Frontostriatal Dysfunction During Decision Making in Attention-Deficit/Hyperactivity Disorder and Obsessive-Compulsive Disorder

Background The aim of the current paper is to provide the first comparison of computational mechanisms and neurofunctional substrates in adolescents with attention-deficit/hyperactivity disorder (ADHD) and adolescents with obsessive-compulsive disorder (OCD) during decision making under ambiguity. Methods Sixteen boys with ADHD, 20 boys with OCD, and 20 matched control subjects (12–18 years of age) completed a functional magnetic resonance imaging version of the Iowa Gambling Task. Brain activation was compared between groups using three-way analysis of covariance. Hierarchical Bayesian analysis was used to compare computational modeling parameters between groups. Results Patient groups shared reduced choice consistency and relied less on reinforcement learning during decision making relative to control subjects, while adolescents with ADHD alone demonstrated increased reward sensitivity. During advantageous choices, both disorders shared underactivation in ventral striatum, while OCD patients showed disorder-specific underactivation in the ventromedial orbitofrontal cortex. During outcome evaluation, shared underactivation to losses in patients relative to control subjects was found in the medial prefrontal cortex and shared underactivation to wins was found in the left putamen/caudate. ADHD boys showed disorder-specific dysfunction in the right putamen/caudate, which was activated more to losses in patients with ADHD but more to wins in control subjects. Conclusions The findings suggest shared deficits in using learned reward expectancies to guide decision making, as well as shared dysfunction in medio-fronto-striato-limbic brain regions. However, findings of unique dysfunction in the ventromedial orbitofrontal cortex in OCD and in the right putamen in ADHD indicate additional, disorder-specific abnormalities and extend similar findings from inhibitory control tasks in the disorders to the domain of decision making under ambiguity.


Supplementary Methods
Supplementary Figure S1. Schematic representation of the Iowa gambling task (IGT). On each trial of the IGT, participants chose from one of four decks by pressing the spatially corresponding button on an MR-compatible button box. The decision phase was followed by an anticipation phase (6 s) before the outcome evaluation phase (3 s) displayed the outcome of the decision (win/loss and magnitude). Each trial culminates with a blank screen that took the total trial duration to 15 s. The loan (red bar) and the current running total (green bar) were presented at the bottom of the task display (1)(2)(3).

Computational Modelling Prospect Valence Learning Models
For all three models, outcome evaluation is assessed according to the non-linear prospect utility function which takes into account a diminishing sensitivity to increasing levels of magnitude, as well as a different sensitivity to loss and gain outcomes (4,5). The utility u(t) on trial t of each outcome x(t) is expressed as: The α (0<α<2) parameter determines the shape of the utility function and represents feedback sensitivity. A higher value represents higher sensitivity to the relative magnitudes of outcomes. The λ (0<λ<10) parameter represents loss-aversion, which determines the participant's sensitivity to losses compared to gains. As λ approaches 0, losses are experienced as neutral events. For λ=1, losses and gains have the same impact. When λ>1, losses have greater impact than gains on the subjective utility of an outcome, i.e. the participant prefers to select the alternative with a decreased probability of losses despite lower expected gains.
Ej(t) refers to the expectancy for deck j on trial t. In this model, A is a recency parameter that indicates the weight given to recent outcomes when updating expectancies for the chosen deck. The variable δj(t) is a dummy variable, which is 1 if deck j is chosen or else is equal to 0.
In the decay model, a decay learning rule (7) is used in which expectancies of all unchosen decks are discounted on each trial, and the expectancy of the chosen deck is updated by the current outcome utility. In this model, A is a decay parameter, which indicates the degree to which expectancy is discounted: Next, the PVL models assume that the participants' choices on the trial t+1 are guided by the expectancies of each deck. This is formalized using a softmax function with an inversetemperature parameter: This function contains a trial-independent sensitivity parameter θ that indexes the extent to which trial-by-trial choices match the expected deck utilities, and which is set by the fourth model parameter, which is c (0≤c ≤ 5) or response consistency. Large values of c indicate a tendency to select decks with the highest learned expectancies (exploitation), whereas small values indicate a tendency explore decks with lower expected utilities (exploration).

Value Plus Perseverance Model
Evidence suggests that participants often use a WSLS strategy during IGT (8). In other words, participants sometimes adopt a perseverative strategy that uses information only about the most recent outcome for the most recently chosen deck. A hybrid Value Plus Perseverance (VPP) model was developed combining the PVL-Delta learning rule and WSLS heuristic (8).
This model assumes that individuals track both expectancies (Ej(t)) and perseverance strengths (Pj(t)); expectancies are computed using the learning rule of the PVL-Delta model, and three additional perseverance parameters are included: ω is the reinforcement learning (RL) weight (0<ω<1); a low ω indicates the subject relies less on RL/more on perseverance. Choice probability was again computed using the softmax function, but with Vj(t+1):

Model Comparison
Posterior inference for all models was performed via Markov Chain Monte Carlo (MCMC) sampling implemented in RStan (v2.1.063) (9) (http://mc-stan.org/interfaces/rstan). RStan uses a specific probabilistic sampler called Hamiltonian Monte Carlo (HMC) to sample from the posterior distribution. A total of 3000 samples per chain were drawn after 1000 burn-in samples with 4 chains (10,11). All model parameters of all models had R^ values around 1.00, indicating that chains converged to the target distributions (10). Post-hoc model comparison was conducted using Wanatabe-Akaike Information Criterion (WAIC) (12). This index is obtained by computing the summed point-wise log-likelihood per participant, accounting for the fact that in the IGT, choices on a given trial are dependent on previous choices (13).
Smaller WAIC scores denote better model-fit, and overall fit was assessed by adding WAIC scores from each group for each model (12). Consequently, the model parameters from the winning VPP model were used to compare groups.

fMRI Image Acquisition
The fMRI images were acquired at King's College London, Institute of Psychiatry's Centre for Neuroimaging Sciences on a 3T General Electric Signa Horizon HDx MRI scanner (GE Healthcare, UK) using the body coil for radio frequency transmission and a quadrature birdcage headcoil for reception. In each of 22 non-contiguous planes parallel to the anteriorposterior commissure, 480 T2*-weighted MR images depicting BOLD contrast covering the whole brain were acquired with echo time TE=30 ms, TR=1.5 s, flip angle=60°, in-plane voxel size=3.75 mm, slice thickness=5.0 mm, slice skip=0.5 mm). A whole-brain high resolution structural scan (inversion recovery gradient echo planar image) used for standard space normalisation was also acquired in the inter-commissural plane with TE=40 ms, TR=3 s, flip angle=90°, number of slices: 43, slice thickness=3.0 mm, slice skip=0.3 mm, in-plane voxel size=1.875 mm, providing complete brain coverage. As collected data in patients with OCD included only adolescents aged 12-18 years old, data from control and ADHD groups for older (adults) and younger (pre-adolescents) age ranges were excluded so that groups could be matched on age. One patient with OCD was excluded due to poor quality fMRI data.
Participants completed other tasks during the same session, which have been published elsewhere (1-3, 14-23).

XBAM Analysis
In contrast to theory-based inference, XBAM (24-26) minimizes assumptions and uses median based statistics which controls for outliers. The test statistic is computed by standardizing for individual differences in residual noise before performing second-level analysis using permutation tests. XBAM employs a mixed-effects method that has been recommended following analysis of the validity and impact of theory-based inference in fMRI (27).
Following motion correction, global detrending, spin-excitation history correction and smoothing (full-width at half-maximum (FWHM) 7.2 mm), the residual effects of motion were regressed out from the time series (using the estimated motion parameters) before a standard general linear modelling approach was used to obtain estimates of the response size (beta) to each of the events of interest (advantageous choices, disadvantageous choices, anticipation period, win outcomes, loss outcomes). The anticipation and outcome phases are likely not entirely separable, due to the lack of jitter between the two conditions in the task design. Moreover, it is important to note that on all trials the probability of winning was 50/50, which means that there is no equivalent to a "wins" versus "loss" anticipation contrast in our version of the IGT, as has been examined in tasks such as the monetary incentive delay task (28). Therefore we did not examine the anticipation phase in the current study, and must note as a limitation that brain activation during the outcome phase may be contaminated by brain activation from the anticipation phase due to the haemodynamic delay.
Time series analysis for each subject was conducted using wavelet-based resampling (29). We The appropriate null distribution for assessing significance of any given SSQ ratio was established using a wavelet-based re-sampling method and applying the model-fitting process to the resampled data. The aim was to achieve a global (image-wide) permutation based threshold for forming clusters at p<0.05 at the first stage of cluster analysis. We did this by permuting 20 times per voxel and combining permutations over the whole brain, resulting in 20 null parametric maps of SSQ ratios for each subject, which were combined to give the overall null distribution of SSQ ratio (26,29). As there are approximately 50,000 intra-cerebral voxels in our analysis, this means that we used around 1,000,000 combined permutations to form a global threshold for the first stage of our cluster analysis (18,29). This same permutation strategy was applied at each voxel to preserve spatial correlation structure in the data. Individual SSQ ratio maps were then affine transformed into standard space by normalising onto a Talairach template (30).
For the group-level comparisons, group activation maps were produced for each contrast by calculating the median observed SSQ-ratio over all subjects at each voxel in standard space and testing them against the null distribution of median SSQ-ratios computed from the identically transformed wavelet re-sampled data (24). ANCOVA analyses with group as factor and head displacement in Euclidian 3-D space and age as covariates were performed to compare groups. The necessary combination of voxel and cluster level thresholds was not assumed from theory but rather was determined by direct permutation for each dataset (26).
The voxel-level threshold was first set to p<0.05 to give maximum sensitivity and to avoid Type II errors, as in order to maximize detection power we used the highest threshold that we have shown empirically to give good type I error control at cluster level under the null hypothesis using our permutation-based method (24,26,29). A cluster-level threshold was computed from the data for the resulting three-dimensional voxel clusters such that the final expected number of type I error clusters was <1. Cluster mass rather than a cluster extent threshold was used to minimize discrimination against possible small, strongly responding foci of activation (26).  : -40,-7,-2, voxels=175, p=0. : 4,-4,31, voxels=329, p=0.003) showing a significant negative correlation between age and brain activation during decision making (advantageous>disadvantageous choices) across all participants. Thresholded at p<0.05 for voxel and expected number of type I error clusters was <1. : 11,-74,-29, voxels=495, p=0.003) and right vmOFC/MPFC/IFG (Peak Talairach Coordinates : 29,52,-2, voxels=256, p=0.004) regions showing a significant negative correlation between age and brain activation during outcome processing (wins>losses) across all participants. Thresholded at p<0.05 for voxel and expected number of type I error clusters was <1.

Exploratory Brain-Behaviour and Brain-Performance Correlations
Statistical BOLD response from regions that showed significant group differences were extracted, and correlated with task performance (net score, and VPP parameters shown to differ between groups: c, ω, α) and symptoms scores (ADHD: Conner's T; OCD: CY-BOCS).
Within controls, net score was correlated with activation in vmOFC (BF10=11.7, r(20)=0.61, p=0.005) during the choice phase, such that greater activation during advantageous relative to disadvantageous choices was associated with a greater proportion of advantageous choices.
In addition, using NHST, in controls VS activation was correlated with net score (r(20)=0.49, In ADHD and OCD patients, there were significant correlations between SMA/PCC/precuneus activation and net score. In patients with ADHD, greater activation to disadvantageous choices was associated with a higher net score (r(16)=-0.57, p=0.02), while in OCD patient greater activation to advantageous choices was associated with a higher net score (r (20)=0.58, p=0.01). However, crucially, for all these findings BF10<10, suggesting weak evidence for the tested model over the null hypothesis. Therefore these results should be interpreted with caution. There were no credible or significant correlations between symptoms and brain activation within either ADHD or OCD patient groups (all BF10<10, p>0.05).   Table S3. ANCOVA differences in brain activation between adolescents with ADHD and OCD and healthy comparison adolescents at a liberal whole-brain cluster threshold (p<.05).  Figure S4. Exploratory ANCOVA results for the between-group differences in brain activation for contrast comparing advantageous and disadvantageous choices at a liberal whole-brain cluster threshold (p<.05). Red indicates significant between-group differences in activation between adolescents with ADHD, adolescents with OCD, and healthy comparison adolescents. Talairach z-coordinates are indicated for slice distance (in mm) from the intercommissural line. The right side of the brain corresponds to the right side of the image.

Supplementary
Supplementary Figure S5. Exploratory ANCOVA results for the between-group differences in brain activation for contrast comparing wins and losses at a liberal whole-brain cluster threshold (p<.05). Red indicates significant between-group differences in activation between adolescents with ADHD, adolescents with OCD, and healthy comparison adolescents. Talairach z-coordinates are indicated for slice distance (in mm) from the intercommissural line. The right side of the brain corresponds to the right side of the image.