practice-related changes in neural activation patterns

r Human Brain Mapping 000:000–000 (2012) r

Practice-Related Changes in Neural ActivationPatterns Investigated via Wavelet-Based

Clustering Analysis

Jinae Lee,1 Cheolwoo Park,1 Kara A. Dyckman,2,3,4 Nicole A. Lazar,1

Benjamin P. Austin,5 Qingyang Li,6 and Jennifer E. McDowell2,3,4*

1Department of Statistics, University of Georgia (UGA), Athens, Georgia2Department of Psychology, University of Georgia (UGA), Athens, Georgia

3Department of Neuroscience, University of Georgia (UGA), Athens, Georgia4The Bio-Imaging Research Center, University of Georgia (UGA), Athens, Georgia

5UW Cardiovascular Research Center, University of Wisconsin School of Medicine and Public Healthand theDepartment of Veterans Affairs (VA) Geriatric Research, Education and Clinical Center

(GRECC), Madison, WI, USA6Center for the Developing Brain, Child Mind Institute, 445 Park Ave, New York, NY 10022, USA

r r

Abstract: Objectives: To evaluate brain activation using functional magnetic resonance imaging (fMRI)and specifically, activation changes across time associated with practice-related cognitive control duringeye movement tasks. Experimental design: Participants were engaged in antisaccade performance (gener-ating a glance away from a cue) while fMR images were acquired during two separate test sessions: (1)at pre-test before any exposure to the task and (2) at post-test, after 1 week of daily practice on antisac-cades, prosaccades (glancing toward a target), or fixation (maintaining gaze on a target). Principal obser-vations: The three practice groups were compared across the two test sessions, and analyses wereconducted via the application of a model-free clustering technique based on wavelet analysis. This se-ries of procedures was developed to avoid analysis problems inherent in fMRI data and was composedof several steps: detrending, data aggregation, wavelet transform and thresholding, no trend test, prin-cipal component analysis (PCA), and K-means clustering. The main clustering algorithm was built inthe wavelet domain to account for temporal correlation. We applied a no trend test based on waveletsto significantly reduce the high dimension of the data. We clustered the thresholded wavelet coeffi-cients of the remaining voxels using PCA K-means clustering. Conclusion: Over the series of analyses,we found that the antisaccade practice group was the only group to show decreased activation frompre-test to post-test in saccadic circuitry, particularly evident in supplementary eye field, frontal eyefields, superior parietal lobe, and cuneus. Hum Brain Mapp 00:000–000, 2012. VC 2012 Wiley Periodicals, Inc.

Keywords: antisaccades; bootstrap; clustering; fMRI; plasticity; principal component analysis;saccades; wavelets

r r

Additional Supporting Information may be found in the onlineversion of this article.

Contract grant sponsor: NIMH; Contract grant number: MH01852.

*Correspondence to: Jennifer E. McDowell, Department of Psy-chology, UGA Psychology Building, Athens, GA 30602.E-mail: [email protected]

Received for publication 20 September 2011; Revised 2 February2012; Accepted 3 February 2012

DOI: 10.1002/hbm.22066Published online in Wiley Online Library (wileyonlinelibrary.com).

VC 2012 Wiley Periodicals, Inc.

INTRODUCTION

Cognitive control mediates the process through whichtasks transition from new or unfamiliar to learned orskilled. With practice, a task becomes less effortful, result-ing in improved performance and a reduced need for cog-nitive control [e.g., Chein and Schneider, 2005; Jansmaet al., 2001; Schneider and Chein, 2003; Schneider and Shif-frin, 1977]. Modifications in the neural circuitry supportingtask performance also occur following practice [e.g., seeChein and Schneider, 2005; Kelly and Garavan, 2005 forreviews]. These changes in brain activation likely includechanges in the specific neural circuitry supporting theresponse of interest as well as general changes reflecting adecreased need for cognitive control.

In this study, changes in brain activation across timewere examined before and after daily exposure to a set ofeye movement tasks. Among eye movements, saccadesrapidly redirect gaze to a location of interest. ‘‘Prosac-cades’’ are more reflexive, as a person simply redirectsgaze to a newly appearing visual stimulus. ‘‘Antisaccades’’are more complex and volitional, as a person redirectsgaze to the mirror image location (same amplitude, oppo-site side) of a newly appearing peripheral visual cue. Assuch, a correct antisaccade response requires the inhibitionof a glance toward the cue and generation of a voluntarysaccade to an unmarked location in the opposite visualfield.

Saccadic performance is supported by a network of sub-cortical and cortical regions, as identified via neuroimag-ing and other techniques [Camchong et al., 2008; Dyckmanet al., 2007; Ford et al., 2005; Keedy et al., 2006; McDowellet al., 2008; Muri et al., 1998; O’Driscoll et al., 1995; Paus,1996; Raemaekers et al., 2002, Sweeney et al., 2007].Although the basic circuitry is the same for prosaccadesand antisaccades, the increased complexity of antisaccadesis supported by increased activation of existing circuitryand/or the recruitment of additional neural regions intothe circuitry. As such, the saccade circuitry provides a spe-cific, well-studied system for understanding changes inbrain activation associated with cognitive control andpractice.

This study used functional magnetic resonance imaging(fMRI) to evaluate the changes across time in neural sys-tem activation after participants practiced either specific ornonspecific eye movement tasks. Participants wereengaged in antisaccade performance during fMR image ac-quisition at two test sessions: (1) at pre-test before any ex-posure to the task and (2) at post-test after 1 week of dailypractice on eye movement tasks. Each participant prac-ticed a single type of eye movement only: (a) antisaccades,(b) prosaccades, or (c) fixation (maintaining gaze on a tar-get). It was hypothesized that participants who had task-consistent practice (i.e., the antisaccade practice group)would show changes in neural circuitry (as measured bythe blood oxygenation level-dependent (BOLD) signal) notobserved in the two task-inconsistent practice groups (the

prosaccade and fixation groups) who practiced eye move-ment control tasks generally, but not antisaccades specifi-cally. It was further hypothesized that (i) for saccade tasks,saccade-related circuitry would show decreased BOLD sig-nal over time as a result of the circuitry becoming moreefficient and (ii) in the antisaccade practice group only,prefrontal cortex (PFC), frontal eye fields (FEFs), and stria-tum would show decreased BOLD signal due to decreasedneed for higher level cognitive control [Camchong et al.,2008, Munoz and Everling, 2004, Raemaekers et al., 2002].

The comparison of practice groups across test sessionswas conducted using a combination of statistical techni-ques that were selected specifically to circumvent com-monly acknowledged methodological problems that areassociated with the structure of the data generated viafMRI and/or the model-based general linear model (GLM)analyses as they are typically applied. Features of fMRIdata that present difficulties for analysis include the accu-rate modeling of the BOLD response, massive-sized datathat are ill-balanced, and the influence of spatial and tem-poral correlations in the data. These statistical challengeshave been addressed in the past using a variety of meth-ods that fall into two broad categories: model-based andmodel-free analyses.

A typical model-based analysis is to build activationmaps by testing each voxel separately using a statisticalmodel (e.g., a general linear model). Under the specifiedmodel, a P-value is calculated and used to determinewhether the corresponding voxel is activated [see, e.g.,Worsley and Friston, 1995 for early work in this direction].The flexibility of the linear model allows for easy incorpo-ration of temporal correlation via specification of anappropriate error structure. However, this approach(including the standard GLM-type analysis) may not beoptimal. It is prone to the multiple testing problem as in-ference is performed at the voxel level [Forman et al.,1995; Genovese et al., 2002; Worsley, 2003]. The number oftrue signals (active voxels) is typically small comparedwith the number tested, which exacerbates the detectionissue. Furthermore, model misspecification, especially ofthe dependence structure of the fMRI time series, is apotentially serious issue. If the error structure is incorrect,estimated standard errors are also incorrect and inferencewill be severely impacted.

As an alternative, wavelet approaches have been appliedto build brain activation maps. S� endur et al. [2005]pointed out that the wavelet transform increases the sig-nal-to-noise ratio (SNR) and substantially decreases the se-rial correlation in the data. Van De Ville and Unser [2008]constructed a statistical parametric map by applying wave-lets to both time and space domains and adapted the falsediscovery rate [FDR; Benjamini and Hochberg, 1995]method to address the multiple testing problem. By con-trast, a standard model-free approach such as clustering[e.g., Balslev et al., 2002; Fadili et al., 2000; Goutte et al,1999; Windischberger et al., 2003] groups together voxelswith similar activation profiles without assuming any

r Lee et al. r

r 2 r

particular model or functional form in advance. Therehave been fewer attempts in model-free clustering, how-ever, that account for the temporal correlation inherent inthese data. One example is Ye et al. [2009], who took ageostatistical approach and clustered fMRI time seriesby comparing their correlation structures. Finally, compro-mises between a massively univariate model-basedanalysis and model-free clustering are possible. Helleret al. [2006], for instance, proposed a cluster-based thresh-olding approach, which creates groups of contiguous vox-els and tests at the cluster, rather than at the individual,level.

The large number of task-irrelevant voxels poses chal-lenges for both categories of analysis. In the model-basedapproaches, the multiplicity problem is greatly exacer-bated by the presence of these voxels. Hence, an increas-ingly common first step is to estimate the number of ‘‘truenulls’’ (see Benjamini and Hochberg, 2000; Turkheimeret al., 2001 for early work in this area; much research onthis problem has been carried out in both neuroimagingand genetics contexts) and test only those voxels that areretained after that initial screening. For clustering meth-ods, so-called ‘‘ill-balanced data’’, wherein one class ismuch more prevalent than the others, are problematic,because many techniques will tend in this situation tocluster all cases into the dominant class. It is, therefore, acommon practice before clustering to reduce the data sizeand to solve the ill-balanced data problem. For instance,Fadili et al. [2000] used a priori anatomical informationand reduced the number of voxels by testing the temporalautocorrelation in each voxel’s time series and consideringthe spatial distribution of the activated regions. Balslevet al. [2002] used an F test to discard voxels with no over-all effect.

Our approach to treating these issues was based on theapplication of clustering routines used in correspondencewith wavelet analysis. Transforming to the wavelet do-main accounts for the temporal correlation in the voxeltime series. We also used a wavelet-based test to detecttask-irrelevant voxels, thereby mitigating the ill-balancednature of the data. Our analysis was composed of severalnotable steps, including detrending, data aggregation,decorrelation, no trend test [Park et al., 2011], principalcomponent analysis [PCA, Jolliffe, 2002], and K-meansclustering [Hartigan and Wong, 1979]. We performed ananalysis of variance (ANOVA) test to evaluate the maineffects of group and time as well as their interaction. Fur-thermore, using regions of interest (ROIs) we identifiedregions that showed statistical differences between thegroups and test sessions based on the bootstrap method.

Importantly, the strategies presented above eliminated theneed to create an a priori model of the ‘‘expected’’ activationacross time, reduced the massive, ill-balanced data set, andaccounted for temporal autocorrelations. As such, the differ-ences between practice groups across time were tested in arigorous manner. These analyses enhanced our ability toevaluate practice-related brain activity changes across time.

METHODS

Participants

Data were collected from 37 right-handed undergradu-ate women (mean age ¼ 19.5 years and standard deviation(SD) ¼ 1.8) recruited from the Undergraduate ResearchPool at the University of Georgia (UGA) or via flyersposted on campus. All participants were free (by self-report) of psychiatric illness, history of head injury, anddrug and alcohol abuse. Participants provided writteninformed consent before the study, which was approvedby the UGA Institutional Review Board.

Procedure

fMRI data during antisaccade task performance wereacquired at two test sessions: (1) pre-test (before any taskexposure) and (2) post-test. Between pre-test and post-testfMRI sessions, participants practiced one type of randomlyassigned eye movement daily for 1 week: (1) antisaccades(glances away from a cue, N ¼ 12), (2) prosaccades (glan-ces toward a cue, N ¼ 14), or (3) fixation (maintains gazeat a central target, N ¼ 11). All participants completedboth test sessions.

fMRI test sessions

fMRI data were collected using a 1.5 T GE Signal HorizonLX Scanner (GE Medical Systems, Waukesha, WI). Taskinstructions were given before entering the scan room. Oncein the room, a participant was positioned on a gurney withher head stabilized with foam padding and a restraint strapacross her forehead. A mirror placed over the participant’shead allowed her to view the stimuli on a rear-projectionscreen placed near her feet. Eye movements were recordedusing an fMRI compatible iView X System (SensoMotoricInstrument; 60-Hz sampling rate) eye tracking system.

Two localizer images were taken at the beginning of thesession to ensure optimal brain coverage for each partici-pant. A high-resolution image was then obtained (spoiledgradient recalled (SPGR) protocol: sagittal, 2 number ofexcitations (NEX) 0.9375 � 0.9375 � 1.5 mm3, 124 slices,echo time (TE) ¼ 2.8 ms, repetition time (TR) ¼ 10.8 ms,flip angle ¼ 20�, and scan time ¼ 5 min 41 s). After thisstructural image was acquired, functional imaging wasconducted (spiral scan with two interleaves, 24 continuousaxial slices, 3.75 � 3.75 � 4 mm3, TE ¼ 40 ms, TR ¼ 1,912ms, flip angle ¼ 77�, and scan time ¼ 5 min 1 s).

The functional run consisted of 13 blocks (22.4 s each)alternating between fixation and antisaccade trials. Stimuliconsisted of 1� circles filled with different colors. A fixa-tion block was indicated by a purple stimulus at the cen-tral location, and participants were instructed to fixate onthe target for the duration of the block. An antisaccadeblock consisted of seven individual trials, each of whichstarted with a blue stimulus at the central location (2,000ms). The stimulus was extinguished and 200 ms later a

r Wavelet-Based Cluster Analysis During Saccades r

r 3 r

cue was presented 5� to the left or the right (half in eachvisual field) on the horizontal axis (1,000 ms). Participantswere instructed to look at the blue stimulus only when itwas in the middle of the screen, and when it was pre-sented in the periphery, to look to the mirror image ofthat stimulus (opposite direction, same amplitude) asquickly and accurately as possible.

Behavioral practice sessions

After the fMRI pre-test, each participant practiced herassigned task (antisaccade, prosaccade, or fixation) in thelaboratory each of 4 days (weekends were excluded). Prac-tice sessions were conducted using a hand-held devicewith an LCD screen known as ‘‘Fix-Train’’ that wasdesigned for the purpose of providing practice with eyemovement tasks [see Fischer et al., 2000]. For each task,the goal of the participant was to determine the orientationof a ‘‘T’’ symbol and to press the corresponding arrowkey. The bottom leg of the ‘‘T’’ symbol could be facing up,down, left, or right. This manual task served as a proxyfor measuring eye movements because the location andduration of the ‘‘T’’ presentation were designed so that theparticipant must execute the correct eye movement to seethe stimulus and discern its orientation.

For fixation training, the symbol remained in the centerof the screen and changed direction two to five timesbefore disappearing. Subjects pressed the arrow key corre-sponding to the final orientation of the symbol before itdisappeared from the screen. For prosaccade training, thetask was identical except that the symbol jumped to oneside after changing orientation in the middle, and partici-pants pressed the arrow key corresponding to the orienta-tion of the symbol once it jumped to the periphery. Forantisaccade training, a star appeared in the center of thescreen to provide a fixation target. Then, the star jumpedto one side of the screen and the ‘‘T’’ symbol appeared 70ms later on the opposite side of the screen. Participantswere instructed to look to the opposite side of the star andpress the arrow key that corresponded to the orientationof the symbol. If participants made a saccade toward thestar before looking away from it, they would miss thesymbol as it was only visible for 100 ms. Prosaccade andantisaccade training consisted of an equal number of trialsto the left and right. Each participant completed 200 trialsa day of her specific task (fixation, prosaccade, or antisac-cade) and her data were recorded on the device. The num-ber of correct trials and the mean reaction time for eachpractice session were recorded by the Fix-Train and down-loaded to a computer for further analysis.

ANALYSES

Behavioral Data From Practice and Test Sessions

We analyzed all behavioral data using repeated-meas-ures ANOVA with between-subjects factor practice group

(antisaccade, prosaccade, and fixation). The practice datahad the within-subjects factor of days (1–4), whereas thetest data were analyzed with the within-subjects factor oftest session (pre- and post-).

Practice data consisted of percent correct (number ofcorrect identifications of the direction of the symbol/num-ber of total trials) and reaction time (how long it took theparticipant to press the button after the disappearance ofthe symbol). Although these measures were of hand move-ments rather than eye movements, they were an adequateproxy because the participant had to have her eyes in thecorrect location to see the symbol orientation (see Fischeret al., 2000 for a similar report).

Test session data consisted of eye movements recordedwhile the participants were engaged in antisaccade tasks.Data were analyzed using MATLAB (The Mathworks,Natick, MA) as described in previous manuscripts [Dyck-man et al., 2007]. Briefly, trials with blinks and trials withno saccades were eliminated. Saccades were scored fordirection and reaction time. ‘‘Percent correct’’ antisaccadeperformance was calculated by dividing the number of tri-als made in the correct direction by the total number ofscoreable trials and multiplying by 100.

fMRI Data

The data analysis consisted of standard fMRI prepro-cessing steps before the novel application of our combina-tion of statistical techniques: detrending and dataaggregation, decorrelation, no trend test, and clusteringthe processed voxel BOLD time series using PCA and K-means. The resulting clustered maps between the two scantest points (pre-test and post-test) and across the threepractice groups (antisaccade, prosaccade, and fixation)were then compared.

Preprocessing

We preprocessed the data using standard methods inAFNI [Cox, 1996] including motion correction, slice timingcorrection, head motion correction, spatial and temporalfiltering, and normalization into Talairach space [Talairachand Tournoux, 1988].

Detrending and data aggregation

All 37 participants had data from 38 slices (40 � 48 vox-els in size) covering the full brain. Voxels outside of thebrain were masked by applying a predetermined thresholdvalue. Visual inspection showed that activation occuredpredominantly in the 11th–34th slices (Talairach z ¼ �20to þ72); therefore, we selected this subset of the full vol-ume for the remaining analyses. We excluded the first fivetime points from the analysis to allow for stabilization ofthe magnetization. Thus, the final data set contained 37participants, each with data from 24 slices and 70 timepoints. Linear and quadratic trends in the voxel time series

r Lee et al. r

r 4 r

were removed from all subject data before any furtheranalysis.

Current software is not capable of clustering massivedata sets such as the one we obtained from this study [37subjects; 40 � 48 � 24 voxels; Mumford and Poldrack,2007, Wanga et al., 2007]. To make clustering more man-ageable, it was therefore necessary to first reduce the sizeof the data. This is not without precedent in the literature.For example, Balslev et al. [2002] averaged together theresults of 18 subjects before clustering. To validate theaggregation of the 37 subjects, we evaluated two centermeasures (mean and median) and three variation meas-ures (standard deviation, range, and interquartile range) ateach voxel for each subject. The distributions of thesemeasures across all subjects were similar. Therefore, wecombined the subjects for each group at each voxel andeach time point separately using a representative value, inthis case the median, and then clustered the combinedbrain. In what follows, we present results from themedian; note, however, that the clustering results (as seen,for example, in Figs. 3 and 4) are similar with the othersummary measures. An advantage of the median was thatit reduced the effect of heavy outliers found in some vox-els. Another advantage of this aggregating approach wasthat it increased the SNR, hence it produced sharper clus-tering brain maps. After aggregation we have a singleimage per group (40 � 48 � 24 voxels in size) for each ofthe 70 time points.

Decorrelation of fMRI time series

The existence of temporal correlation is inherent in fMRItime series and has been evaluated using different types oftime series models ranging from autoregressive of loworder [see, for example, Gautama and Van Hulle, 2005;Luo and Nichols, 2003; Woolrich et al., 2001] to long-rangedependence [see, for example, Bullmore et al., 2003;Park et al., 2010]. This dependence structure makes clus-tering analysis more challenging and should be accountedfor because it is nontrivial to distinguish deterministicpatterns from dependence artifacts in fMRI time series. Atypical approach to address this issue in clustering analy-sis is to assume a temporal model for characterizing thedependence structure in fMRI time series [e.g. Worsley,2003]. Instead of modeling the temporal correlation explic-itly, we used the (model-free) wavelet transformationto take advantage of its decorrelation and denoisingproperties.

Let w be a wavelet function with vanishing moment L[Daubechies, 1992], that is,

Zu1wðuÞdu ¼ 0; l ¼ 0; 1; : : : ; L� 1;

and wj;kðuÞ ¼ 2jwð2ju� kÞ; j; k 2 Z, be a shifted and scaledversion of w. If we denote a fMRI time series by Yt at time

t, then the wavelet transformation at scale j and time loca-tion k is defined as

bj;k ¼Z

YtðuÞwj;kðuÞdu;

and the bj,ks are called the wavelet coefficients of Yt.One critically important advantage of using the wavelet

transformation in fMRI analysis is that it removes the tem-poral correlation so that the transformed data have lessserial correlation in the wavelet domain for large L [forexample, see Stoev et al., 2005]. If we then cluster the vox-els in the wavelet domain rather than in the time domain,the time correlation effect inherent in fMRI data isreduced. Veitch and Abry [1999] pointed out, however,that the theoretical improvement with large L is balancedby an increase in the number of wavelet coefficients influ-enced by the boundary effect, because of the finite lengthof the time series. They used L ¼ 3 for internet traffic datato reduce strong serial correlation and also remove possi-ble linear and quadratic trends in the data. In our analysis,as we removed those trends in advance, we used a Daube-chies wavelet function [Daubechies, 1992] with L ¼ 2,which would increase the number of available waveletcoefficients, and applied the wavelet transformation toeach (median-valued) voxel. We also compared clusteringmaps to check significant changes from L ¼ 2 to 4 andfound them similar to each other.

In addition, we applied a wavelet hard thresholdingtechnique to obtain a sparse representation of a signal bysetting the wavelet coefficients less than the threshold sug-gested by Donoho and Johnstone [1994] to zero. Thisdenoising process amplifies important signals and removesnoise in the observed data. We used the Wavelab package,available at http://stat.stanford.edu/�wavelab/, for wave-let analysis.

No trend test

For ill-balanced data in which there is a severe lack of bal-ance between the different classes of voxels, clustering tech-niques generally do not perform well as the identifiedclusters tend to be dominated by the one large class. Becausethe minority of voxels is task related in fMRI studies, cluster-ing results can be dominated by the majority of nontask-related voxels. Goutte et al. [1999] applied a loose statisticaltest such as F-test as a reduction tool. Fadili et al. [2000]removed white noise voxels and further reduced the num-ber of voxels by considering only the gray matter.

We performed a no trend test based on wavelets, whichis more rigorous than a simple F-test as it accounts fortemporal correlation. Park et al. [2011] proposed the adapt-ive pivot test based on wavelets to detect any kind oftrend in a time series. The key idea was to convert theproblem of testing a trend of a function into high-dimen-sional normal mean inference in the wavelet domain. Weconducted this test at each voxel to identify voxels with no


r 5 r

trend (i.e., voxels that were nontask related). Details ofthis procedure are provided in the Appendix. Becauseeach voxel was tested using this approach, adjustment formultiple testing was necessary and we applied the FDRwith q ¼ 0.15. This was a relatively generous threshold, tohelp ensure that truly interesting voxels were not removedfrom the rest of the analysis.

For each session and each group, we started with 46,080(40 � 48 � 24) voxels. After the masking procedure,14,743–15,319 voxels were categorized as brain. Applica-tion of the no trend test with FDR for multiplicity retained1,852–3,329 voxels that were categorized as antisaccadetask related, categorizing the other voxels as nontaskrelated.

Clustering methods

We clustered the wavelet coefficients of the remainingvoxels after applying the no trend test using both K-meansand PCA K-means clustering. The K-means method startsfrom specification of the number of clusters, K, in advanceand assigns each voxel to the cluster with the nearest cent-roid as follows [Johnson and Wichern, 2002]:

1. Partition each wavelet-transformed voxel arbitrarilyinto K initial clusters and calculate the cluster centroids.

2. Reassign each voxel into the nearest cluster based onits distance from the K centroids.

3. Update the cluster centroid and reassign each voxelinto the nearest cluster by recalculating the distanceof each voxel from the updated centroids.

4. Repeat Step 3 until there is no more reassignment.

There are various choices for a distance metric including

the Euclidean, Minkowski, Mahalanobis, and correlation.

Because the data are suitably standardized by detrending

and are approximately independent by the wavelet trans-formation, we use the Euclidean metric for computing the

distance between points and cluster centroids, which is a

typical choice for K-means clustering analysis.PCA [Jolliffe, 2002] is widely used as a data dimension

reduction technique. The basic idea is that if linear rela-tionships exist between variables, analyses will be effectiveif only a smaller number M of uncorrelated variables,called principal components (PC), are used. The clusteringprocedure for PCA K-means is as follows:

1. For wavelet-transformed voxels, obtain the completeset of PCs via PCA.

2. Draw the scree plot in which the horizontal axis containsthe PCs sorted by decreasing fraction of total varianceexplained and identify theMmost important PC scores.

3. Apply K-means clustering to the M selected PC scoresof the wavelet-transformed voxels.

On the basis of the scree plots for pre-test and post-test forthe three groups (Fig. 1), we selected 15 PCs. The 15 compo-nents explained at least 80% of the total variance for eachtest session and each practice group. The results for the threegroups were similar to each other so that they were indistin-guishable at each time point (pre-test and post-test).

In both methods, the choice of K is an important issue. It canbe determined using criteria such as silhouette values [Ye,2003], a likelihood approach with cross validation [Balslevet al., 2002], a hierarchical approach [Stanberry et al., 2003], or

Figure 1.

Scree plots of PCA for each test session [(a) pre-test and (b)

post-test] and each practice group. Each curve represents one

of the practice groups (antisaccade, prosaccade, or fixation), and

the heights of the vertical bars on the bottom represent the

contributions of each PC for that group. Note that there are lit-

tle differences across the groups so that lines are indistinguish-

able. On the basis of these plots, we selected 15 PCs (dashed

vertical line) because they explained at least 80% of the total

variance.

r Lee et al. r

r 6 r

unsupervised fuzzy c-means analysis [Fadili et al., 2000]. Inour analysis, we considered different numbers of clusters, K¼3, 4, 5, and 7, and chose K ¼ 3 because the larger numbers ofclusters only broke the noise cluster into multiple clusters anddid not provide additional information. Another issue in K-means clustering is that initial partitions can affect the finalclusters. We therefore ran the clustering algorithm multipletimes for the same number of clusters and obtained similarclustering results to those presented in Figures 3 and 4.

In what follows, we summarize the data analysis strategy:

1. Preprocessing.2. Masking, linear and quadratic detrending, and data

aggregation using median.3. Wavelet transformation of the aggregated voxels for

decorrelation and wavelet thresholding for denoising.4. No trend test and multiple testing adjustments using

FDR to remove nontask-related voxels.5. PCA analysis to find important PC scores of the

wavelet coefficients obtained from Step 3 in theremaining voxels after Step 4.

6. K-means clustering analysis of the PC scores selectedfrom Step 5.

RESULTS

Behavioral Data From Practice and Test Sessions

Practice sessions

Practice data obtained from the Fix-Train consisted of(a) the mean percentage correct and (b) reaction time val-

ues for each group across the four practice days for thespecific practice task assigned (Fig. 2).

For percentage correct, we found a main effect of group[F(2,30) ¼ 8.7, P ¼ 0.001], a main effect of time [F(3,88) ¼9.84, P < 0.0001], and a significant group � time interaction[F(6,88) ¼ 3.8, P ¼ 0.002]. The antisaccade practice groupshowed more correct antisaccades across practice days,whereas the prosaccade and fixation groups remained at thesame (high) level of performance across time.

For reaction time, there were main effects of group[F(2,31) ¼ 4.44, P ¼ 0.02] and time [F(3,91) ¼ 79.32, P <0.001], but no significant group � time interaction [F(6,91)¼ 0.39, P ¼ 0.89]. All three groups decreased their reactiontimes in their assigned task across practice days.

Test sessions

Eye tracking data during fMR imaging (pre-test and post-test days) were available from 86% of the participants. Datafrom five participants (two in the antisaccade and fixationpractice groups and one in the prosaccade practice group)were not recorded because of technical limitations. On anti-saccade measures, there were no significant main effects ofgroup or test session and no group � test session interaction(P-values > 0.18). The mean percent correct for all partici-pants over both days was 82% (SD¼ 11.3).

fMRI Data From Clustering Analysis

The clustering results by the two methods (K-means andPCA K-means) were similar, and the results from PCA K-

Figure 2.

Behavioral data (percent correct and reaction time) from each groups’ practice sessions. Black

lines represent the antisaccade practice group, dark gray lines represent the prosaccade practice

group, and light gray lines represent the fixation practice group.


r 7 r

Figure 3.

Three clusters (cluster 1 is shown in black, cluster 2 is shown in dark gray, and cluster 3 is

shown in light gray) for pre-test and post-test for antisaccade (a, b), prosaccade (c, d), and fixa-

tion groups (e, f). Black dashed lines represent the stimulus timing with values on the x-axis

being the number of whole-brain collections (TRs) across time.

r Lee et al. r

r 8 r

means clustering with K ¼ 3 are presented for purposes ofdemonstration. For each group and each test session (pre-and post-), the brain was partitioned into three clusters(Fig. 3) and compared with the behaviorally based refer-ence wave. For each plot at least one cluster displays thisshape, suggesting the presence of antisaccade task-relatedvoxels. Note that the amplitude of this primary clusterlooks attenuated at post-test compared with pre-test, espe-cially for the antisaccade group.

To evaluate where in the brain the task-related clusterswere located and to determine whether amplitude changeoccurred between pre-test and post-test in regions associ-ated with saccadic performance, the voxels making up thetwo clusters showing greatest task-related activation wereplotted in the brain map (Fig. 4). These data for the anti-saccade practice group clearly identified the well-knowncircuitry supporting saccadic performance that includesFEFs, supplementary eye fields (SEFs), and posterior parie-tal cortex (PPC).

Next, we considered in more detail the four superior sli-ces (Talairach coordinates z ¼ 52, 56, 60, and 64) thatclearly contained the well-known saccadic circuitry tomore specifically investigate changes from pre-test to post-test within and between practice groups. The locations ofvoxels at each time point were color coded to identifythose showing antisaccade-related activity at pre-test only,at both pre-test and post-test, or at post-test only (Fig. 5).The numbers of voxels in each category in the four sliceswere tabulated (Table I). The number of voxels activatedat post-test was smaller for the antisaccade practice group

compared with the other groups. This is consistent withthe observation in Figure 3b that the box-car shaped timeseries is attenuated for the antisaccade group at post-test.

To evaluate the visual observations from the clusteranalysis, we performed an ANOVA using the proportionof activated voxels based on Table I as the response vari-able, test session (pre-test and post-test), and practice con-dition (antisaccade, prosaccade, and fixation) as factors,location (after dividing the brain into five regions basedon Fig. 5) as a block variable, and slice (z ¼ 52, 56, 60, and64) as replications. There was a main effect of day [F(1,98)¼ 5.06, P ¼ 0.027] that showed more activated voxels atpost-test, but practice condition [F(2,98) ¼ 1.29, P ¼ 0.281]and location [F(4,98) ¼ 1.68, P ¼ 0.162] were not statisti-cally significant. There was, however, a significant interac-tion between day and practice condition [F(2,98) ¼ 50.44,P < 0.0001]. Post hoc testing showed that the antisaccadegroup had fewer activated voxels at post-test than at pre-test, whereas the other groups had more. Only the antisac-cade group had a decreased proportion of activation atpost-test (Fig. 6).

ROI Analysis

The previous ANOVA revealed an attenuation at post-test for the antisaccade practice group in the saccade cir-cuitry constrained to superior cortex (slices z ¼ 52, 56, 60,and 64). This analysis, however, did not provide informa-tion about how specific neural regions within whole-brain

Figure 4.

Brain maps of pre-test and post-test for the antisaccade practice group with two primary clus-

ters. Voxels shown in green represent cluster 1 and those shown in blue represent cluster 2.

The slices were arranged starting from inferior (Talairach z ¼ 20, upper left) and moving supe-

rior (z ¼ 64, lower right).


r 9 r

saccadic circuitry may have been differentially affected.Therefore, we conducted an additional analysis using 11neural ROIs previously identified [Dyckman et al., 2007]including SEF, bilateral FEF, left and right PFC, right infe-rior frontal cortex (IFC), bilateral superior parietal lobe

(SPL), bilateral cuneus, bilateral inferior parietal lobe (IPL),bilateral middle occipital gyrus (MOG), bilateral striatum,and bilateral thalamus.

To identify the specific ROIs that showed attenuation atpost-test for the antisaccade group, we performed a

Figure 5.

Overlay plots for (a) antisaccade, (b) prosaccade, and (c) fixation groups for four slices inferior

to superior (Talairach z ¼ 52, 56, 60, and 64). Voxels in red are activated at pre-test but not at

post-test, those shown in orange are activated at both pre-test and post-test, and those shown

in yellow are activated only at post-test. Activation is observed in typical saccadic circuitry at

this level of the brain, which includes bilateral FEFs, SEF at the midline, and bilateral PPC.

r Lee et al. r

r 10 r

bootstrap resampling [Efron, 1979]-based test by buildinga 95% confidence band for the difference of pre-test andpost-test. More specifically,

1. For a given ROI and a given practice group, obtainthe difference between pre-test and post-test timecourses for each subject.

2. Obtain a bootstrap sample of the difference curveswith the size corresponding to the number of subjectsof the given practice group and calculate the averagecurve.

3. Repeat Step 2 1,000 times.4. Use the 1,000 average curves from Step 3 to construct

a 95% confidence band.5. Repeat Steps 1–4 for each practice group and each

ROI.

If pre-test and post-test for an ROI showed similar acti-vation patterns, the confidence band for the differencecurve would be expected to include 0 for most time points.If attenuation occurred at post-test the confidence bandswould miss 0 at peaks (with a positive sign) and valleys(with a negative sign). Application of this bootstrapapproach suggested that four ROIs showed a strongattenuation at post-test for the antisaccade practice grouponly: SEF, FEF, SPL, and cuneus (see Figs. 7–10 and Sup-porting Information for the plots for the entire group ofROIs). The time course for each ROI and each practicegroup is shown as solid lines for pre-test and dashed linesfor post-test. Time points where statistically significantattenuations occurred at post-test are identified by darkgray bands. Visual inspection of these four ROIs showedthat attenuation at post-test was characteristic in theseregions only for the antisaccade practice group.

The converse of attenuation at post-test was also possi-ble and such an amplification effect was indicated by con-fidence bands that would miss 0 at peaks (with a negativesign) and valleys (with a positive sign). Time points thatshow statistically significant amplification at post-test areidentified by light gray bands in the plots (Figs. 7–10 and

Supporting Information). Although the prosaccade andfixation practice groups showed some periods of ampli-fication, there was little evidence of amplification in theantisaccade group at post-test.

The findings from this ROI analysis are summarized inTable II. The values indicate differences between the pro-portion of time points (out of 70) that had significantattenuations (dark gray bands) and the proportion of timepoints showing amplification (light gray bands) for eachpractice group at each ROI. A general pattern emergedshowing that in the ROIs identified a priori as saccadic cir-cuitry [Dyckman et al., 2007] there was a general attenua-tion (positive values showing less activation at post-test)associated with task-specific practice (i.e., the antisaccadegroup). There appeared to be little dramatic change in ei-ther direction in the prosaccade practice group. The onlygroup that showed evidence of consistent amplification(negative values showing more activation at post-test)across time was the fixation practice group. This patternreplicates that shown in Figure 6 but with increased speci-ficity and improved ability to evaluate which regions ofthe circuitry may be crucial for practice-related changes inperformance.

CONCLUSIONS

This study evaluated brain activation associated withpractice-related cognitive control during eye movementtasks. All participants were tested on antisaccade perform-ance at pre-test and post-test while fMRI data wereacquired. In between pre-test and post-test, participants

TABLE I. Number of activated voxels by slice for pre-

test only (red), both pre-test and post-test (orange),

and post-test only (yellow) for the three practice

groups: antisaccade (Anti), prosaccade (Pro), and

fixation (Fix) corresponding to Figure 5

Slice

Pre-test only(red)

Pre-test andpost-test (orange)

Post-test only(yellow)

Anti Pro Fix Anti Pro Fix Anti Pro Fix

29 43 41 16 51 16 31 39 33 9230 57 15 16 61 46 41 29 34 8331 38 16 9 43 22 33 33 40 6332 34 16 6 24 18 14 15 22 54Total 172 88 47 179 102 119 116 129 292

Figure 6.

Interaction plot with 95% error bars between test session (pre-

test and post-test) and proportion of activated voxels (i.e., the

number of activated voxels/total number of voxels in a slice) by

practice group: Antisaccade in black, prosaccade in dark gray,

and fixation in light gray.


r 11 r

engaged in daily practice of one assigned eye movementtask, antisaccade, prosaccade, or fixation. Over the seriesof analyses conducted, task-consistent practice was associ-

ated with decreased activation from pre-test to post-test.The antisaccade practice group was the only group toshow consistently decreased antisaccade-related activation

Figure 7.

SEF: The average time series plots for pre-test (solid) and post-test (dashed) are drawn for the

three practice groups. Black dashed lines at the bottom show the timing of antisaccade task

blocks. Dark gray bands represent time points where attenuations at post-test were statistically

significant, and light gray bands represent time points where amplifications at post-test were stat-

istically significant by the bootstrap resampling method.

Figure 8.

FEF: The average time series plots for pre-test (solid) and post-test (dashed) are drawn for the





r Lee et al. r

r 12 r

in saccadic circuitry, particularly evident in FEF, SEF, SPL,and cuneus. This pattern may be associated with increasedefficiency such that fewer neural resources were necessary

to support the response due to task-specific practice. Asimilar pattern has been reported in numerous studiesshowing decreased activity associated with learning

Figure 9.

SPL: The average time series plots for pre-test (solid) and post-test (dashed) are drawn for the





Figure 10.

Cuneus: The average time series plots for pre-test (solid) and post-test (dashed) are drawn for

the three practice groups. Black dashed lines at the bottom show the timing of antisaccade task





r 13 r

[Garavan et al., 2000, Jansma et al., 2001, Koch et al., 2006,Poldrack et al., 2005].

A distinctly different pattern of neural activationresponse was demonstrated by the two groups who prac-ticed test-inconsistent tasks (prosaccade and fixation). Theprosaccade practice group was most likely to showunchanged activation levels between pre-test and post-test.Regional analyses suggested some minor decreases in acti-vation at post-test, but they were neither as widespreadnor as strong as those observed in the antisaccade practicegroup. Thus, the brain activation across time in the prosac-cade group was relatively stable. The fixation practicegroup was the only one that demonstrated generallyincreased activation at post-test. The persistent increases inthe fixation group may suggest strengthening of the neuralcircuitry that inhibits rapid responses away from the fixa-tion target [Leigh and Zee, 2006]. Although both prosac-cade and fixation practice groups constituted controlconditions, their different task requirements would predictdifferent brain activation patterns.

In summary, these data demonstrated that changes inbrain activation patterns of the well-identified [McDowellet al., 2008, Munoz and Everling, 2004, Sweeney et al.,2007] neural circuitry supporting antisaccade performancediffered between task-specific and task-irrelevant practice.The changes in brain activation were observed despite noquantifiable differences in antisaccade performance atpost-test, which may imply that brain activation patterns

are more sensitive than behavioral measures as indicationsof plasticity. Of course, this intervention was merely 1week in duration. It may take longer for changes in the be-havioral component of the task to be manifest if they are aless sensitive measure of changes in brain function.

A notable component of this study was the applicationof a novel combination of statistical techniques selectedspecifically to circumvent typically acknowledged meth-odological problems associated with the structure of thedata generated via fMRI and/or the model-based GLManalyses as they are commonly applied to fMRI data.These problems include massive-sized data that areill-balanced and influenced by temporal autocorrelations,which we successfully moderated via the integration andapplication of clustering routines and wavelet analyses.The difficulties inherent in accurate modeling of the BOLDresponse, which is standard in GLM analyses, was like-wise avoided in this article by using a data-driven tech-nique. This improved analysis successfully identified thewell-known brain activation patterns known to supportantisaccade performance. Not only was the circuitry iden-tified but also the analysis of practice-related changescould be evaluated with greater confidence because thesemethods were more resistant to common problems. Assuch, the differences between practice groups across timewere tested in a rigorous manner and revealed task-spe-cific brain activation changes in cognitive control acrosstime.

ACKNOWLEDGMENTS

The authors thank Jazmin Camchong for help with datacollection. They also thank the referees for helpful com-ments. The Franklin Research Seed Grant, UGA Psychol-ogy Department.

REFERENCES

Balslev D, Nielsen F, Frutiger S, Sidtis J, Christiansen T, Svarer C,Strother S, Rottenberg D, Hansen L, Paulson O, Law I (2002):Cluster analysis of activity-time series in motor learning. HumBrain Mapp 15:135–145.

Benjamini Y, Hochberg Y (1995): Controlling the false discoveryrate: A practical and powerful approach to multiple testing. JR Stat Soc Ser B 57:289–300.

Benjamini Y, Hochberg Y (2000): On the adaptive control of thefalse discovery rate in multiple testing with independent statis-tics. J Educ Behav Stat 25:60–83.

Bullmore ET, Suckling J, Zelaya F, Long C, Honey GD, RoutledgeC, Ng V, Fletcher PC, Brown J, Williams SCR (2003): Waveletsand statistical analysis of functional magnetic resonanceimages of the human brain. Stat Methods Med Res 12:375–399.

Camchong J, Dyckman K, Austin B, Clementz B, McDowell J(2008): Common neural circuitry supporting volitional sac-cades and its disruption in schizophrenia patients and rela-tives. Biol Psychiatry 64:1042–1050.

Chein JM, Schneider W (2005): Neuroimaging studies of practice-related change: fMRI and meta-analytic evidence of a domain-

TABLE II. Differences of the proportions of statistically

significant attenuations (positive values) and

amplifications (negative values) for the three groups at

each ROI

ROI Anti Pro Fix

SEF 0.16 �0.01 �0.01FEF 0.19 0.04 �0.01SPL 0.11 0.06 �0.07

Cuneus 0.09 �0.03 �0.07

Thalamus 0.04 0.03 �0.09

IPL �0.03 0.00 �0.14

PFC—L �0.04 0.07 0.00PFC—R 0.04 �0.01 �0.01Striatum 0.06 0.04 �0.01MOG �0.01 �0.01 0.01IFC 0.01 0.03 0.00

These values are calculated as: [(proportion of time points attenu-ated (i.e., number of time points significantly decreased at post-test) � proportion of time points amplified (i.e., number of timepoints significantly increased at post-test))/70]. Values greaterthan or equal to 0.07 are boldfaced. [SEF (Talairach x ¼ 0, y ¼ 1, z¼ 56); FEF (LH �34, 5, 53; RH 34, 7, 52); SPL (LH �22, �59, 55;RH 15, �60, 55); cuneus (LH �14, �79, 7; RH 10, �78, 7); thala-mus (LH �13, �13, 12; RH 10, �13, 14); IPL (LH �47, �31, 25;RH 54, �42, 23); PFC (LH �34, 34, 38; RH 36, 43, 26); striatum(LH �18, 1, 17; RH 15, 2, 18); MOG (LH �29, �85, 4; RH 24, �87,4); and IFC (RH 37, 32, 1)].

r Lee et al. r

r 14 r

general control network for learning. Cogn Brain Res 25:607–623.

Cox RW (1996): AFNI: Software for analysis and visualization offunctional magnetic resonance neuroimages. Comput BiomedRes 29:162–173.

Daubechies I (1992): Ten Lectures on Wavelets. Philadelphia: Soci-ety for Industrial and Applied Mathematics.

Donoho D, Johnstone J (1994): Ideal spatial adaptation by waveletshrinkage. Biometrika 81:425–455.

Donoho D, Johnstone J (1995): Adapting to unknown smoothnessvia wavelet shrinkage. J Am Stat Assoc 90:1200–1224.

Dyckman K, Camchong J, Clementz B, McDowell J (2007): Aneffect of context on saccade-related behavior and brain activity.Neuroimage 36:774–784.

Efron B (1979): Bootstrap methods: Another look at the jackknife.Ann Stat 7:1–26.

Fadili M, Ruan S, Bloyet D, Mazoyer B (2000): A multistep unsu-pervised fuzzy clustering analysis of fMRI time series. HumBrain Mapp 10:160–178.

Fischer B, Hartnegg K, Mokler A (2000): Dynamic visual percep-tion of dyslexic children. Perception 29:523–530.

Ford A, Goltz C, Brown G, Everling S (2005): Neural processesassociated with antisaccade task performance investigatedwith event-related fMRI. J Neurophysiol 94:429–440.

Forman SD, Cohen JD, Fitzgerald M, Eddy WF, Mintun MA,Noll,DC (1995): Improved assessment of significant activationin functional magnetic resonance imaging (fMRI): Use of acluster-size threshold. Magn Reson Med 33:636–647.

Garavan H, Kelley D, Rosen A, Rao SM, Stein EA (2000): Practice-related functional activation changes in a working memorytask. Microsc Res Tech 51:54–63.

Gautama T, Van Hulle MM (2005): Estimating the global order ofthe fMRI noise model. Neuroimage 26:1211–1217.

Genovese CR, Wasserman L (2005): Confidence sets for nonpara-metric wavelet regression. Ann Stat 33:698–729.

Genovese CR, Lazar NA, Nichols TE (2002): Thresholding of sta-tistical maps in functional neuroimaging using the false dis-covery rate. Neuroimage 15:870–878.

Goutte C, Toft P, Rostrup E, Nielsen F, Hansen L (1999): On clus-tering fMRI time series. Neuroimage 9:298–310.

Hartigan A, Wong A (1979): Algorithm AS 136: A K-means clus-tering algorithm. J R Stat Soc Ser C 28:100–108.

Heller R, Stanley D, Yekutieli D, Rubin N, Benjamini Y (2006):Cluster based analysis of fMRI data. Neuroimage 33:599–608.

Jansma J, Ramsey N, Slagter H, Kahn RS (2001): Functional ana-tomical correlates of controlled and automatic processing. JCogn Neurosci 13:730–743.

Johnson A, Wichern W (2002): Applied Multivariate StatisticalAnalysis, 5th ed. New York: Prentice-Hall.

Jolliffe T (2002): Principal Component Analysis, 2nd ed. NewYork: Springer.

Keedy SK, Ebens CL, Keshavan MS, Sweeney JA (2006): Func-tional magnetic resonance imaging studies of eye movementsin first episode schizophrenia: Smooth pursuit, visually guidedsaccades and the oculomotor delayed response task. PsychiatryRes 146:199–211.

Kelly A, Garavan H (2005): Human functional neuroimaging ofbrain changes associated with practice. Cereb Cortex 15:1089–1102.

Koch K, Wagner G, von Consbruch K, Nenadic I, Schultz C, EhleC, Reichenbach J, Sauer H, Schlosser R (2006): Temporalchanges in neural activation during practice of information re-

trieval from short-term memory: An fMRI study. Brain Res1107:140–150.

Leigh J, Zee D (2006): The Neurology of Eye Movements. NewYork: Oxford University Press.

Luo W, Nichols T (2003): Diagnosis and exploration of massivelyunivariate neuroimaging models. Neuroimage 19:1014–1032.

McDowell J, Dyckman K, Austin B, Clementz A (2008): Neuro-physiology and neuroanatomy of reflexive and volitional sac-cades: Evidence from studies of humans. Brain Cogn 38:255–270.

Mumford JA, Poldrack RA (2007): Modeling group fMRI data. SocCogn Affect Neurosci 2:251–257.

Munoz P, Everling S (2004): Look away: The anti-saccade task andthe voluntary control of eye movement. Nat Rev Neurosci5:218–228.

Muri RM, Heid O, Nirkko AC, Ozdoba C, Felblinger J, Schroth G,Hess CW (1998): Functional organisation of saccades and anti-saccades in the frontal lobe in humans: A study with echo pla-nar functional magnetic resonance imaging. J NeurolNeurosurg Psychiatry 65:374–377.

O’Driscoll A, Alpert M, Matthysse W, Levy L, Rauch L, HolzmanS (1995): Functional neuroanatomy of antisaccade eye move-ments investigated with positron emission tomography. ProcNatl Acad Sci USA 92:925–929.

Park C, Lazar NA, Ahn J, Sornborger (2010): A multiscale analysisof the temporal characteristics of resting-state fMRI data. JNeurosci Methods 193:1407–1433.

Park C, Ahn J, Hendry M, Jang W (2011): Analysis of long periodvariable stars with nonparametric tests for trend detection. JAm Stat Assoc 106:832–845.

Paus T (1996): Location and function of the human frontal eyefield: A selective review. Neuropsychologia 34:475–483.

Poldrack RA, Sabb FW, Foerde K, Tom SM, Asarnow RF, Book-heimer SY, Knowlton BJ (2005): The neural correlates of motorskill automaticity. J Neurosci 25:5356–5364.

Raemaekers M, Jansma JMJ, Cahn W, Van der Geest JN, Van derLinden JA, Kahn RS, Ramsey NF (2002): Neuronal substrate ofthe saccadic inhibition deficit in schizophrenia investigatedwith 3-dimensional event-related functional magnetic reso-nance imaging. Arch Gen Psychiatry 59:313–320.

Schneider W, Chein J (2003): Controlled and automatic processing:Behavior, theory and biology. Cogn Sci 27:525–559.

Schneider W, Shiffrin R (1977): Controlled and automatic humaninformation processing. I. Detection, search, and attention. Psy-chol Rev 84:1–66.

S� endur L, Maxim V, Whitcher B, Bullmore E (2005): Multiple hy-pothesis mapping of functional MRI data in orthogonal andcomplex wavelet domains. IEEE Trans Signal Process 53:3413–3426.

Stanberry L, Nandy R, Cordes D (2003): Cluster analysis of fMRIdata using dendrogram sharpening. Hum Brain Mapp 20:201–219.

Stoev S, Taqqu M, Park C, Marron J (2005): On the wavelet spec-trum diagnostic for Hurst parameter estimation in the analysisof internet traffic. Comp Netw 48:423–445.

Sweeney JA, Luna B, Keedy SK, McDowell JE, Clementz BA(2007): fMRI studies of eye movement control: Investigatingthe interaction of cognitive and sensorimotor brain systems.Neuroimage 36:T52–T60.

Talairach J, Tournoux P (1988): Co-planar Stereotaxic Atlas of theHuman Brain: 3-Dimensional Proportional System—An Approachto Cerebral Imaging. New York: ThiemeMedical Publishers.


r 15 r

Turkheimer FE, Smith CB, Schmidt K (2001): Estimation of thenumber of ‘‘true’’ null hypotheses in multivariate analysis ofneuroimaging data. Neuroimage 13:920–930.

Van De Ville D, Unser M (2008): False discovery rate for wavelet-based statistical parametric mapping. IEEE J Sel Top SignalProcess 2:897–906.

Veitch D, Abry P (1999): Wavelet analysis of long-range depend-ence traffic. IEEE Trans Inform Theory 44:2–15.

Wanga Z, Childressc AR, Wanga J, Detrea JA (2007): Support vec-tor machine learning-based fMRI data group analysis. Neuro-image 36:1139–1151.

Windischberger C, Lamm C, Bauer H, Mosera E (2003): Humanmotor cortex activity during mental rotation. Neuroimage20:225–232.

Woolrich W, Ripley D, Brady M, Smith M (2001): Temporal auto-correlation in univariate linear modeling of fMRI data. Neuro-image 14:1370–1386.

Worsley KJ (2003): Detecting activation in fMRI data. Stat Meth-ods Med Res 12:401–418.

Worsley KJ, Friston KJ (1995): Analysis of fMRI time-series revis-ited—Again. Neuroimage 2:173–181.

Ye N (2003): The Handbook of Data Mining, Mahwah. New Jer-sey: Lawrence Erlbaum Associates.

Ye J, Lazar N, Li Y (2009): Geostatistical analysis in clusteringfMRI time series. Stat Med 28:2490–2508.

APPENDIX: NO TREND TEST

We summarize the no trend test proposed by Park et al.[2011]. Let Yit denote the BOLD signal value at the ithvoxel and the tth time point. Then, we can set up a regres-sion model as follows. Suppose that

Yit ¼ fit

n

� �þ eit

,where fi is a trend function in the ith voxel, n is the totalnumber of time points, and eit is the error term in the ithvoxel time series with mean 0 and standard deviation r.Our interest is to test H0 : fi ¼ 0 for each i, which impliesthat the corresponding voxel is not active. The proposedapproach transforms this null hypothesis to the waveletdomain. To simplify notation, we drop i from here on.

Let u and w denote scaling and wavelet functions,respectively, and define their dilated and translated ver-sions as

uj;kðuÞ ¼ 2j=2uð2ju� kÞ; wj;kðuÞ ¼ 2j=2wð2ju� kÞ; j; k 2 Z;

where j and k represent the scale and location, respec-tively. Assume that a trend function f[L2[0,1], then one canexpand f with the wavelet basis functions:

f ðxÞ ¼X2J0�1

k¼0akuJ0 ;kðxÞ þ

X1j¼J0

X2j�1

k¼0bj;kwj;kðxÞ;

where J0 is a positive integer, ak ¼ hf ;uJ0 ;ki; bj;k ¼ hf ; wj;ki,and h�; �i denote the inner product in L2[0,1]. Inpractice, the series is truncated based on the samplesize n ¼ 2J1ðJ1 > J0Þ. In our analysis, we choose J0 ¼ 3and J1 ¼ 6 by binning the data to avoid boundaryeffects. The wavelet coefficients can be estimated asfollows:

~ak ¼ 1

n

Xn

t¼1YtuJ0;kðt=nÞ and ~bJ;k ¼

1

n

Xn

t¼1Ytwj;kðt=nÞ:

Denoting ln ¼ ðl1; : : : ;lnÞ ¼ ða0; : : : ; a2J0 � 1;bJ0;0; : : : ;bJ02J0�1; ; : : : ;bJ1�1;2J1�1�1Þ the null hypothesis H0 : f ¼ 0 wetest becomes H0 : ln ¼ 0;, i.e., H0 : all the wavelet coeffi-cients are 0.The standard deviation of the error can be esti-mated using the wavelet coefficient estimates

r2 ¼ 2Xn

t¼n2þ1

~l2t ;

where ð~l1; : : : ; ~lnÞ ¼ ð~a0; : : : ; ~bJ1�1;2J1�1 �1

Þ. Xe consider thesoft thresholding estimators to increase the power of thetest. Define ðl1; : : : ; lnÞ ¼ ða0; : : : ; bJ1�1�1Þ, where ak ¼~aksgn(bJ;k ¼ sgnð~bJ;kÞðj~bj;kj � kÞþ. Here, ð�Þþ indicates thatthe negative pieces are set to zero. For k, one can use eitherthe universal threshold, k ¼ r

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi2log n=n

p, or the levelwise

SureShrink rule [Donoho and Johnstone, 1995], whichminimizes

SnðkjÞ ¼ r2

n2J0 þ

XJ1�1

j¼J0SjðkjÞ;

where

SjðkjÞ ¼X2J0�1

k¼0

r2

n� 2

r2

nIf~jbJ;kj � kjg þminð~b2J;k;k2j Þ

� �;

J0 � j � J1 � 1:

Under H0 : ln ¼ 0, Genovese and Wasserman [2005]showed that

ffiffiffin

p ðPnt¼1 l

2t � SnðkÞÞffiffiffi

2p

r2� Nð0; 1Þ;

which leads to the rejection region

Xnt¼1

l2t > r2 zaffiffiffiffiffiffiffiffin=2

p þ SnðkÞ

where za stands for the 100ð1� aÞ% percentile of thestandard normal distribution.

r Lee et al. r

r 16 r

practice-related changes in neural activation patterns

Documents