NeuroImage 59 (2012) 695–707

Contents lists available at ScienceDirect

NeuroImage

j ourna l homepage: www.e lsev ie r.com/ locate /yn img

Fear-relevant outcomes modulate the neural correlates of probabilisticclassification learning

Steven E. Prince, Laura A. Thomas, Philip A. Kragel, Kevin S. LaBar ⁎

⁎ Corresponding author at: Duke University, Box 909USA.

E-mail address: klabar@duke.edu (K.S. LaBar).

1053-8119/$ – see front matter © 2011 Elsevier Inc. Aldoi:10.1016/j.neuroimage.2011.07.027

a b s t r a c t

a r t i c l e i n f o

Article history:Received 12 March 2011Revised 1 June 2011Accepted 7 July 2011Available online 23 July 2011

Keywords:Functional magnetic resonance imagingEmotionFearProcedural memoryProbabilistic learningStriatumMedial temporal lobe

Although much work has implicated the contributions of frontostriatal and medial temporal lobe (MTL)systems during probabilistic classification learning, the impact of emotion on these learning circuits isunknown. We used a modified version of the weather prediction task in which two participant groups werescanned with identical neutral cue cards probabilistically linked to either emotional (snake/spider) or neutral(mushroom/flower) outcomes. Owing to the differences in visual information shown as outcomes, analyseswere restricted to the cue phase of the trials. Learning rates did not differ between the two groups, althoughthe Emotional group was more likely to use complex strategies and to respond more slowly during initiallearning. The Emotional group had reduced frontostriatal and MTL activation relative to the Neutral group,especially for participants who scored higher on snake/spider phobia questionnaires. Accurate performancewas more tied to medial prefrontal activity in the Emotional group early in training, and to MTL activity in theNeutral group later in training. Trial-by-trial fluctuations in functional connectivity between the caudate andMTL were also reduced in the Emotional group compared to the Neutral group. Across groups, reaction timeindexed a switch in learning systems, with faster trials mediated by the caudate and slower trials mediated bythe MTL and frontal lobe. The extent to which the caudate was activated early in training predicted laterperformance improvements. These results reveal insights into how emotional outcomes modulate procedurallearning systems, and the dynamics of MTL-striatal engagement across training trials.

99, Durham, NC 27708-0999,

l rights reserved.

© 2011 Elsevier Inc. All rights reserved.

Introduction

Recent advances in the neuroscience of emotional memory havebeen made by investigating how emotional stimuli modulatedeclarative memory systems and by revealing the mechanisms ofconditioned emotional learning (for a review see LaBar and Cabeza,2006). However, much less is known about the impact of emotion onthe neural systems supporting procedural or habit learning, in whichcognitive or motor task performance improves gradually withfeedback and practice. One influential paradigm for investigatingcognitive aspects of procedural learning is probabilistic classificationlearning (PCL). In a typical PCL task, subjects predict potentialoutcomes based on the information provided by cues shown on agiven trial. For instance, in the weather prediction task (Knowltonet al., 1994, 1996), subjects are shown combinations of 1 to 3 cue cards(out of 4 total) that are probabilistically associated with a binaryoutcome (rain/shine), and they learn what the classifications arethrough performance feedback on a trial-by-trial basis. In these tasks,performance improvement from chance levels is dependent on theintegrity of the striatum, as evidenced by impaired performance in

Parkinson's disease patients (Knowlton et al., 1996). Functionalmagnetic resonance imaging (fMRI) studies have consistentlyassociated striatal regions, in particular the caudate nucleus, withfeedback-based learning in PCL and related tasks (Aron et al., 2004;Poldrack et al., 2001; Shohamy et al., 2004; Tricomi et al., 2004, 2006).

Although the caudate is likely to mediate important aspects oflearning, neuroimaging studies have suggested interactions betweenMTL and/or frontal regions and the striatum during the learningprocess (Moody et al., 2004; Poldrack et al., 2001). Task performancein Parkinson's disease patients may be improved by recruiting theseadditional learning circuits. Indeed, an fMRI study of mild Parkinson'spatients showed that successful performance recruited prefrontal andmedial temporal lobe (MTL) activation while left caudate activationwas reduced (Moody et al., 2004). Although the MTL is implicated infMRI studies of PCL learning (Foerde et al., 2006; Poldrack et al., 2001),evidence from amnesic patients performing PCL tasks is equivocal.One study reports no deficits relative to controls (Reber et al., 1996)with another study reporting deficits only later in learning (Knowltonet al., 1994) and a third study reporting deficits throughout the courseof learning (Hopkins et al., 2004). Differences in task difficulty, patientage, and etiology of amnesia may account for these disparate findings.These studies suggest that the MTL is likely to contribute to the PCLtask itself (perhaps without an impact on performance) but mayprovide additional benefits in later learning or in the transfer oflearned contingencies to new situations.

696 S.E. Prince et al. / NeuroImage 59 (2012) 695–707

Amechanism proposed in several studies is that theMTL facilitatesflexible or conceptual knowledge about the learned task that isexpressed in transfer tasks (Bayley et al., 2005; Foerde et al., 2006;Kumaran et al., 2009; Reber et al., 1996). For example, Bayley et al.found that amnesics could successfully learn (over several weeks) aspecific object discrimination task. Despite similar performance onthe original learned task, only control subjects could apply theirknowledge to a modified version of the task. They conclude thatwithout the function of the MTL, a slower acquisition of knowledge,mediated by the basal ganglia in the form of habit memory cansupport task performance, but the information acquired is rigidlyorganized. Together, the existing findings suggest that while thestriatum provides a critical contribution to the rigid learning aspectsof feedback-based tasks, other regions, including the MTL, facilitateflexible learning that can be applied in other contexts.

Emotional stimuli have the capacity to alter both attention andmemory (Anderson, 2005; LaBar and Cabeza, 2006) and typicallyenhance declarative memory and MTL activation (Dolcos et al., 2004,2005; Kensinger and Corkin, 2004). Survival is likely to benefit fromboth enhanced memory for emotional information as well asprediction of emotional outcomes based on environmental contin-gencies, to the extent that these effects can guide future behavior. Inother cases, however, the presence of emotional stimuli can be apotent distraction and interfere with brain regions subservingongoing tasks (Dolcos and McCarthy, 2006; Morey et al., 2009;Wang et al., 2008). In the context of nondeclarative memory, task-irrelevant emotional stimuli interspersed within an ongoing PCL taskhave been found to interfere with immediate task performance (Steidlet al., 2006). In a behavioral PCL study with emotional and neutraloutcomes (Thomas and LaBar, 2008), we found a deficit in earlylearning in fearful subjects (those with high scores on phobiaquestionnaires) who were shown fear-relevant outcomes, but not infearful subjects shown neutral outcomes. Control subjects showed nosuch impairment with fear-relevant outcomes. Fear-relevant out-comes also promoted the use of complex strategies on the task.

The present study was designed to investigate brain activationduring a PCL task with fear-relevant (snakes or spiders) or neutral(mushrooms or flowers) outcomes. Fear-relevant stimuli were usedbecause theymay provide amodel for understanding habit learning inanxiety disorders. Prediction of outcomes in PCL tasks has been linkedto nondeclarative as well as declarative memory processes (Poldracket al., 2001). Based on the prior literature, a priori regions of interest(ROIs) included the striatum (bilateral caudate and putamen) andMTL (bilateral amygdala, hippocampus, parahippocampal gyrusincluding the entorhinal and perirhinal cortex). The modulationhypothesis (McGaugh, 2004) predicts that emotionally arousingexperiences modulate memory. In human studies, performance ondeclarative tasks has been shown to benefit memory for emotionalinformation (Dolcos et al., 2004, 2005). Performance benefits onnondeclarative tasks might not be reflected in terms of overallaccuracy. Indeed, we predicted equivalent accuracy levels for subjectsassigned to emotional and neutral outcome conditions, based onresults from our prior behavioral study using the same design(Thomas and LaBar, 2008). Nonetheless, the neural mechanismsunderlying learning are hypothesized to be modulated by thepresence of emotional outcomes. Furthermore, individual differencesin fear-relevancy are hypothesized to modulate MTL and striatallearning-related activity.

Because previous classification learning studies have suggestedthat cognitive and neural mechanisms differ between early and latelearning (Knowlton et al., 1994; Poldrack et al., 2001), we wereparticularly interested in investigating group activation differences asa function of learning stage and changes in brain activation over time.Prior research has suggested that the contribution of striatal and MTLmechanisms change over the course of learning, with hippocampalactivity emerging early and dissipating over time whereas the

striatum shows the opposite pattern, perhaps reflecting the relativedegrees of declarative and procedural learning processes engagedthrough competitive interactions (Poldrack et al., 2001; Poldrack andPackard, 2003).Wewere interested in tackling this issue from anotherangle by using RT as a proxy for the relative amount of deliberativeprocessing engaged on a trial-by-trial basis. Thus we predicted thatMTL activity would be stronger on long RT trials whereas striatalactivity would be stronger on short RT trials, reflecting the relativeweighting of these regions in extracting underlying contingenciesversusmore automatic expression of learned relationships. Finally, weused activation in these ROIs during early learning to predict themagnitude of performance improvement later in learning. It would bereasonable to speculate that both the MTL and striatum may set thestage for improved learning over time. However, a previous fMRIstudy that analyzed the cue portion of probabilistic classification trialsreported high levels of caudate activation early in learning that werereduced later, once learning had occurred (Delgado et al., 2005). Wetherefore speculated that caudate nucleus activation early in learningmay predict improvement later in learning.

Methods

Subjects

Participants were local residents recruited through the BrainImaging and Analysis Center at Duke University Medical Center andwere reimbursed at the rate of $20/h. The Institutional Review Boardat Duke University approved the experimental protocol and humansubject procedures. A total of 43 subjects provided informed writtenconsent to participate in this study and were randomly assigned toeither the Emotional or Neutral group. A between-groups design wasnecessary to employ because of strong practice effects withprobabilistic classification learning and because a single, binary choicetask across emotional and neutral outcomeswould be uninterpretable(due to the probabilistic structure, all cue cards would be associated tosome degree with both emotional and neutral outcomes) (Thomasand LaBar, 2008). All participants were screened by a self-reportquestionnaire for history of neurologic and psychiatric illness,substance abuse, current psychotropic medication use, and fordepression by the Beck Depression Inventory (Beck et al., 1996). Nosubject scored within 2 SD's of the phobic norms on questionnairesassessing attitudes towards snakes and spiders (Klorman et al., 1974).Following Aron et al. (2004), individuals who did not score abovechance after the first 50 trials were not included in the final analyses(‘non-learners’, 8 from the Emotional group, 7 from the Neutralgroup). To ensure that the groups did not differ on other emotionalcharacteristics, questionnaires were administered assessing emotion-al experience (Positive and Negative Affect Schedule; (Watson et al.,1988)), affect intensity (Affect Intensity Measure; Larsen, 1984), andcurrent stress levels (Daily Stress Inventory; Brantley & Jones, 1993),which showed no differences between groups (all F'sb2.0). Data from3 subjects were discarded because of MRI signal problems, leaving 12subjects in the emotional condition (Mage=22.1, 7 females) and 13subjects in the neutral condition (Mage=23.0, 3 females). Age was notsignificantly different between groups (t(23)=−0.6, p=.55). A chi-square test showed that the gender balance was not significantlydifferent across groups (Fisher's exact p=.11). Due to the smallsample size, gender was not probed further in the analyses.

Stimuli

The card cue and outcome stimuli are the same as described in ourprevious behavioral study (Thomas and LaBar, 2008). Briefly, eachcard cue contained a unique shape (square, diamond, circle, circlewith arrow in center) arranged in a 5×3 grid of rows and columns.Auditory feedback (described below) and visual feedback followed

697S.E. Prince et al. / NeuroImage 59 (2012) 695–707

the response to the cue. Visual feedback consisted of viewing one ofsix exemplars from each category; different exemplars were used inorder to minimize habituation effects over the 100 training trials.Low-level visual properties of the pictures were equated across theexemplars. Subjects in the emotional condition saw snakes andspiders whereas subjects in the neutral condition saw mushroomsand flowers. Pictures were selected based on normative values (Langet al., 1997) to be of lower valence for the emotional condition thanthe neutral condition. Valence ratings were provided following theexperiment (unavailable from 1 subject in each condition) and therewere no significant differences within groups between the specificstimulus categories used (snakes versus spiders, t(10)=1.24, p=.24,or mushrooms versus flowers t(11)=0.99, p=.34).

Study design

The task design was modeled after that used by Aron et al. (2004)(Fig. 1). Between one and three (out of four) cue cards appeared onthe screen at a time, comprising 14 possible cue patterns. Thesepatterns were associated with two outcome categories in a probabi-listic manner. For example, one pattern had cue cards 2, 3, and 4present, and appeared 4 times (4% of the total trials). The probabilitythat outcome A occurred with this pattern was 75%, whereas theprobability that outcome B occurred was 25%. Since outcome Aoccurred over 50% of the time, this outcome was considered ‘correct’.Participants were randomly assigned to receive either the fear-relevant (snake/spider) or neutral (flower/mushroom) outcomes.Participants completed two runs of 50 trials each.

On each trial, one of the 14 card patterns appeared and remainedon the screen for 4 s, at which time the subject was prompted torespond with a left button press for outcome A and a right buttonpress for outcome B. Participants then heard a high-frequencyfeedback tone (duration=1000 ms) when they predicted the trialoutcome and four 100 ms bursts of white noise at 80 dB when theydid not predict the trial outcome (Aron et al., 2004; Knowlton et al.,1996; Shohamy et al., 2004). Similar to Thomas and LaBar (2008), theoutcome photo was displayed in a dynamic fashion, first appearingsmall in the center of the screen for 200 ms and then appearing at fullscreen for 800 ms to create a looming effect toward the viewer. This

Fig. 1. An example trial for the Neutral and Emotional groups. Subjects were assigned to eith(snake/spider outcome photographs). The flower and snake images shown are from the publdetermined by the subject, with a maximum of 3000 ms. Feedback consisted of 1000 ms alooming manner (200 ms small, 800 ms large). Trials were separated by an inter-trial inter

was done in order to increase the emotional impact of the outcome.Following outcome presentation, there was a jittered 2–5 s fixationscreen inter-trial interval (Fig. 1). The first 25 trials were pseudo-randomized such that an equal number of patterns appeared thatwere ‘easy’ (highly predictive) or ‘hard’ (less predictive). Thisprocedure was conducted to reduce the number of non-learners, asindicated by pilot testing. The following 75 trials were fullyrandomized.

Procedure

Before scanning, subjects briefly practiced 5 random PCL trials tofamiliarize themwith the task requirements. Instructions appeared onthe screen prior to the practice trials. After the instructions, structuralMRI scans were obtained. Then the two functional scans were run (50trials each, 10.5 min duration maximum per run), with a short breakbetween scans. Subjects used right index and middle fingers to pressbuttons on the MR-compatible button box.

MRI acquisition

Scanning was performed on a General Electric 4 T LX Nvi MRIscanner system equipped with 41 mT/m gradients (General Electric,Waukesha, Wisconsin, USA). Scanner noise was reduced withearplugs, and head motion was reduced with foam pads. Stimuliwere presented with liquid-crystal display goggles (ResonanceTechnology, Northridge, CA), and behavioral responses were recordedwith a four-key fiber-optic response box (Resonance Technology). Aquadrature birdcage radio frequency (RF) head coil was used totransmit and receive. Sixty-eight high-resolution structural imageswere acquired using a 3D fast SPGR pulse sequence (TR=500 ms;TE=20 ms; FOV=24 cm; image matrix=2562; voxel size=1 mm×1 mm×1.9 mm). These structural images were acquired inthe near axial plane defined by the anterior and posterior commissures.Whole brain functional images were acquired using a gradient-recalledinward spiral pulse sequence sensitive to blood oxygenation leveldependent (BOLD) contrast (TR=2000 ms; TE=31ms; FOV=24 cm;image matrix=642; α=60°; voxel size=3.75×3.75×3.8 mm; 34contiguous axial slices). This protocol is effective at reducing MRI-

er the Neutral group (flower/mushroom outcome photographs) or the Emotional groupic domain and are for illustration purposes only. The duration of the response phase wasuditory (tone for correct, white noise bursts for incorrect) and 1 s visual display in aval (ITI) of 2000–5000 ms with a fixation cross in the center of the screen.

Table 1Behavioral results: mean (SD).

Subjects Emotional Neutral

12 13

Performance (% correct)Early learning 64.4 (10.5) 65.1 (9.7)Late learning 68.4 (12.5) 65.6 (13.5)Difference 3.9 (10.1) 0.05 (13.8)

Reaction time (ms)Early learning 765 (159) 621 (78)Late learning 637 (185) 593 (108)Difference 128 (156) 28 (105)

Strategy (% complex/%simple/%unknown)Early learning 75/25/0 53.8/38.5/7.7Performance early 65.7 (11.1)/60.7(9.2) 72(4.4)/58.8(7)/48Late learning 91.7/8.3/0 61.5/38.5/0Performance late 69.9(12)/52 72.5(12.3)/54.6(5.6)

Phobic scoreSnake (out of 30) 4.25 (4.3) n/aSpider (out of 31) 6.17 (7.2) n/aComposite (out of 61) 10.42 (9.5) n/a

698 S.E. Prince et al. / NeuroImage 59 (2012) 695–707

induced signal artifacts in frontolimbic regions at high field strength.These functional images were acquired in a similar orientation to thestructural images. A semi-automated high-order shimming programensured global field homogeneity. Runs consisted of the acquisition of310 brain volumes and began with 4 discarded RF excitations to allowfor steady state equilibrium.

FMRI data analysis

Statistical Parametric Mapping (SPM5, http://www.fil.ion.ucl.ac.uk/spm/software/) was used for preprocessing and analysis. Theimages were realigned and spatially smoothed using an 8-mm full-width half-maximum Gaussian kernel. Translational movementparameters never exceeded 0.5 of a voxel in any subject per run. Anonlinear high-pass filter with a 128-s cut-off was used to temporallyfilter the data. A twelve-step affine linear transformation procedurewas used and functional images were registered to standard MontrealNeurologic Institute (MNI) space.

After preprocessing, statistical analyses were performed at thesingle-subject level by using the general linear model within SPM.Each cue card presentation was modeled as an impulse convolvedwith a canonical hemodynamic response function (HRF). Feedbackwas modeled separately from the cue. The inter-trial interval, whichwas not explicitly modeled, served as an intrinsic baseline. Specificcomparisons of interest were tested by using linear contrasts. Afteranalysis at the individual level, the results were spatially normalizedto the MNI template using SPM's registration tool for group effectanalyses. Mixed-effects group analyses were performed for eachcontrast by using SPM random effects with factorial designs andregressionwith individual differencemeasures. Anatomical ROIs wereinvestigated using the Wake Forest University PickAtlas (Maldjianet al., 2003; Tzourio-Mazoyer et al., 2002). These included theamygdala, hippocampus, parahippocampal gyrus, and putamen andwhich were selected from the AAL atlas and the caudate nucleuswhich was selected from the TD brodmann areas+atlas because offurther separation into head, body and tail of the caudate. Functionalconnectivity analyses used ROIs from the AAL atlas for the caudate andMTL. All analyses were conducted with particular focus on thehypothesis driven ROIs and whole-brain analyses additionallyreported for exploratory purposes.

The contrasts reported here include: 1) overall activations anddeactivations in early (first 50 trials) and late (last 50 trials) learningruns and differences between Emotional and Neutral group activations;2) correlation with phobia questionnaire composite score in theEmotional group; 3) correct versus incorrect responses (based on themajority outcome percentage as described in Study design section);4) parametricmodulation of reaction time; 5) early learning correlationwith performance change; and 6) correlation with average perfor-mance. For clusters in our a priori areas of interest (MTL, dorsalstriatum),we used pb .05 family-wise error correctionwithin individuallateralized structures. Whole-brain exploratory analyses were thre-sholded at pb .001 uncorrected, k=5 voxels. Calculations for spatialextent correction for multiple comparisons were done using the RESTAlphaSim utility (www.restfmri.net; toolkit V1.3), which performssimulations in the same manner as the AFNI software version. Alphalevels were computed for each contrast mask with 10,000 Monte Carlosimulations, an individual voxel threshold probability of .001, a clusterconnection radius of 5, and a 4 mm FWHM smoothness. In all cases, theresults of these simulations yielded anαb .001 FWE rate. The contrast ofcorrect versus incorrect trials was additionally inclusively masked withthe group effect (pb .05) that was presumed to be driving theinteraction, which is more statistically conservative. For example,Emotional (correctN incorrect) versus Neutral (correctN incorrect) wasmaskedwithEmotional (correctN incorrect) tomitigate thepotential forincorrect Neutral trials to drive group differences. The parametricallymodulatedRT contrastswere basedon entering a regressor (normalized

by SPM) for each subject, for each trial (correct and incorrect), acrossearly and late learning. Random-effects analysis, which included groupas a factor were constructed to weight the regressor in the positivedirection (+1 for both groups=weighted towards increasing RTs) ornegative direction (−1 for both groups=weighted towards decreas-ing RTs). Functional connectivity analyses were conducted usinganatomical masks of the left and right caudate nucleus with MTLsubstructures bilaterally. Single-trial analysis time-locked to cue cardonset was used to extract beta parameter estimate values and computecorrelation scores between structures (one score for the first 50 andanother for the second 50 trials). These scores were then Fisher-transformed and entered into separate repeated-measures ANOVAswith factors for group (Emotional, Neutral), hemisphere (right, left),learning phase (early, late), and structure (amygdala, hippocampus,parahippocampus).

Results

Behavioral results

Table 1 lists reaction time, performance information, strategypercentage and performance given strategy used, for both theEmotional and Neutral groups. It also contains the phobic score(snake, spider, and composite) for the subjects in the Emotionalgroup. Composite phobic score is the index of individual differences infear-relevancy in this group.

Reaction time (RT)RT data (across correct and incorrect trials) were analyzed using

a mixed ANOVA with Run (1 or 2) as a within-subjects factor andgroup (Emotional or Neutral) as a between-subjects factor. Therewas a significant main effect of run, F(1, 23)=8.71, pb .01, withsubjects performing faster in Run 2 than Run 1. This was especiallytrue for the Emotional group, t(11)=−2.80, pb .05 but not theNeutral group, t(12)=−0.96, p=.36. Correlation analysis withphobic score revealed no significant effects for Run 1, Run 2, or thedifference between runs (all pN0.87).

Learning rateA mixed ANOVA with run as a within-subjects factor and group as

the between-subjects factor revealed no significant effects (all Fsb1).

699S.E. Prince et al. / NeuroImage 59 (2012) 695–707

For both groups, comparing performance to chance level yieldedsignificant effects (all pb0.001) for both Run 1 and Run 2.

Implicit learning strategyImplicit learning strategies were evaluated in a similar manner to a

previous behavioral study (Thomas and LaBar, 2008). Briefly, usingmathematical models to fit each subject's data, performance wascompared to the ideal use of three strategy types (simple, complex,nonidentifiable) which vary in the amount of integrative processingand number of cues used. In particular, simple strategy useencompassed both singleton and one-cue strategies whereas complexstrategies included bothmultimatch and multimax strategies (Lagnadoet al., 2006). Separate strategy analyses were conducted for each run ineachgroup.A greaterproportionof subjects in theEmotional groupusedcomplex vs. simple strategies in both early (75% vs. 25%) and latelearning (91.7% vs. 8.3%). In the neutral group, complex vs. simplestrategy use was more similar in early learning (54% vs. 39%, 1 subjecthad a nonidentifiable strategy)with a greater proportion using complexstrategies in late learning (61% vs. 39%).

To test for the influence of strategy use on performance, weexamined whether subjects using simple and complex strategiesperformed equivalently. Separate ANOVAs conducted for each run,with both group and strategy as between-subjects factors, revealed that

Table 2Brain regions associated with A) group differences and B) within-group regression of phob

A. Cue-related activity

Early learning

Emotional vs. Neutral H BA x y z

ROI No suprathreshold clustersWhole brain No suprathreshold clusters

Early learning

Neutral vs. Emotional H BA x y z

ROICaudate body L – −15 15 1Amygdala R – – – –

Amygdala L – – – –

Whole brainDorsal PFC L 9 −15 53 3Temporal R 39 53 −75 1Temporal R 22 49 −56 1

B. Phobia composite score regression (emotional group only)

Early learning

Correlation H BA x y z

Positive-ROI No suprathreshold clusters

Positive-whole brainOccipital R 19 38 −64 4Parietotemporal L 40 −53 −53 2Temporal R 39 38 −49 8Dorsal frontal R 6 19 15 6Temporal R 21 45 −34 −

Negative-ROICaudate body L – −15 19 8Caudate head L −15 23 4Amygdala L – – – –

Negative-whole brainCaudate L – −15 19 8Occipital R 18 – – –

Occipital R 17 – – –

Occipital R 19 – – –

H = hemisphere, BA = Brodmann area (nearest gray matter), PFC = prefrontal cortex, x, y⁎ FWEb .05.

subjects using complex strategies performed better than those usingsimple strategies. For Run 1, F(1, 21)=5.70, pb .03, Msimple=60%,Mcomplex=68%; for Run 2, F(1, 21)=7.30, pb .02, Msimple=54%,Mcomplex=71%.

fMRI results

Standard analysesContrasts included cue-related activity for all trials, with one contrast

for early and one for late learning. Across all subjects, early learningversus the intrinsic baseline (fixation ITIs) yielded a broad set of regionsin visual, frontoparietal, anterior cingulate insular cortices, posteriorhippocampus, right caudate and bilateral putamen. The reverse contrastyielded default mode network regions including medial parietal,prefrontal cortices, mid-hippocampus and parahippocampal regions(see Supplemental Table 1).Within theROIs, theNeutral groupdisplayedgreater activation within the left caudate nucleus compared to theEmotional group (see Table 2A). Repeated measures ANOVAs on betavalues extracted across both early and late learning from the caudatepeak revealed a significant main effect of group (pb .05) and nosignificant group×time interaction. Contrasts from late learning yieldedvery similar results to early learning, across all subjects (see Supple-mental Table 1). However, within the MTL/striatum regions of interest,

ia composite score.

Late learning

Z value x y z Z value

No suprathreshold clustersNo suprathreshold clusters

Late learning

Z value x y z Z value

5 2.96⁎ – – – –

– 23 0 −19 2.74⁎

– −30 0 −23 2.58⁎

8 3.65 No suprathreshold clusters9 3.459 3.37

Late learning

Z value x y z Z value

No suprathreshold clusters

4.55 No suprathreshold clusters7 4.05

3.621 3.5811 3.51

3.64⁎ −15 11 19 2.98⁎

3.26⁎

– −19 −4 19 3.55⁎

3.64 – – – –

– 19 −90 15 3.51– 15 −90 −4 3.43– 41 −83 −15 3.29

, z coordinates in MNI space.

700 S.E. Prince et al. / NeuroImage 59 (2012) 695–707

group differences (Neutral greater than Emotional) were found in theright and left amygdala (see Table 2A). Repeated measures ANOVAs onbeta values extracted across both early and late learning from theamygdala peaks revealed a significant main effect of group, and agroup×time interaction in the right amygdala (both p'sb .05) and trendsfor main effect of group, and a group×time interaction in the leftamygdala (both p'sb .10). In summary, analyses revealed group

Early learningx=-15, y=15, z=15

x=-15, y=19, z=8

Emotional Group-Phob

Neutral vs. Emotion

A B

C D

Fig. 2. Activation in early (2A and 2C) and late (2B and 2D) learning, which correspond to fMfor purposes of display at pb0.01, extent threshold of 3 voxels, within region of interest masputamen. A) Group comparison, NeutralNEmotional in early learning, focused on a regionfocused on bilateral amygdala. C) Within-group (Emotional) negative correlation with phob(Emotional) negative correlation with phobia score in late learning, showing a region of th

differences in the caudate and amygdala, with greater activation for theNeutral than the Emotional group.

Correlation with composite phobia score in emotional group subjectsSimple regression contrasts (within Emotional group subjects)

included the composite phobia score to assess the relationshipbetween individual differences in fear relevancy and cue activity.

Late learningx=23, y=0, z=-19

ic Score Correlation

x=-15, y=-4, z=-19

al Group

RI runs comprising the first 50 and second 50 trials, respectively. All contrasts are shownks consisting of bilateral amygdala, hippocampus, parahippocampal gyrus, caudate, andof the left caudate nucleus. B) Group comparison, NeutralNEmotional in late learning,ia score in early learning, showing a region of the left caudate nucleus. D) Within-groupe left caudate nucleus and left amygdala.

701S.E. Prince et al. / NeuroImage 59 (2012) 695–707

During early learning (Run 1), several regions positively correlatedwith phobia score, including temporal and frontoparietal regions (seeTable 2B). Negative correlations were found in the left caudate (seeFig. 2C). During late learning (Run 2), left caudate, left amygdala andseveral areas in right visual cortex were negatively correlated withphobia score (see Fig. 2D). No suprathreshold voxels were found forpositive correlations with phobia score in late learning. Thus, reducedactivation in the caudate and amygdala for the Emotional group(Standard analyses section) is further magnified by individualdifferences in fear relevancy.

Correct versus incorrect trialsComparing activation on correct vs. incorrect responses provides a

neural metric of learning accuracy, which in turn may be influencedby strategy use. Neural activation related to accurate performance(based on the majority outcome percentage as described in Studydesign section) was assessed for the Emotional versus Neutral group,both for early and late learning. Medial prefrontal (mPFC) activationwas differentially associated with correct performance in theEmotional group in early learning (see Table 3, Fig. 3). Precuneusand right anterior temporal lobe activations were differentiallyassociated with correct performance in the Neutral group in latelearning. Within the ROIs, MTL regions were differentially related tocorrect performance in the Neutral group (Fig. 3B). In sum, accurateperformance was associated with mPFC for the Emotional group inearly learning and the MTL, precuneus and anterior temporal lobe forthe Neutral group in late learning.

Functional connectivitySingle-trial responses (beta parameter estimates) were extracted

from bilateral caudate andMTL ROIs in order to test whether functionalconnectivity differed for the Emotional and Neutral groups. Repeated-measures ANOVAs for left caudate-MTL connectivity and right caudate-MTL connectivity both yielded a significant main effect of group(NeutralNEmotional) (both pb0.02) (Fig. 4). Post-hoc analyses bysubregion revealed reduced connectivity between the right caudate andamygdala, relative to the hippocampus (pb .001) and parahippocampalgyrus (pb .01). To ensure that global correlations were not driving anygroup differences, we used the same methods to compare connectivitybetween the left and right putamen, which revealed no significanteffect for group (p=.88). In order to test for a relationship between

Table 3Brain regions showing group differences in responses to correct vs. incorrect trials.

Correct vs. incorrect trials Early learning

Emotional vs. Neutral H BA x y z Z

ROI No suprathreshold clusters

Whole brainMedial PFC M 10 4 56 4 3Dorsomedial PFC M 9\10 0 56 30 3

Early learning

Neutral vs. Emotional H BA x y z Z

ROIHippocampus L – – – – –

Amygdala L – – – – –

Parahippocampus R – – – – –

Whole brainAnterior temporal lobe R 21 – – – –

Precuneus M 31 – – – –

Anterior cingulate L 32 – – – –

Precuneus R 31 – – – –

Dorsomedial PFC M 9\10 – – – –

H = hemisphere, M = midline (x value between ±8), BA = Brodmann area (nearest gray⁎ FWEb .05.

connectivity and behavioral performance, separate post-hoc correlationanalyses were conducted for Emotional and Neutral groups for eachlearning stage. Significant correlations between connectivity andperformance were only found for the Neutral group in early learning(left caudate-MTL with performance r(11)=.58, pb .05, right caudate-MTL with performance r(11)=.56, pb .05). Overall, connectivitybetween the caudate and MTL was reduced in the Emotional grouprelative to the Neutral group, and a significant relationship betweencaudate-MTL connectivity values and performance was only found forthe Neutral group in early learning.

Parametric modulation by reaction timeAcross learning trials, changes in RT could reflect shifts in cognitive

processing from more deliberate to more automatic expressions oflearned relationships. We therefore combined all subjects and alltrials in order to assess global activation changes related to faster orslower responses. To account for potential differences in RT, theparametrically modulated SPM analyses were computed in a factorialdesign, with group (Emotional, Neutral) as the factor.Within the ROIs,the left caudate (see Table 4 and Fig. 5A) was associated with fasterRTs and bilateral MTL regions (see Table 4 and Fig. 5B) wereassociated with slower RTs. Repeated measures ANOVAs on betavalues extracted (from the standard analyses) across both early andlate learning from the caudate peak revealed no main effect of group(p=.93) but a trend toward a group×time interaction (pb .06)whereas there were no significant effects for the MTL peaks (allp'sN .18). Paired t-tests revealed a trend for greater activation in lateversus early learning in the caudate peak in the Emotional group(pb .09), while the Neutral group exhibited no difference (p=.86). Inthe whole brain analysis of parametric modulation for decreasing(faster) RT only the left caudate was significant. The reverse contrastfor increasing (slower) RT revealed primarily temporal and frontalregions (see Table 4). Analysis of all learning trials across both groupsrevealed the caudate and MTL to be modulated by RT in opposingdirections Table 4.

Correlation of early learning activation with performance changeActivation during early learning may set the stage for future

learning and reflect individual differences in performance improve-ment/decline. This contrast employed regression with the magnitudeof performance change (late learning percent correct minus early

Late learning

value x y z Z value

No suprathreshold clusters

.40 No suprathreshold clusters

.26

Late learning

value x y z Z value

−23 −8 −19 3.60⁎

−23 −4 −19 3.32⁎

26 −23 −27 3.20⁎

45 8 −38 4.24−4 −68 27 3.6815 38 11 3.6711 −56 30 3.614 56 34 3.36

matter), PFC = prefrontal cortex, x, y, z coordinates in MNI space.

B. Neutral vs. Emotional (Correct>Incorrect trials)in Late Learning

A. Emotional vs. Neutral (Correct>Incorrect trials)in Early Learning

x=-1, y=52, z=11

x=-27, y=-19, z=-19

Fig. 3. Interaction between group (Emotional, Neutral) and correctness (correct,incorrect). A) Emotional vs. Neutral group for correctN incorrect trials in early learning,showing two regions in medial PFC (mPFC). Image shown at pb0.001, extent thresholdof 5 voxels. B) Neutral vs. Emotional group for correctN incorrect trials in late learning,showing bilateral MTL regions. Contrast shown for purposes of display at pb0.01,extent threshold of 3 voxels, within a region of interest mask consisting of bilateralamygdala, hippocampus, parahippocampal gyrus, caudate, and putamen.

702 S.E. Prince et al. / NeuroImage 59 (2012) 695–707

learning percent correct) across all subjects to assess the relationshipbetween specific activations and the expression of task learning.Within the ROIs, left and right caudate nucleus were positivelycorrelated with performance change (Fig. 5C). In the whole-brainanalysis, regions that positively correlated with performance change(see Table 5) included left visual cortex, right cingulate, and frontalcortex. No suprathreshold clusters were found for the negativecorrelation. In sum, across all subjects, greater activation in the

caudate nucleus early in learning was related to greater improvementin task performance.

Correlation of activation with average performanceIndividual differences in performancemayaffect brain activation over

the entire course of learning. This contrast employed regressionwith theaverage performance for each run (for each subject) across all subjects.Within the ROIs, the left hippocampus was negatively correlated withaverage performance (see Fig. 5D). In the whole-brain analysis, occipitalcortexwas positively correlatedwith average performancewhile severalfrontal and parietal regions and the cerebellum were negativelycorrelated with average performance (see Table 6). The peak was verysimilar to the overall deactivations reported in the hippocampus in earlyand late learning (see Supplemental Table 1). Extracted beta values fromthe hippocampal peak of this analysis were negative (mean=−.51,Emotional=−.73, Neutral=−.31). A t-test of group values showed atrend for an effect of greater deactivation in the Emotional than Neutralgroup, t(23)=−2.1, p=.05.

Discussion

Emotion effects

To test the effects of emotion on neural responses duringclassification learning, we compared data from the cue period of anidentical PCL task across groups who received either Emotional(snakes and spiders) or Neutral (mushroom, flowers) outcomes. Thepresence of fear relevant outcomes during the PCL task was associatedwith reduced recruitment of the caudate and amygdala (Figs. 2A andB). Notably, individual differences in fear relevancy were associatedwith reduced activation in similar regions. In the Emotional group,negative correlations were found between the composite phobiascore and left caudate activation in early learning and left amygdalaand caudate in late learning (Figs. 2C and D). Correct responses wereassociated with mPFC activation in the Emotional group in earlylearning (Fig. 3A) and the MTL in the Neutral group in late learning(Fig. 3B). Furthermore, the Emotional group displayed significantlyweaker connectivity between the caudate and MTL compared to theNeutral group (Fig. 4). Taken together, the results suggest thatcircuitry typically associated with PCL tasks is disrupted when fearrelevant outcomes are integrated into the task.

Emotion and MTL/striatum activation differences

The Emotional group showed decreased activation in ROIs in bothearly and late learning, characterized as decreased left caudateresponse in early learning and decreasedMTL (amygdala in particular)in late learning. The caudate nucleus has been associated broadly withhabit learning and value-based decision making (Daw et al., 2005,2006; Delgado et al., 2005, 2008; Graybiel, 2005; Shohamy et al., 2008;Tricomi et al., 2004, 2006). Although the amygdala result may seemsurprising at first glance, recent studies in humans and animals havereported a role for the amygdala in the coding of upcoming feedback inresponse to neutral cues (Bischoff-Grethe et al., 2009; Kahn et al.,2002; Paton et al., 2006; Spiegler andMishkin, 1981). Another study inrats using a stimulus–response habit task found a double dissociationbetween amygdala and striatal function (McDonald and Hong, 2004).While the dorsal striatumwas required for habit learning to occur, theamygdala was required in order to develop a preference for thereinforced stimulus. In the present study, the Emotional group had asignificantly lower overall amygdala response (Fig. 2B), and reducedassociationwith correctness (Fig. 3B) during late learning. The currentdata support the idea that during early learning (first 50 trials), bothbehaviorally and neurally, the group receiving emotional outcomephotographs was not engaged in habit-based learning to the samedegree as the group receiving neutral outcome photographs. The fact

0

0.15

0.3

0.45

0.6

0.75

0.9

1.05

1.2

Fis

her

-tra

nsf

orm

ed C

orr

elat

ion

Va

lue

Left Caudate Right Caudate

Emotional

Neutral

**

Caudate-MTL Connectivity

Fig. 4. Functional connectivity between left caudate nucleus (anatomical mask) andMTL regions and right caudate nucleus andMTL regions, as a function of group. Separate ANOVAswere run for the left and right caudate, with factors for hemisphere (left MTL, right MTL), learning phase (early, late), structure (amygdala, hippocampus, and parahippocampalgyrus), and group (Emotional, Neutral). No significant main effects were found for hemisphere or learning phase. A significant effect of structure for the right caudate is detailed inthe text. However, no significant interactions were found for either the left or right caudate. For simplicity, groupmain effects are shown averaged across hemisphere, learning phaseand structure.

703S.E. Prince et al. / NeuroImage 59 (2012) 695–707

that an aversive visual stimulus was the outcome in the Emotionalgroup, whether the subject's response was correct or not, may haveremoved an incentive of feedback prediction and thus blunted theamygdala's response. For the Neutral group, it is possible that theamygdala and other associated MTL regions performed a similarfunction in this task as they might in the episodic domain, namely toassociate trial level information during the cue with a particularoutcome. In both early and late learning, despite equivalent behavioralperformance, activation levels in regions strongly linked to feedbackbased habit learning were reduced in the Emotional group.

Emotion and functional connectivity between the MTL and striatum

Given the evidence for striatal and MTL system interactions duringlearning, we assessed group differences in trial-by-trial functionalconnectivity between the caudate nucleus and substructures of theMTL.Bilateral caudate connectivity with MTL substructures was significantlyreduced for theEmotional group relative to theNeutral group.Disruptedfunctional connectivity could mean that the Emotional group was lessable to maximize the contribution of each system. Task performance in

the first 50 trials was significantly correlated with functional connec-tivity in the Neutral but not Emotional group, providing evidence for aninfluence of emotion on interactions between the striatal and MTLsystems related to performance in PCL tasks. The connectivity analysessuggest that a pattern that was beneficial to the Neutral group had norelation to performance and was not expressed to the same degree bythe Emotional group.

Emotion and task behavior

A previous behavioral study of emotional versus neutral informationin a PCL task reported initial impairments that equalized with morelearning (Steidl et al., 2006). Their paradigmused interspersed emotionalpictures during the weather prediction task and therefore may have hada distracting influence (Anderson, 2005; Dolcos and McCarthy, 2006). Inour previous behavioral study, fearful subjects shown emotionaloutcomes, but not neutral outcomes, had deficits in early learning andaltered strategy use (Thomas and LaBar, 2008). The emotionalinformation in our paradigm was integrated into the flow of the task,potentially reducing any distracting influence in non-fearful subjects;

x=-13 y=-23, z=-23

A. Faster RTs B. Slower RTs

C. Positive correlation between early learning activation and

performance change

y=-19, z=-15

D. Negative correlation between activation and performance

x=-15

0

0.5

1

1.5

2

2.5

3

3.5

0

0.5

1

1.5

2

2.5

3

3.5

0

0.5

1

1.5

2

2.5

3

3.5

Fig. 5. Additional activations as a function of behavioral metrics. Top panel: Analysis of all subjects with parametric modulation by reaction time (RT). Bottom panel: Regressionanalysis of early learning activation with behavioral performance change. Contrasts are shown for purposes of display at pb0.01, extent threshold of 3 voxels, within a region ofinterest mask consisting of bilateral amygdala, hippocampus, parahippocampal gyrus, caudate, and putamen. A) Contrast weighted toward faster RTs, showing a region of the leftcaudate nucleus. B) Contrast weighted toward slower RTs, showing bilateral medial temporal lobe (MTL) regions. C) Positive correlation shown from contrast of early learningactivation (first 50 trials) correlated with change in behavioral performance (late percent correct minus early percent correct). D) Negative correlation shown from contrast of allactivation (100 trials) correlated with behavioral performance.

704 S.E. Prince et al. / NeuroImage 59 (2012) 695–707

however, as described above, the presence of the fear-relevant outcomesmay have blunted the amygdala's response to feedback incentives. Thestrategy analyses in the present study suggest that the Emotional groupwas able to use complex strategies even early in learning despite anassociated RT cost. Therefore, the neural circuitry typically employed insolving PCL tasks may have been relied on to a lesser extent, with aconcomitant shift to alternate neural circuitry, as discussed furtherbelow. In other tasks, multiple memory systems are tested after theinduction of stress, with a common finding of bias towards proceduralresponse strategies (Schwabe et al., 2010). In contrast to those findings,our fear-relevancy manipulation might bias subjects toward morecomplex strategies. Although the neural correlates associated with thetask differ between groups, any costs of activation differences are

ameliorated in terms of overall behavioral performance and in thisregard, fear-relevant stimuli may provide a benefit.

Performance measures from our behavioral study (Thomas andLaBar, 2008) suggest that there may be an optimal level of fearfulnesstowards fear-relevant stimuli, similar to the relationship expressed bythe Yerkes–Dodson law (Yerkes and Dodson, 1908). In the non-fearfulsubjects of the present study, only negative correlations were foundwithin the striatal and MTL ROIs. However, at the level of the wholebrain, the composite phobia scorewas positively correlatedduring earlylearningwith regionsassociatedwith enhancedattention. These includeoccipital, temporal and frontal regions implicated in spatial attentionand workingmemory (Corbetta et al., 1998; Corbetta et al., 2002; LaBaret al., 1999). Furthermore, differential mPFC recruitment by the

Table 4Brain regions, across all subjects, associated with trial reaction time as a parametricregressor.

Parametric modulation by reaction time

Faster H BA x y z Z value

ROICaudate body K – −8 15 11 3.14⁎

Caudate body L – −11 19 8 2.93⁎

Whole brainCaudate L – −11 23 8 3.45Slower H BA x y z Z valueROIParahippocampus R – 26 −23 −23 3.54⁎

Parahippocampus L – −26 −23 −19 3.14⁎

Hippocampus L – −23 −19 −19 3.13⁎

Whole brainLateral parietal L 39\40 −60 −53 30 4.71Lateral temporal L 39 −53 −71 15 3.98Ventrolateral PFC L 47 −41 30 −19 3.81Dorsal PFC L 9 −11 56 30 3.77Cuneus R 18 11 −71 0 3.66Ventrolateral PFC L 44\45\47 −45 23 8 3.63Parahippocampus R 35 26 −23 −23 3.54Ventrolateral PFC R 47 38 26 −23 3.51Orbitofrontal M 11 −8 49 −11 3.51

H = hemisphere, M = midline (x value between ±8), BA=Brodmann area (nearestgray matter), PFC = prefrontal cortex, x, y, z coordinates in MNI space.⁎ FWEb .05.

Table 6Brain regions, across all subjects, associated with average performance over all trials.

Correlation with average performance values

Positive H BA x y z Z value

ROI No suprathresholdclusters

Whole BrainOccipital M 17 4 −94 −4 3.29

Negative H BA x y z Z value

ROIHippocampus L – −23 −19 −15 3.50⁎

Whole BrainPosterior cingulate R 31 11 −45 23 4.75⁎

Anterior PFC L 10 −26 53 15 4.55Ventromedial PFC M 10\11 8 41 −11 4.40Precentral gyrus R 4\6 11 −30 65 4.22Dorsal frontal L 6 −19 15 61 4.09Cerebellum L – −19 −49 −27 3.85Ventral parietal R 40 45 −26 27 3.67Anterior PFC R 10 11 60 15 3.65Posterior cingulate M 31 4 −34 38 3.61Dorsal frontal R 6 15 11 61 3.56Hippocampus L – −23 −19 −15 3.50Precentral gyrus R 6 23 −23 61 3.44Precentral gyrus R 6 41 −11 49 3.43Precentral gyrus L 4 −11 −38 72 3.38Dorsolateral PFC L 9\46 −34 26 30 3.30Anterior cingulate L 32 −11 19 46 3.28

H = hemisphere, M = midline (x values between ±8), BA = Brodmann area (nearestgray matter), PFC = prefrontal cortex, x, y, z coordinates in MNI space.⁎ FWEb .05.

705S.E. Prince et al. / NeuroImage 59 (2012) 695–707

Emotional group for correct responses in early learningmay be a neuralcorrelate of an alternate approach to the task. The mPFC has beenimplicated in abstract strategizing, predictability, and the emergence ofconceptual knowledge in decision making tasks (Hampton et al., 2006;Koch et al., 2008; Kumaran et al., 2009; Tsuchida et al., 2010). Together,these findings suggest that emotional outcomes can enhance activationin regions associated with attention and strategic processing in earlylearning.

It is unclear whether the altered neural correlates had any impacton the rigidity or flexibility of learning between groups. In the future,the flexibility of learning could be assessed by using transfer tests(Reber et al., 1996) to further probe for differences betweenEmotional versus Neutral groups. Another potential impact of alteredprocessing could be on long-term retention of learning, as suggestedby Steidl et al. (2006). Future studies could assess group differencesafter a larger delay (weeks to months). Although there were nosignificant differences in behavioral performance (percent correct)between groups, the Emotional group had a statistically significantspeeding of their responses from early to late learning as well as a

Table 5Brain regions in early learning, across all subjects, associated with later change inperformance.

Correlation of early learning activation with performance change

Positive H BA x y z Z value

ROICaudate head L – −15 23 4 3.42⁎

Caudate head L – −4 4 4 2.52⁎

Whole BrainCuneus L 19 −23 −86 34 3.88Anterior cingulate R 32 26 30 30 3.81Inferior PFC R 45 30 30 15 3.65Negative H BA x y z Z valueROI No suprathreshold clustersWhole Brain No suprathreshold clusters

H = hemisphere, BA = Brodmann area (nearest gray matter), PFC = prefrontal cortex,x, y, z coordinates in MNI space.⁎ FWEb .05.

greater proportional use of complex strategies. Attenuation of regionstypically associated with PCL tasks such as the striatum and MTL maybe counteracted by the increased use of alternate learning mecha-nisms such as mPFC. Fear relevant outcomes are likely to alter boththe cognitive and neural mechanisms engaged in learning.

Beyond emotion

The differences between groups provide information about uniqueneural substrates. However, by investigating performance relatedmeasures, regardless of group assignment, we may be able to betterdelineate the role of structures involved in habit learning. Across allsubjects, the left caudate was associated with faster RTs in aparametric analysis (Fig. 5A). Although caudate activation wasgenerally related with faster RTs, ANOVAs on the extracted betavalues from this peak suggest a trend whereby subjects in theEmotional groupmay increase activation in this region over time (latevs. early) to a greater extent than the Neutral group. Additionally, leftcaudate activation magnitude during early learning correlated withdegree of performance change (late minus early) (Fig. 5C). An accountthat ties these findings together is that the caudate is optimized toextract information that is easy to attain within the structure of thetask, and that once this information is learned, the caudate is neededto a lesser extent (Delgado et al., 2005; Mattfeld and Stark, 2011). Interms of RTs, when information is readily available, RTs are likely to bereduced. In terms of performance improvement, later learning is likelyto benefit from an early foundation on which more complexcontingencies can be extracted. The caudate activation was consis-tently localized across all analyses conducted — correlations betweenactivation levels (parameter estimate beta values) from the uniquepeaks in the caudate nucleus reported in all the analyses rangedbetween 0.43 and 0.94 (all pb .05).

The striatal learning mechanism is hypothesized to provideinformation about outcomes across multiple trials, perhaps without

706 S.E. Prince et al. / NeuroImage 59 (2012) 695–707

conscious awareness (Knowlton et al., 1994, 1996). However, fMRIresults from PCL tasks (Dickerson et al., 2011; Foerde et al., 2006;Moody et al., 2004; Poldrack et al., 1999, 2001), provide substantialevidence that the striatum andMTL both contribute to these tasks andsubjects may rely on an interaction between multiple learningmechanisms.

The present study showed that, in contrast to the caudate, the MTLwas associated with slower RTs across all subjects (Fig. 5B) and theleft hippocampus was negatively associated with average perfor-mance (Fig. 5D). In patients with MTL damage, PCL studies using theweather prediction task have suggested impairments in late learning(Knowlton et al., 1994) or across the entire learning session (Hopkinset al., 2004). A role for MTL regions in response learning has beenreported in animal research (Atallah et al., 2008; Packard, 1999;Packard and McGaugh, 1996). Models of striatal function posit acompetitive, but interactive, relationship with the MTL memorysystem (Packard and Knowlton, 2002; Poldrack and Packard, 2003).However, in addition to the interactive memory model, the comple-mentary learning systems view (Atallah et al., 2004; McClelland,1998; McDonald and Hong, 2004; White and McDonald, 2002) positsthat these systemsmay compete or cooperate during the independentprocessing of information. According to White and McDonald'stheory, the hippocampus, dorsal striatum and amygdala are special-ized to represent stimulus events, responses, and reinforcers,respectively. The present findings are consistent with this interactiveview insofar as the MTL systemmay learn complex associations at thetrial level (which would require more processing time) while thestriatal system may facilitate responses to rapidly learnable material(which would be expressed on trials with faster responses). However,consistent with the directionality of effects reported by Poldrack et al.(2001) decreases in left hippocampal activation are associated withoverall task performance. In the Neutral group, the amygdala wasmore responsive on correct trials (where positive feedback was morelikely to be presented), which is consistent with a role in codingreinforcement. Finally, functional connectivity between the caudateand MTL was correlated with early learning, but only in the neutralgroup. These separate systems complement each other by extractingunique information about the task at hand, with several accountsarguing for rigid (striatal) versus flexible (MTL) applications ofknowledge gained (Bayley et al., 2005; Foerde et al., 2006; Kumaran etal., 2009). The present findings added a unique perspective to theprevailing interacting systems view by linking striatal-MTL connec-tivity to early learning, which suggests a cooperative relationshipbetween these systems across trials in traditional PCL tasks. However,the results also link faster RTs to the caudate and slower RTs to theMTL on a trial-by-trial basis, consistent with a temporary biasing ofone system over the other. Together, the results suggest that the MTLand striatal systems interact both cooperatively and competitivelyacross different timescales.

Conclusions

PCL tasks are thought to mimic procedural learning in real-worldscenarios where events can cue emotional outcomes probabilistically.However, the standard weather prediction task does not contain astrongly affective outcome so the generalization of probabilisticlearning mechanisms to emotionally arousing scenarios is unknown.A modification to the standard weather prediction task involving thepresentation of fear-relevant outcomes resulted in reduced recruit-ment of the caudate nucleus and several MTL substructures. Thisfinding was bolstered by the fact that the left caudate nucleus andbilateral amygdala were found in both the group comparison analysis(Neutral vs. Emotional) and the within-group correlation of phobiascore (Emotional group). Thus, the reductions in activation arepotentially related to individual differences in fear relevancy. Inearly learning, emotional outcomes were also associated with greater

activation in regions associated with attentional and strategicprocessing, including the mPFC. This finding suggests the use of analternate learning mechanism, especially early in the course oflearning. In addition to group differences in BOLD response, functionalconnectivity analyses revealed a reduction in striatal-MTL connectiv-ity for the Emotional group. Furthermore, the connectivity measurewas significantly correlated with performance in the Neutral but notthe Emotional group. Together, the results suggest that emotionaloutcomes in category learning alter the cognitive and neural learningmechanisms employed, with a shift in the connectivity and regionssupporting accurate task performance.

Additional analyses that took behavioral measures such as RT andfuture performance improvement into account corroborated the roleof the caudate nucleus and MTL in category learning. Interestingly,however, the RT analysis indicated thatMTL regionswere recruited onslow response trials whereas the caudate was recruited during rapidtrials, suggesting an alternative view to competitive models (Poldrackand Packard, 2003). Instead, the results suggest that these learningsystems may be complementary (Atallah et al., 2004; McClelland,1998; White and McDonald, 2002), with the caudate nucleus settingthe stage for enhanced learning by extracting readily availableinformation (Delgado et al., 2005). As this information accumulatesover learning trials, more complex relationships can be extractedbased on slower learning associated with individual trials. Together,the results of this study demonstrate how emotion modulatesprocedural learning and further delineates striatal and MTL contri-butions to specific aspects of procedural learning.

Acknowledgments

We thank Matt Fecteau and Steven Green for their assistance withdata collection and analysis. This work was supported in part by NSFgrant 0745919, and NIH grants R01 DA027802 and 2 P01 NS041328.

Appendix A. Supplementary data

Supplementary data to this article can be found online at doi:10.1016/j.neuroimage.2011.07.027.

References

Anderson, A.K., 2005. Affective influences on the attentional dynamics supportingawareness. J. Exp. Psychol.-Gen. 134, 258–281.

Aron, A.R., Shohamy, D., Clark, J., Myers, C., Gluck, M.A., Poldrack, R.A., 2004. Humanmidbrain sensitivity to cognitive feedback and uncertainty during classificationlearning. J. Neurophysiol. 92, 1144–1152.

Atallah, H.E., Frank, M.J., O'Reilly, R.C., 2004. Hippocampus, cortex, and basal ganglia:insights from computational models of complementary learning systems. Neuro-biol. Learn. Mem. 82, 253–267.

Atallah, H.E., Rudy, J.W., O'Reilly, R.C., 2008. The role of the dorsal striatum and dorsalhippocampus in probabilistic and deterministic odor discrimination tasks. Learn.Mem. 15, 294–298.

Brantley, P.J., Jones, G.N., 1993. Daily stress and stress-related disorders. Ann. Behav.Med. 15, 17–25.

Bayley, P.J., Frascino, J.C., Squire, L.R., 2005. Robust habit learning in the absence ofawareness and independent of the medial temporal lobe. Nature 436, 550–553.

Beck, A.T., Steer, R.A., Brown, G.K., 1996. Manual for the Beck Depression Inventory-II.Psychological Corporation, San Antonio, TX.

Bischoff-Grethe, A., Hazeltine, E., Bergren, L., Ivry, R.B., Grafton, S.T., 2009. The influenceof feedback valence in associative learning. NeuroImage 44, 243–251.

Corbetta, M., Akbudak, E., Conturo, T.E., Snyder, A.Z., Ollinger, J.M., Drury, H.A.,Linenweber, M.R., Petersen, S.E., Raichle, M.E., Van Essen, D.C., Shulman, G.L., 1998.A common network of functional areas for attention and eye movements. Neuron21, 761–773.

Corbetta, M., Kincade, J.M., Shulman, G.L., 2002. Neural systems for visual orienting andtheir relationships to spatial working memory. J. Cogn. Neurosci. 14, 508–523.

Daw, N.D., Niv, Y., Dayan, P., 2005. Uncertainty-based competition between prefrontaland dorsolateral striatal systems for behavioral control. Nat. Neurosci. 8,1704–1711.

Daw, N.D., O'Doherty, J.P., Dayan, P., Seymour, B., Dolan, R.J., 2006. Cortical substratesfor exploratory decisions in humans. Nature 441, 876–879.

Delgado, M.R., Miller, M.M., Inati, S., Phelps, E.A., 2005. An fMRI study of reward-relatedprobability learning. NeuroImage 24, 862–873.

707S.E. Prince et al. / NeuroImage 59 (2012) 695–707

Delgado, M.R., Li, J., Schiller, D., Phelps, E.A., 2008. The role of the striatum in aversivelearning and aversive prediction errors. Philos. Trans. R. Soc. Lond. B. Biol. Sci. 363,3787–3800.

Dickerson, K.C., Li, J., Delgado, M.R., 2011. Parallel contributions of distinct humanmemory systems during probabilistic learning. Neuroimage 55, 266–276.

Dolcos, F., McCarthy, G., 2006. Brain systems mediating cognitive interference byemotional distraction. J. Neurosci. 26, 2072–2079.

Dolcos, F., LaBar, K.S., Cabeza, R., 2004. Interaction between the amygdala and themedial temporal lobe memory system predicts better memory for emotionalevents. Neuron 42, 855–863.

Dolcos, F., LaBar, K.S., Cabeza, R., 2005. Remembering one year later: role of theamygdala and the medial temporal lobe memory system in retrieving emotionalmemories. Proc. Natl. Acad. Sci. U. S. A. 102, 2626–2631.

Foerde, K., Knowlton, B.J., Poldrack, R.A., 2006. Modulation of competing memorysystems by distraction. Proc. Natl. Acad. Sci. U. S. A. 103, 11778–11783.

Graybiel, A.M., 2005. The basal ganglia: learning new tricks and loving it. Curr. Opin.Neurobiol. 15, 638–644.

Hampton, A.N., Bossaerts, P., O'Doherty, J.P., 2006. The role of the ventromedialprefrontal cortex in abstract state-based inference during decision making inhumans. J. Neurosci. 26, 8360–8367.

Hopkins, R.O., Myers, C.E., Shohamy, D., Grossman, S., Gluck, M., 2004. Impairedprobabilistic category learning in hypoxic subjects with hippocampal damage.Neuropsychologia 42, 524–535.

Kahn, I., Yeshurun, Y., Rotshtein, P., Fried, I., Ben-Bashat, D., Hendler, T., 2002. The role ofthe amygdala in signaling prospective outcome of choice. Neuron 33, 983–994.

Kensinger, E.A., Corkin, S., 2004. Two routes to emotional memory: distinct neuralprocesses for valence and arousal. Proc. Natl. Acad. Sci. U. S. A. 101, 3310–3315.

Klorman, R., Weerts, T.C., Hastings, J.E., Melamed, B.G., Lang, P.J., 1974. PsychometricDescription of Some Specific-Fear Questionnaires. Behav. Ther. 5, 401–409.

Knowlton, B.J., Squire, L.R., Gluck, M.A., 1994. Probabilistic classification learning inamnesia. Learn. Mem. 1, 106–120.

Knowlton, B.J., Mangels, J.A., Squire, L.R., 1996. A neostriatal habit learning system inhumans. Science 273, 1399–1402.

Koch, K., Schachtzabel, C., Wagner, G., Reichenbach, J.R., Sauer, H., Schlosser, R., 2008.The neural correlates of reward-related trial-and-error learning: an fMRI studywith a probabilistic learning task. Learn. Mem. 15, 728–732.

Kumaran,D., Summerfield, J.J., Hassabis, D.,Maguire, E.A., 2009. Tracking the emergence ofconceptual knowledge during human decision making. Neuron 63, 889–901.

LaBar, K.S., Cabeza, R., 2006. Cognitive neuroscience of emotional memory. Nat. Rev.Neurosci. 7, 54–64.

LaBar, K.S., Gitelman, D.R., Parrish, T.B., Mesulam, M., 1999. Neuroanatomic overlap ofworking memory and spatial attention networks: a functional MRI comparisonwithin subjects. NeuroImage 10, 695–704.

Lagnado, D.A., Newell, B.R., Kahan, S., Shanks, D.R., 2006. Insight and strategy inmultiple-cue learning. J. Exp. Psychol.-Gen. 135, 162–183.

Lang, P.J., Bradley, M.M., Cuthbert, B.N., 1997. International Affective Picture System(IAPS): Technical Manual and Affective Ratings. NIMH Center for the Study ofEmotion and Attention, University of Florida, Gainesville, FL.

Larsen, R.J., 1984. Theory and measurement of affect intensity as an individualdifference characteristic. Dissertation Abstracts International, 45, 07B. Ph.D. thesis,University of Illinois at Urbana-Champaign, IL.

Maldjian, J.A., Laurienti, P.J., Kraft, R.A., Burdette, J.H., 2003. An automated method forneuroanatomic and cytoarchitectonic atlas-based interrogation of fMRI data sets.NeuroImage 19, 1233–1239.

Mattfeld, A.T., Stark, C.E., 2011. Striatal and medial temporal lobe functionalinteractions during visuomotor associative learning. Cereb. Cortex 21, 647–658.

McClelland, J.L., 1998. Complementary learning systems in the brain. A connectionistapproach to explicit and implicit cognition and memory. Ann. N. Y. Acad. Sci. 843,153–169.

McDonald, R.J., Hong, N.S., 2004. A dissociation of dorso-lateral striatum andamygdala function on the same stimulus–response habit task. Neuroscience124, 507–513.

McGaugh, J.L., 2004. The amygdala modulates the consolidation of memories ofemotionally arousing experiences. Annu. Rev. Neurosci. 27, 1–28.

Moody, T.D., Bookheimer, S.Y., Vanek, Z., Knowlton, B.J., 2004. An implicit learning taskactivates medial temporal lobe in patients with Parkinson's disease. Behav.Neurosci. 118, 438–442.

Morey, R.A., Dolcos, F., Petty, C.M., Cooper, D.A., Hayes, J.P., LaBar, K.S., McCarthy, G.,2009. The role of trauma-related distractors on neural systems for workingmemory and emotion processing in posttraumatic stress disorder. J. Psychiatr. Res.43, 809–817.

Packard, M.G., 1999. Glutamate infused posttraining into the hippocampus or caudate-putamen differentially strengthens place and response learning. Proc. Natl. Acad.Sci. U. S. A. 96, 12881–12886.

Packard, M.G., Knowlton, B.J., 2002. Learning and memory functions of the BasalGanglia. Annu. Rev. Neurosci. 25, 563–593.

Packard, M.G., McGaugh, J.L., 1996. Inactivation of hippocampus or caudate nucleuswith lidocaine differentially affects expression of place and response learning.Neurobiol. Learn. Mem. 65, 65–72.

Paton, J.J., Belova, M.A., Morrison, S.E., Salzman, C.D., 2006. The primate amygdalarepresents the positive and negative value of visual stimuli during learning. Nature439, 865–870.

Poldrack, R.A., Packard, M.G., 2003. Competition among multiple memory systems:converging evidence from animal and human brain studies. Neuropsychologia 41,245–251.

Poldrack, R.A., Prabhakaran, V., Seger, C.A., Gabrieli, J.D., 1999. Striatal activation duringacquisition of a cognitive skill. Neuropsychology 13, 564–574.

Poldrack, R.A., Clark, J., Pare-Blagoev, E.J., Shohamy, D., Creso Moyano, J., Myers, C., Gluck,M.A., 2001. Interactive memory systems in the human brain. Nature 414, 546–550.

Reber, P.J., Knowlton, B.J., Squire, L.R., 1996. Dissociable properties of memory systems:differences in the flexibility of declarative and nondeclarative knowledge. Behav.Neurosci. 110, 861–871.

Schwabe, L., Wolf, O.T., Oitzl, M.S., 2010. Memory formation under stress: quantity andquality. Neurosci. Biobehav. Rev. 34, 584–591.

Shohamy, D., Myers, C.E., Grossman, S., Sage, J., Gluck, M.A., Poldrack, R.A., 2004. Cortico –striatal contributions to feedback-based learning: convergingdata fromneuroimagingand neuropsychology. Brain 127, 851–859.

Shohamy, D., Myers, C.E., Kalanithi, J., Gluck, M.A., 2008. Basal ganglia and dopaminecontributions to probabilistic category learning. Neurosci. Biobehav. Rev. 32, 219–236.

Spiegler, B.J., Mishkin, M., 1981. Evidence for the sequential participation of inferiortemporal cortex and amygdala in the acquisition of stimulus-reward associations.Behav. Brain Res. 3, 303–317.

Steidl, S., Mohi-uddin, S., Anderson, A.K., 2006. Effects of emotional arousal on multiplememory systems: evidence from declarative and procedural learning. Learn. Mem.13, 650–658.

Thomas, L.A., LaBar, K.S., 2008. Fear relevancy, strategy use, and probabilistic learning ofcue-outcome associations. Learn. Mem. 15, 777–784.

Tricomi, E.M., Delgado, M.R., Fiez, J.A., 2004. Modulation of caudate activity by actioncontingency. Neuron 41, 281–292.

Tricomi, E., Delgado, M.R., McCandliss, B.D., McClelland, J.L., Fiez, J.A., 2006. Performancefeedback drives caudate activation in a phonological learning task. J. Cogn.Neurosci. 18, 1029–1043.

Tsuchida, A., Doll, B.B., Fellows, L.K., 2010. Beyond reversal: a critical role for humanorbitofrontal cortex in flexible learning from probabilistic feedback. J. Neurosci. 30,16868–16875.

Tzourio-Mazoyer, N., Landeau, B., Papathanassiou, D., Crivello, F., Etard, O., Delcroix, N.,Mazoyer, B., Joliot, M., 2002. Automated anatomical labeling of activations in SPMusing a macroscopic anatomical parcellation of the MNI MRI single-subject brain.NeuroImage 15, 273–289.

Wang, L., LaBar, K.S., Smoski, M., Rosenthal, M.Z., Dolcos, F., Lynch, T.R., Krishnan, R.R.,McCarthy, G., 2008. Prefrontal mechanisms for executive control over emotionaldistraction are altered in major depression. Psychiatry Res. 163, 143–155.

Watson, D., Clark, L.A., Tellegen, A., 1988. Development and validation of brief measures ofpositive and negative affect: the PANAS scales. J. Pers. Soc. Psychol. 54, 1063–1070.

White, N.M., McDonald, R.J., 2002. Multiple parallel memory systems in the brain of therat. Neurobiol. Learn. Mem. 77, 125–184.

Yerkes, R.M., Dodson, J.D., 1908. The relation of strength of stimulus to rapidity of habit-formation. J. Comp. Neurol. Psychol. 18, 459–482.