFC may provide a new target for optimised treatment and intervention, such as non-invasive stimulations using transcranial direct current stimulation.
Data Availability
All data produced in the present study are available upon reasonable request to the corresponding author.
Competing interests
The authors report no competing interests.
Acknowledgements
We would like to thank the patients and their families as well as the healthy controls for their time and effort. Furthermore, we would like to thank Dr. Felix Brandl and Dr. Christian Sorg for their insights and support during data acquisition.
References
