Working Memory Coding of Analog Stimulus Properties in the Human Prefrontal Cortex

Bernhard Spitzer1, Matthias Gloel2, Timo T. Schmidt2,3 and Felix Blankenburg1,3

1Dahlem Institute for Neuroimaging of Emotion, Freie Universität Berlin, Berlin 14195, Germany, 2Bernstein Center forComputational Neuroscience, Berlin 10115, Germany and 3Center for Adaptive Rationality (ARC), Max Planck Institute forHuman Development, Berlin 14195, Germany

Address Correspondence to B. Spitzer, Dahlem Institute for Neuroimaging of Emotion, Freie Universität Berlin, Habelschwerdter Allee 45, 14195Berlin, Germany. Email: bernhard.spitzer@fu-berlin.de

Building on evidence for working memory (WM) coding of vibro-tactile frequency information in monkey prefrontal cortex, recentelectroencephalography studies found frequency processing inhuman WM to be reflected by quantitative modulations of prefrontalupper beta activity (20–30 Hz) as a function of the to-be-maintainedstimulus attribute. This kind of stimulus-dependent activity hasbeen observed across different sensory modalities, suggesting ageneralized role of prefrontal beta during abstract WM processingof quantitative magnitude information. However, until now the avail-able empirical evidence for such quantitative WM representationremains critically limited to the retention of periodic stimulus fre-quencies. In the present experiment, we used retrospective cueingto examine the quantitative WM processing of stationary (intensity)and temporal (duration) attributes of a previously presented tactilestimulus. We found parametric modulations of prefrontal betaactivity during cued WM processing of each type of quantitativeinformation, in a very similar manner as had before been observedonly for periodic frequency information. In particular, delayed pre-frontal beta modulations systematically reflected the magnitude ofthe retrospectively selected stimulus attribute and were functionallylinked to successful behavioral task performance. Together, thesefindings converge on a generalized role of stimulus-dependent pre-frontal beta-band oscillations during abstract scaling of analogquantity information in human WM.

Keywords: EEG, oscillations, somatosensory, stimulus coding, workingmemory

Introduction

It is widely accepted that prefrontal cortex (PFC) plays a keyrole in working memory (WM), that is, in operations enablingthe maintenance and online processing of information that nolonger exists in the environment (for reviews, see Miller andD’Esposito 2005; Pasternak and Greenlee 2005; D’Esposito2007). Studies of human WM, mostly in the visual and verbaldomains, routinely reported PFC activity in terms of hemody-namic responses (Wager and Smith 2003; Nee et al. 2012), butalso in terms of changes in oscillatory activity as assessed byelectro- or magnetoencephalography (EEG/MEG) (e.g., Tallon-Baudry et al. 1998; for review, see Benchenane et al. 2011).Thereby, WM function was typically examined by contrastingconditions of high versus no (or lower) WM processing, yield-ing insights into task-dependent activity associated with differ-ent WM operations and materials (Gazzaley and Nobre 2012).

Another line of research in the somatosensory domain hasrecently succeeded in delineating direct, stimulus-dependentsignatures of WM processing in the human PFC. In particular,complementing evidence for parametric WM coding of vibro-tactile frequency information in the monkey PFC (Romo et al.

1999; Barak et al. 2010), human EEG studies found WM pro-cessing of different frequencies to be systematically reflectedby parametric modulations of right prefrontal oscillatoryactivity in the upper beta band (20–30 Hz; Spitzer et al. 2010;Spitzer and Blankenburg 2011). Recently, this prefrontalactivity has also been observed in other sensory modalities(Spitzer and Blankenburg 2012), suggesting a general role ofparametric beta oscillations during WM processing of analogstimulus quantities. However, the currently available empiri-cal evidence for such parametric WM processing remainsrestricted to tasks that require the maintenance of periodicfrequency information, thus leaving open whether modu-lations of prefrontal beta may represent other stimulusattributes, as well. Moreover, the precise nature of the infor-mation, which is parametrically encoded during WM proces-sing of different frequencies, is not clear yet. Periodic stimuliof different temporal frequency, as defined by the number ofcyclic repeats per second, differ in their fine-grained temporalproperties and might be evaluated in terms of the perceivedspeed of the stimulus dynamics. However, different stimulusfrequencies may also be evaluated with respect to the inferredintensity of stimulation, with faster oscillating stimulisuggesting the presence of higher physical energy. It there-fore remains unknown to what extent frequency-dependentparametric modulations of prefrontal beta activity mightreflect a unique concept of periodic stimulus frequency, orthe processing of temporal structure or physical energy, or amore general abstract representation of sensory quantity.

To shed light on these questions, in the present experiment,we systematically examined the WM processing of temporal(duration) and stationary (intensity) quantitative attributes ofto-be-maintained vibrotactile stimuli. We used a modifieddelayed-match-to-sample (DMTS) task (Fig. 1A), which involvedretrospective selection of the task-relevant stimulus attribute forsustained maintenance, allowing us to disambiguate the activeWM processing of either stimulus dimension after identical en-coding conditions. Examining prefrontal EEG oscillations, wehypothesized that if parametric WM effects (Romo et al. 1999;Spitzer et al. 2010; Spitzer and Blankenburg 2012) were un-iquely linked to a proprietary concept of periodic frequencies,no parametric modulations, neither by duration nor by inten-sity, should be observed. If, on the other hand, frequency-dependent modulations relate to the processing of temporalstimulus properties, then prefrontal beta activity might bemodulated—if at all—by stimulus duration, but not by intensity.In contrast, if the previous frequency-dependent modulationsintrinsically reflect a scaling of perceived physical energy, inde-pendent of its precise temporal structure, then a modulation byintensity only is expected. If however, as proposed recently(Spitzer and Blankenburg 2012), parametric WM processing inthe PFC relies on a generalized WM representation of quantity,

© The Author 2013. Published by Oxford University Press. All rights reserved. For Permissions, please e-mail: journals.permissions@oup.com

Cerebral Cortexdoi:10.1093/cercor/bht084

Cerebral Cortex Advance Access published March 31, 2013 at U

niversity of Chicago on M

ay 14, 2014http://cercor.oxfordjournals.org/

Dow

nloaded from

then prefrontal upper beta may be modulated by either of the 2stimulus attributes, when their abstract magnitude is activelyprocessed in WM.

Materials and Methods

SubjectsTwenty-six healthy volunteers (20–35 years; 12 females and 14 males)participated in the experiment with written informed consent. Twosubjects were excluded from analysis due to excessive EEG artifactsand deficient performance in the behavioral task (<60% correct). Thestudy was approved by the Ethical Committee of the Charité Univer-sity Hospital Berlin and corresponded to the Human Subjects Guide-lines of the Declaration of Helsinki.

Stimuli and Behavioral TaskVibrotactile stimulation of the left index finger was delivered by a16-dot piezoelectric Braille display (4 × 4 quadratic matrix, 2.5 mmspacing) controlled by a programmable stimulator (Piezostimulator,QuaeroSys, St. Johann, Germany). The pins of the Braille displaywere simultaneously driven by a 123-Hz sinusoidal carrier signal,which was amplitude modulated by a 33-Hz sinus function, creatingthe percept of a “flutter” vibration (Romo and Salinas 2003) at 33 Hz.These carrier and modulation frequencies were chosen to minimizethe risk of artifact contamination at EEG frequencies <90 Hz and toprevent steady-state evoked responses in the EEG frequency range ofmain interest (20–30 Hz). The sound of the tactile stimulation wasmasked by white noise (∼80 dB) played via loudspeakers throughoutthe entire experiment. None of the participants reported hearing anysound attributable to the tactile stimulation.

An illustration of the experimental protocol is shown in Figure 1A.After a variable prestimulus interval (1000–1500 ms), each trial startedwith the presentation of a base stimulus (s1). The amplitude

(vibration intensity) of s1 was randomly chosen to be 39.1%, 53.7%,68.4%, or 83.0% of the stimulator’s maximum capacity (correspondingto 1600, 2200, 2800, and 3400 arbitrary amplitude units of 4095 maxi-mally available units). The resulting stimulation intensities were notpainful, and ranged between the 26- and 56-fold of the average detec-tion threshold for this type of stimuli (1.5% stimulator capacity, or60.3 amplitude units; assessed in a follow-up test on 5 of the partici-pants using the method-of-limits procedure). The duration of s1 wasrandomly chosen to be 350, 450, 550, or 650 ms, independent of s1’sintensity. Eight hundred milliseconds after offset of s1, a visual cue(“A” or “B,” randomly selected on each trial) indicated whether sub-jects should for the remainder of the trial focus on s1’s intensity(A-cue) or duration (B-cue). After a 2500-ms retention interval, acomparison stimulus (s2) was presented, which differed from s1 bothin intensity (by ±7.3% stimulator capacity, corresponding to ±300 am-plitude units) and in duration (by ±150 ms). In the A-cue condition,subjects’ task was to indicate whether s2 was weaker or stronger thans1 by pressing a response button with the right hand either once ortwice. Analogously, in the B-cue condition, subjects were asked toindicate whether s2 was shorter or longer than s1. After several prac-tice trials, each participant completed 7 blocks of 64 trials, for a totalof 224 discrimination trials in each cue condition.

EEG Recording and AnalysisEEG was recorded using a 64-channel active electrode system (Acti-veTwo, BioSemi, Amsterdam, Netherlands), with electrodes placed inan elastic cap according to the extended 10-20 system. Individual elec-trode locations were registered using a stereotactic electrode position-ing system (Zebris Medical GmbH, Isny, Germany). Vertical andhorizontal eye movements were recorded from 4 additional channels.Signals were digitized at a sampling rate of 2048 Hz, off-line band-pass filtered (1–100 Hz), downsampled to 512 Hz, and average refer-enced. All analyses were carried out using SPM8 for MEG/EEG(update revision number 4667, 27 February 2012; Wellcome Depart-ment of Cognitive Neurology, London, UK: www.fil.ion.ucl.ac.uk/spm/) and custom MATLAB code (The Mathworks, Inc., Natick, USA).The EEG was corrected for eye blinks using calibration data to gener-ate individual artifact coefficients and adaptive spatial filtering (fordetails, see Ille et al. 2002). Remaining artifacts were rejected byexcluding all data containing signal amplitudes >80 µV from analysis.

Spectral AnalysisTime frequency (TF) representations of spectral power (5–48 Hz) inthe retention interval (−800–2500 ms, relative to cue onset) were ob-tained by applying a tapered sliding window fast Fourier transform(FFT) with a Hanning taper and an adaptive time window of 7 cycleslength, using the FieldTrip function ft_specest_mtmconvol.m pro-vided with SPM8. This method is very similar to a Morlet wavelettransform (cf. Bruns 2004), but provides additional tapering optionsfor optimum sensitivity at higher frequency ranges. Here, exploratoryanalysis of higher frequency bands (>48 Hz), using a multitaperedFFT, yielded no significant effects. To avoid potential results distor-tions by systematic stimulus-induced effects in the precue interval(Supplementary Fig. 2), the TF spectra were normalized relative to theaverage power in the entire data epoch (−800–2500 ms) to representpercentage power changes over time. This normalization approachretains condition-specific differences in the TF spectras’ temporalprofile only. Control analyses of the average reference spectra usedfor the normalization showed no systematic differences in any of thereported contrasts (all frequency bins P > 0.05), rendering it unlikelythat stationary effects had been masked out in the normalized TFdata. Note that subsequent analysis involved direct comparisons ofthe cue-induced TF profiles between the A- and B-tasks after identicals1 stimulation and encoding conditions (cf. Figs 1 and 2), ruling out apotential contamination of the delay period results by s1-inducedactivity and/or stimulation artifacts. TF bins for which the sliding FFTwindow overlapped with s2 onset, at the end of the analysis window,were excluded from analysis (cf. transparent masks in Figs 1 and 2).

Figure 1. (A) Experimental task. On each trial, a vibrotactile stimulus (s1) of a givenintensity and duration was presented at the left index finger. A visual cue instructedsubjects to retrospectively focus on s1’s intensity (“A”) or duration (“B”) throughout aretention interval (green), for subsequent comparison against the correspondingattribute of s2. (B) Upper left: Time frequency statistical parametric maps ofdifferences in oscillatory activity between the A- and B-tasks, collapsed across allvalues of s1. Dashed gray rectangles indicate TF window and channels where asignificant effect was observed (P<0.05, FWE). Lower left: Time-course of beta(15–25 Hz) activity changes in the A- and B-tasks, for the channels outlined above.Right: SPM 3D source reconstruction of the difference in beta power between 700and 1300 ms.

2 WM Coding of Analog Stimulus Properties • Spitzer et al.

at University of C

hicago on May 14, 2014

http://cercor.oxfordjournals.org/D

ownloaded from

Statistical AnalysisThe statistical design was implemented in SPM8, using the generallinear model (GLM) applied to each subject’s single-trial TF data. Towarrant conformity with the GLM under normal error assumptions(Kiebel et al. 2005), the TF data were convolved with a 3 Hz × 500 ms(full-width at half maximum) Gaussian kernel. The GLM designmatrix (cf. Supplementary Fig. 1A) consisted of 2 dummy variablesspecifying the trials’ cue condition (A or B), 2 parametric regressorscoding the intensity of s1 under the respective cue conditions, and 2parametric regressors analogously specifying the duration of s1. Thevectors coding intensity and duration were zero-centered and normal-ized. Because across trials, intensity and duration were varied inde-pendently of each other, the resulting parametric regressors wereorthogonal, allowing us to estimate their respective impact within thesame regression model. The model was inverted using restrictedmaximum-likelihood estimation as implemented in SPM8, yielding

beta parameter estimates for each model regressor at each TF bin. TFcontrasts of interest were then computed by weighted summation ofindividual regressors’ beta estimates (cf. Supplementary Fig. 1B–D).

The individual contrast spectra were subjected to mass univariateanalysis on the group level, using 1-sample t-tests as implemented inSPM8 (for details, see Kiebel et al. 2005; Litvak et al. 2011). This in-volved computing for each TF bin a t-value that reflected the signifi-cance of the contrast. Family-wise errors (FWEs) in TF space werecontrolled using random field theory (RFT; Worsley et al. 1996; Kilneret al. 2005) to determine the FWE-corrected probability that, at agiven channel, a cluster of adjacent significant TF bins may have beenobtained by chance. A cluster was thereby defined as a group of adja-cent TF bins that all exceeded a threshold of P < 0.005 (uncorrected).RFT approximates the probability that a multidimensional statisticalmap exceeds some height or extent by chance, under the assumptionthat the error terms conform to a smooth random field with a Gaus-sian distribution (Brett et al. 2003). Convolving the normalized datawith a Gaussian smoothing kernel prior to analysis ensures this as-sumption. The FWE-corrected significance of a given cluster (Pcluster)is then determined using the distribution of the expected Eulercharacteristic, given the smoothness of the map, under the nullhypothesis of a continuous random field. Analysis of parametric TFcontrasts (Fig. 2, cf. Supplementary Fig. 1C,D) a priori focused onright prefrontal channels near electrode F2, where parametric modu-lations of upper beta activity were most consistently observed in ourprevious vibrotactile frequency work. If significant (Pcluster < 0.05,FWE) cue-induced parametric modulations were identified at F2, and/or at least 2 immediately neighboring right prefrontal channels, wecontinued with analyses of the condition-specific modulations in therespective frequency band, averaged over those adjacent channelsshowing a similar effect (cf. Fig. 2). Unless noted otherwise, onlycorrect discrimination trials were included in the reported analyses.Where applicable, error trials were analyzed separately.

Source ReconstructionThe sources of EEG activity were modeled using 3-dimensional (3D)source reconstruction as implemented in SPM8 (Friston et al. 2006).For each participant, a forward model was constructed, using an 8196vertex template cortical mesh coregistered to the individual electrodepositions via 3 fiducial markers. The forward model’s lead field wascomputed using the 3-shell boundary element method EEG headmodel available in SPM8 (Phillips et al. 2007). Prior to model inver-sion, the data were band-pass filtered around the frequency band ofinterest. Source estimates were then computed on the canonical meshusing multiple sparse priors (Friston et al. 2008) under group con-straints (Litvak and Friston 2008), including the data from all con-ditions of interest. For TF windows where significant effects wereidentified in the channel-level analysis, condition-specific TF contrastswere used to summarize oscillatory source power for specific fre-quency bands and at specific times as 3D images. This entailed con-volving the single-trial source activity with a series of Morletprojectors and weighting the average power from each condition witha Gaussian window centered on the time interval of interest. Thesource reconstructions were then analyzed on the group level usingconventional statistical parametric mapping procedures, using adisplay threshold of P < 0.001 (uncorrected). Additional nonsignifi-cant source reconstruction results are illustrated descriptively in Sup-plementary Figure 3. The SPM anatomy toolbox (Eickhoff et al. 2005)was used to establish cytoarchitectonic reference where possible.

Results

Behavioral ResultsOn average, participants correctly discriminated 69.3% of thestimulus intensities in the A-task, compared with 77.6% of thestimulus durations in the B-task (P < 0.001, 2-sided pairedt-test). A detailed description of the behavioral data is given inTable 1. As expected by Weber’s law (Fechner 1966),

Figure 2. Parametric modulations of prefrontal oscillatory activity by thetask-relevant stimulus attribute. (A) Modulations by stimulus intensity. Upper left:Time frequency statistical parametric map of the difference in parametric modulationby s1 intensity during retention in the A-cue condition compared with the B-cuecondition, for representative right prefrontal channels outlined by dashed rectangle inupper right. Upper right: Descriptive topographical map (color scaling as in left) of theparametric modulation outlined by dashed rectangle in upper left. Lower left:Time-courses of the parametric modulations by s1 intensity in the 2 task conditions,for the frequency band and channels outlined in upper left. “MI”: maintain intensity(A-cue); “MD”: maintain duration (B-cue). Lower right: 3D source reconstruction ofthe parametric modulation by intensity in the A-cue condition, for the TF windowoutlined in upper left. (B) Analogous to A, for the modulation by s1 duration in theB-cue condition compared with the A-cue condition. Source reconstruction failed toproduce statistically reliable results for the TF window outlined in upper left. Adescriptive illustration of the reconstructed prefrontal source is provided inSupplementary Figure 3.

Cerebral Cortex 3

at University of C

hicago on May 14, 2014

http://cercor.oxfordjournals.org/D

ownloaded from

discrimination accuracy in the A-task decreased with increas-ing s1 stimulus intensity (Table 1, upper; P < 0.001, lineartrend analysis, Bonferroni corrected). In contrast, accuracy inthe B-task was not systematically affected by s1 duration(P > 0.70). Thus, the lower mean accuracy in the A-task com-pared with the B-task was in particular due to poor discrimi-nation of high s1 intensities, compared with long s1 durations(cf. Table 1). In neither task did accuracy covary with vari-ations in the uncued stimulus dimension (i.e., duration in theA-task, or intensity in the B-task, both P’s > 0.05).

Average response times (RTs; measured relative to s2offset) were longer in the B-task (745 ms) when comparedwith the A-task (656 ms; P < 0.001). This is expected becauseduration judgments (B-task) require evaluation of s2 in fulllength. Likewise, RTs in both tasks decreased with increasingstimulus duration (Table 1, lower; both P’s < 0.001), butremained stable across stimulus intensities (both P’s > 0.50).

Oscillatory EEG ResponsesFirst, we explored possible differences in overall cue-inducedTF activity between the 2 task conditions during cued reten-tion (0–2500 ms, Fig. 1B, left). Statistical analysis indicatedthat, between 700 and 1300 ms after cue onset, activity in thetraditional sensorimotor beta band (15–25 Hz) over rightcentro-parietal channels (C4, C6, CP4, and CP6) was signifi-cantly decreased in the A-task compared with the B-task(Pcluster < 0.05, FWE; maximum at channel CP6, 20 Hz, 1050ms, t = 4.26). SPM 3D source reconstruction attributed thiseffect to the right primary somatosensory cortex (SI, areas 3b,1, and 2; Fig. 1B, right; P < 0.001) contralateral to s1 appli-cation (for related lateralization analyses, see Spitzer andBlankenburg 2011). No further differences in overall oscil-latory activity during the retention interval were foundbetween the 2 tasks.

Also evident from the time-courses illustrated in Figure 1B,lower; in both memory tasks, sensorimotor beta activity wasby tendency decreasing throughout the retention interval, inline with previous findings that tactile WM processing per seis typically accompanied by suppression of beta activity inearly somatosensory processing areas (e.g., Spitzer and Blan-kenburg 2011, 2012). Supplementary inspection of this effectin the present data affirmed that the overall power decrease

was attributable to traditional beta-band activity (15–25 Hz) inSI (for details, see Supplementary Fig. 2A).

Stimulus-Dependent Parametric Modulations ofPrefrontal Beta ActivityTo examine parametric modulations of oscillatory activity as afunction of the stimulus attribute processed in WM, we ana-lyzed parametric contrasts reflecting the strength of a linearrelation between TF activity and the different properties of s1(i.e., intensity or frequency, cf. Supplementary Fig. 1). Basedon our previous work, we focused on right prefrontal chan-nels near electrode F2, where delayed parametric modulationsin the upper beta frequency range (20–30 Hz) have been con-sistently found during WM maintenance of vibrotactile fre-quency (Spitzer et al. 2010; Spitzer and Blankenburg 2012,see Spitzer and Blankenburg 2011 for evidence that this typeof activity is typically right lateralized).

We first inspected the cue-induced parametric modulationby intensity in the A-task, compared with the modulationby intensity in the B-task (Fig. 2A). This contrast revealed asignificant (Pcluster < 0.05, FWE) cue-induced parametricmodulation in the upper beta frequency range (20–25 Hz)between 250 and 750 ms after cue onset (maximum atchannel F4, 22 Hz, 400 ms, t = 3.88). Note that the right pre-frontal scalp topography of this effect (Fig. 2A, upper right)might have been reinforced by a priori channel selection.Figure 2A (lower) illustrates the condition-specific time-courses of the average parametric upper beta-band modu-lations by s1 intensity. In the A-cue condition, when subjectsretrospectively focused on s1 intensity, a sustained modu-lation by intensity was evident between 250 and 1200 msafter cue onset (all time bins P < 0.05). No evidence for amodulation by s1 intensity was found in the B-cue condition(all time bins P > 0.05). Source reconstruction (Fig. 2A, lowerright) attributed the parametric modulation by intensity inthe A-cue condition to the inferior frontal gyrus (IFG, area45, P < 0.001) in the right lateral PFC, which conforms withprevious localizations of frequency-dependent WM activity(e.g., Spitzer et al. 2010).

Analogously, we examined the cue-induced parametricmodulation by duration in the B-task, compared with the modu-lation by duration in the A-task (Fig. 2B). This contrast revealeda significant (P < 0.05, FWE) cue-induced parametric modu-lation that was spectrally and topographically similar to themodulation by intensity reported above (23–29 Hz; maximumat channel F2, 27 Hz, 1550 ms, t = 3.26), but emerged in a latertime window (1300–1650 ms). The condition-specific time-courses of the average parametric modulations indicate thatupper beta was modulated by s1’s duration in the B-cue con-dition (Fig. 2B, lower, 1300–1700 ms, all time bins P < 0.05)where subjects retrospectively focused on duration, but not inthe A-cue condition (all time bins P > 0.05), where subjectsfocused on intensity. Source analysis of the modulation by dur-ation in the B-task yielded no results that exceeded the displaythreshold. Below threshold, at a considerably reduced level ofsignificance, the effect was attributed to a very similar source asthe modulation by intensity (IFG, area 45, see SupplementaryFig. 3).

A descriptive summary of the grand-average changes inupper beta activity as identified in the above analyses isshown in Figure 3A. The linearity of the group averaged data

Table 1Behavioral results

Accuracy (%)s1 intensity (a.u.) 1600 2200 2800 3400 Linear slopeA (maintain intensity) 74.1 76.9 66.1 60.1 −0.53*B (maintain duration) 75.0 78.3 77.7 79.1 0.12

s1 duration (ms) 350 450 550 650A (maintain intensity) 69.2 69.7 67.9 69.9 0.00B (maintain duration) 74.0 80.9 80.1 75.4 0.04

RT (ms)s1 intensity (a.u.) 1600 2200 2800 3400 Linear slopeA (maintain intensity) 660 671 660 632 −9.5B (maintain duration) 754 726 746 758 3.2

s1 duration (ms) 350 450 550 650A (maintain intensity) 743 691 603 592 −54.1*B (maintain duration) 794 770 730 688 −35.8*

Note: Discrimination performance. Upper: Mean s1–s2 discrimination accuracy (% correct) in theA- and B-tasks for each s1 intensity and duration, and linear regression slopes. Asterisks indicatesignificance of a linear trend (Bonferroni corrected). Lower: Same as upper, for RTs (ms),measured relative to the offset of s2.

4 WM Coding of Analog Stimulus Properties • Spitzer et al.

at University of C

hicago on May 14, 2014

http://cercor.oxfordjournals.org/D

ownloaded from

points may only approximately mirror the group statisticalanalysis of individual subject linearity derived from single-trialGLM analysis (cf. Fig. 2) and is shown for illustrative purposeonly.

Parametric Modulations on Error TrialsConcluding our analysis, we examined the potential behavior-al relevance of the parametric modulation effects reportedabove by examining the modulations on correct discrimi-nation trials when compared with error trials. To account forthe smaller number of incorrect than correct trials, we usedresampling statistics (Voytek et al. 2010; Spitzer and Blanken-burg 2012) to estimate a surrogate distribution of regressioncoefficients computed from subsets of randomly drawncorrect trials (matched to the number of error trials availableper cue and stimulus conditions). For computational effi-ciency, the resampling analysis was implemented usingsimple regression analysis instead of the full GLM model usedin the main analysis (cf. Supplementary Fig. 1).

Applied to correct trials, single-trial regression replicatedthe cue-specific modulations by intensity and duration out-lined in Figure 2A,B, lower, respectively (mean t-values 2.97and 2.51; both P’s < 0.05). Applied to error trials, in contrast,the analysis showed neither a modulation by intensity in theA condition (t =−0.37; P > 0.70), nor a modulation by durationin the B condition (t =−0.15; P > 0.80). Resampling statisticsbased on 1000 iterations indicated that, relative to the surro-gate distribution’s mean, the observed regression coefficienton incorrect trials was significantly decreased in the intensitytask (Fig. 3B, left; 0.0119 vs. −0.0033; z = 3.59, P < 0.001), and

showed a trend in the same direction for the duration task(Fig. 3B, right, 0.0204 vs. −0.0019, z = 1.72, P < 0.10), overallindicating a reduction in the stimulus-dependent parametricmodulations on error trials, compared with correct trials.

Discussion

The presented findings provide an important extension ofprevious evidence for parametric WM processing of periodicstimulus frequencies (Romo et al. 1999; Barak et al. 2010;Spitzer et al. 2010; Spitzer and Blankenburg 2011, 2012), bydemonstrating that oscillatory delay activity in the PFC can beparametrically driven by other stimulus attributes, as well. Inparticular, during retention of either intensity- or durationinformation, we identified stimulus-dependent modulations ofsimilar spectral profiles, with maximum modulation in theupper beta frequency range (20–30 Hz), thereby strongly re-sembling the frequency-dependent parametric WM effects re-ported recently (Spitzer et al. 2010; Spitzer and Blankenburg2011, 2012). Like in the previous work, the present para-metric EEG modulations were found to relate to successfulDMTS task performance and to critically depend on taskdemands to actively focus on a singular sensory quantity inWM. In particular, the observation of spectrally and topogra-phically similar prefrontal modulations by stationary and non-stationary stimulus attributes indicates that the previouslyreported frequency-dependent parametric WM phenomenamay not be uniquely linked to any specific temporal or non-temporal properties of periodic frequencies. Together, thepresent and previous findings corroborate a view of para-metric WM processing in the PFC as a generalized mechanism

Figure 3. (A) Summary illustration of grand-average cue-induced changes in prefrontal upper beta activity as a function of s1 intensity (left) and as a function of s1 duration(right). (B) Performance-related differences. Left: Parametric modulation by s1 intensity during the maintenance of intensity, computed from correct trials (saturated red)compared with incorrect trials (light red). On incorrect trials, the parametric modulation was significantly reduced (z=3.59, P< 0.001; resampling statistics based on 1000iterations). Right: Same as left, showing a similar trend for the parametric modulation by s1 duration during the maintenance of duration (z= 1.72, P< 0.10).

Cerebral Cortex 5

at University of C

hicago on May 14, 2014

http://cercor.oxfordjournals.org/D

ownloaded from

for internal representation of abstract quantity information,which may be derived for manifold perceptual dimensionsand/or sensory modalities (cf. Spitzer and Blankenburg2012).

Although the parametric modulations by intensity and dur-ation were spectrally and topographically similar, the intensityeffect showed a slightly more anterior prefrontal scalp distri-bution, and by tendency covered lower frequencies within theupper beta range (20–30 Hz). Due to limitations in the spatialprecision of noninvasive EEG recordings, we cannot conclus-ively infer whether the minor differences on the scalp levelmay represent different reflections of the same underlyingcortical process, or may reflect a similar encoding of the2 stimulus attributes in adjacent prefrontal areas, whichmay have been selectively engaged depending on the cue. Atleast suggestively, 3D source localization attributed thepresent prefrontal effects to the IFG in the lateral PFC, whichconforms very well with previous source localizations ofupper beta modulations during periodic frequency mainten-ance (Spitzer et al. 2010; Spitzer and Blankenburg 2011,2012) and with the homolog area found to exhibit quantitat-ive WM coding in monkeys’ lateral PFC (Romo et al. 1999;Barak et al. 2010).

A prominent difference between the oscillatory represen-tations of the 2 stimulus attributes was their differential time-course during cued retention, with the modulation by intensityarising considerably earlier than the modulation by duration.Notably, very similar time-courses of upper beta modulationswere observed in our previous retro-cue study of vibrotactilefrequencies, where a quantitative WM representation of moredifficult to-be-maintained information arose very early (atsimilar times as the present intensity effect), whereas moreeasily remembered information was parametrically representedonly after a delay (at similar times as the present durationeffect; cf. Fig. 3. in Spitzer and Blankenburg 2011). Thispattern is paralleled by our present findings, where s1–s2 dis-crimination of intensity appeared more difficult than discrimi-nation of the stimuli’s duration (cf. Behavioral Results). Wemay speculate that, in the retro-cue paradigm, subjects mayhave put greater emphasis a priori on the more difficult of the2 stimulus attributes (here: intensity), which may have facili-tated an early prefrontal representation of the to-be-maintainedinformation. To test this possibility in a post hoc analysis, weexamined the brain–behavior correlation between the earlycue-induced modulation by intensity (as identified in Fig. 2A,left) and the individual difference in behavioral A- versusB- task performance, which can be seen as a coarse indicatorof the extent to which individual subjects may have prioritizedintensity- over duration-encoding. Indeed, a positive corre-lation was found (RPearson = 0.51, P < 0.01), corroborating thatanticipatory prioritization may have contributed to the rela-tively early onset of the prefrontal representation of intensityinformation.

Independent from the stimulus-dependent modulations ofupper beta (20–30 Hz) in PFC, traditional beta-band power(15–25 Hz) in somatosensory areas was overall decreasingthroughout the retention interval (cf. Fig. 1A, lower and Sup-plementary Fig. 2A), in line with previous findings that tactileWM processing per se is accompanied by suppression of sen-sorimotor beta oscillations (e.g., Spitzer and Blankenburg2011, 2012). Here, the suppression in SI was found to be tran-siently reduced during the maintenance of duration, compared

with intensity (cf. Fig. 1B). Beta suppression in SI, in theabsence of stimulation, is typically related to tactile attention (e.g., Bauer et al. 2006; van Ede et al. 2010; Spitzer and Blanken-burg 2011). The presented pattern of SI activity thus suggests atemporary disengagement of tactile attention during retrospec-tive focusing on s1’s duration. Interestingly, this effect wasevident just before the quantitative representation of durationin the PFC was established (cf. Fig. 1B vs. Fig. 2B), corroborat-ing that quantitative evaluation of the stimuli’s duration in-volved more abstract, nonsensory processing. Aside from thesetask-dependent differences, however, like in our previousinvestigations of vibrotactile frequencies, we found no evi-dence for any delayed parametric representation of thetask-relevant quantitative stimulus attributes in early sensoryareas (for a review of similar findings, see Romo and Salinas2003; but see Harris et al. 2002; Wang et al. 2012). In reverse,aside from the specific stimulus-dependent modulations dis-cussed above, right prefrontal upper beta did not reflect overall(i.e., stimulus-independent) differences between the 2 tasks.Together, the present results converge with previous findingsthat stimulus-dependent prefrontal upper beta modulations arefunctionally dissociable from sensorimotor activity in the tra-ditional beta band (Spitzer and Blankenburg 2011, 2012).

Why and how should prefrontal upper beta temporarily in-crease as a function of the to-be-maintained quantitativestimulus attribute? One potential interpretation might be interms of increased cognitive demands, as might be inferredfrom decreased discriminability of greater sensory magni-tudes, according to the Weber–Fechner law (Weber 1834).However, such view of the present findings appears unlikely,given that s1–s2 discrimination performances declined withincreasing stimulus intensity only, but not with increasingduration (cf. Behavioral Results). Moreover, stimulus-dependent prefrontal modulations during s1 processing wereevident only on correct, but not on incorrect trials, in linewith recent compelling evidence against any direct relationbetween right prefrontal beta amplitudes and task demandsin similar WM tasks (Spitzer and Blankenburg 2011, 2012).An alternative interpretation may be that, in WM, greatersensory magnitudes might appear more salient. In this light,modulations of prefrontal upper beta could be seen as formof mnemonic bottom-up activity, with stimulus-dependentprocessing induced by the memory, rather than the actualpresence, of specific sensory input.

In terms of a more mechanistic account for the presentfindings, we propose that the prefrontal modulations in upperbeta power are more directly linked to the abstract quantityinformation, which is processed in WM. Parametric coding ofanalog sensory quantities has been extensively studied inmonkeys (for review, see Romo and Salinas 2003). In particu-lar, during WM maintenance in PFC, such coding was charac-terized by complex parametric modulations of single cells’firing rates and by dynamic changes of small populationstates over time (Romo et al. 1999; Barak et al. 2010). Thepresent oscillatory modulations in the human PFC may reflecta large-scale correlate of such a dynamically changing quanti-tative representation of the task-relevant WM contents (forrelated evidence, see Haegens et al. 2011; Spitzer and Blan-kenburg 2011). As such, modulations of prefrontal upper betamay reflect a population-level aspect of the neural code thatrepresents the task-relevant quantitative information in WM.We speculate that the parametric (i.e., continuous) nature of

6 WM Coding of Analog Stimulus Properties • Spitzer et al.

at University of C

hicago on May 14, 2014

http://cercor.oxfordjournals.org/D

ownloaded from

such coding corresponds to an abstract concept of analogquantity, which might be necessary to perform the sub-sequent quantitative comparison against s2 (for related dis-cussion, see Nieder and Merten 2007; Verguts 2007).

We recently proposed that parametric prefrontal modu-lations observed in human EEG may particularly indicate activeupdating of WM with the to-be-maintained stimulus attribute’squantitative value on an abstract internal scale (Spitzer andBlankenburg 2011, 2012). In light of such interpretation, anadditional factor may have contributed to the observation thatthe modulation by s1 duration occurred relatively late in theretention interval. Quantitative scaling of s1’s duration mayhave been achieved only after a mental replay of the entirestimulation period (length: 350–650 ms, see Stimuli and Behav-ioral Task). Thus, in addition to a potential reorientation to theless-attended stimulus attribute, the retrospective WM proces-sing, or imagery, of time (cf. Coull et al. 2008) may havefurther delayed the parametric prefrontal representation byduration, compared with the representation by s1’s stationaryintensity. In both memory tasks, however, like in the previousstudies of periodic frequency information (Spitzer and Blan-kenburg 2011, 2012), the parametric EEG modulations weretransient, suggesting that further maintenance of a single quan-tity value may not necessarily require continuous WM updatingthroughout the entire retention period.

In a broader context of previous EEG/MEG investigations ofhuman WM, mostly in the visual and auditory domains, “para-metric” WM processes have typically been assessed by varyingthe number of discrete, simultaneously to-be-maintained items(i.e., the WM “load”). Oscillatory WM effects associated withload manipulations included parametric modulations of alpha-(∼8–13 Hz; e.g., Jensen et al. 2002; Tuladhar et al. 2007), theta-(∼4–8 Hz; e.g., Jensen and Tesche 2002), and gamma-band am-plitudes (>30 Hz; e.g., Howard et al. 2003; Roux et al. 2012). Inparticular, delay activity in the theta- and gamma bands hasbeen related to the active maintenance of “sensory” represen-tations, with additional phase–amplitude interactions betweenthe 2 bands supporting the simultaneous representation ofmultiple items in WM (e.g., Jensen 2006; Sauseng et al. 2009;Axmacher et al. 2010). The present quantitative WM taskrequired the abstraction and maintenance of a single stimulusattribute only, and focal upper beta modulations in PFC selec-tively reflected the one task-relevant quantity in the currentfocus of attention (for similar results, see Spitzer and Blanken-burg 2011). Based on these basic findings, an avenue forfuture research can be the integration of analog quantity infor-mation on the level of multi-item WM, which may involve morecomplex interactions between different frequency bands, and/or brain areas.

To conclude, we demonstrated parametric WM processingin the human PFC for different quantitative stimulus attri-butes, reflected by systematic modulations of upper beta oscil-latory activity. The findings extend previous research on WMmaintenance of periodic frequencies and suggest a general-ized mechanism for the purposeful updating of abstract quan-tity information in human WM.

Supplementary MaterialSupplementary material can be found at: http://www.cercor.oxfordjournals.org/.

Funding

This research was supported by a grant from the GermanFederal Ministry of Education and Research (BMBF) to F.B.

NotesWe thank R. Auksztulewicz and 2 anonymous reviewers for helpfulcomments on the manuscript.

ReferencesAxmacher N, Henseler MM, Jensen O, Weinreich I, Elger CE, Fell J.

2010. Cross-frequency coupling supports multi-item workingmemory in the human hippocampus. Proc Natl Acad Sci USA.107:3228–3233.

Barak O, Tsodyks M, Romo R. 2010. Neuronal population coding ofparametric working memory. J Neurosci. 30:9424–9430.

Bauer M, Oostenveld R, Peeters M, Fries P. 2006. Tactile spatial atten-tion enhances gamma-band activity in somatosensory cortex andreduces low-frequency activity in parieto-occipital areas. J Neuro-sci. 26:490–501.

Benchenane K, Tiesinga PH, Battaglia FP. 2011. Oscillations in theprefrontal cortex: a gateway to memory and attention. Curr OpinNeurobiol. 21:475–485.

Brett M, Penny W, Kiebel S. 2003. An introduction to random fieldtheory. In: Frackowiak RSJ, Friston KJ, Frith CD, Dolan RJ, PriceCJ, Zeki S, Ashburner JT, Penny WD, editors. Human Brain Func-tion II. London (UK): Academic Press.

Bruns A. 2004. Fourier-, Hilbert- and wavelet-based signal analysis:are they really different approaches? J Neurosci Methods.137:321–332.

Coull JT, Nazarian B, Vidal F. 2008. Timing, storage, and comparisonof stimulus duration engage discrete anatomical components of aperceptual timing network. J Cogn Neurosci. 20:2185–2197.

D’Esposito M. 2007. From cognitive to neural models of workingmemory. Philos Trans R Soc Lond B Biol Sci. 362:761–772.

Eickhoff SB, Stephan KE, Mohlberg H, Grefkes C, Fink GR, AmuntsK, Zilles K. 2005. A new SPM toolbox for combining probabilisticcytoarchitectonic maps and functional imaging data. Neuroimage.25:1325–1335.

Fechner G. 1966. Elements of psychophysics. New York: Holt Rine-hart & Winston.

Friston K, Harrison L, Daunizeau J, Kiebel S, Phillips C, Trujillo-BarretoN, Henson R, Flandin G, Mattout J. 2008. Multiple sparse priors forthe M/EEG inverse problem. Neuroimage. 39:1104–1120.

Friston K, Henson R, Phillips C, Mattout J. 2006. Bayesian estimationof evoked and induced responses. Hum Brain Mapp. 27:722–735.

Gazzaley A, Nobre AC. 2012. Top-down modulation: bridging selec-tive attention and working memory. Trends Cogn Sci. 16:129–135.

Haegens S, Nácher V, Hernández A, Luna R, Jensen O, Romo R. 2011.Beta oscillations in the monkey sensorimotor network reflect so-matosensory decision making. Proc Natl Acad Sci USA.108:10708–10713.

Harris JA, Miniussi C, Harris IM, Diamond ME. 2002. Transientstorage of a tactile memory trace in primary somatosensory cortex.J Neurosci. 22:8720–8725.

Howard MW, Rizzuto DS, Caplan JB, Madsen JR, Lisman J,Aschenbrenner-Scheibe R, Schulze-Bonhage A, Kahana MJ. 2003.Gamma oscillations correlate with working memory load inhumans. Cereb Cortex. 13:1369–1374.

Ille N, Berg P, Scherg M. 2002. Artifact correction of the ongoing EEGusing spatial filters based on artifact and brain signal topogra-phies. J Clin Neurophysiol. 19:113–124.

Jensen O. 2006. Maintenance of multiple working memory items bytemporal segmentation. Neuroscience. 139:237–249.

Jensen O, Gelfand J, Kounios J, Lisman JE. 2002. Oscillations in thealpha band (9–12 Hz) increase with memory load during retentionin a short-term memory task. Cerebr Cortex. 12:877–882.

Cerebral Cortex 7

at University of C

hicago on May 14, 2014

http://cercor.oxfordjournals.org/D

ownloaded from

Jensen O, Tesche CD. 2002. Frontal theta activity in humans increaseswith memory load in a working memory task. Eur J Neurosci.15:1395–1399.

Kiebel SJ, Tallon-Baudry C, Friston KJ. 2005. Parametric analysis ofoscillatory activity as measured with EEG/MEG. Hum Brain Mapp.26:170–177.

Kilner JM, Kiebel SJ, Friston KJ. 2005. Applications of random fieldtheory to electrophysiology. Neurosci Lett. 374:174–178.

Litvak V, Friston K. 2008. Electromagnetic source reconstruction forgroup studies. Neuroimage. 42:1490–1498.

Litvak V, Mattout J, Kiebel S, Phillips C, Henson R, Kilner J, Barnes G,Oostenveld R, Daunizeau J, Flandin G et al. 2011. EEG and MEGdata analysis in SPM8. Comput Intell Neurosci. doi:10.1155/2011/852961.

Miller BT, D’Esposito M. 2005. Searching for “the top” in top-downcontrol. Neuron. 48:535–538.

Nee DE, Brown JW, Askren MK, Berman MG, Demiralp E, Krawitz A,Jonides J. 2012. A meta-analysis of executive components ofworking memory. Cereb Cortex. 23:264–282.

Nieder A, Merten K. 2007. A labeled-line code for small and large numer-osities in the monkey prefrontal cortex. J Neurosci. 27:5986–5993.

Pasternak T, Greenlee MW. 2005. Working memory in primatesensory systems. Nat Rev Neurosci. 6:97–107.

Phillips C, Mattout J, Friston K. 2007. Forward models for EEG. In:Friston K, Ashburner J, Kiebel S, Nichols T, Penny W, editors. Stat-istical parametric mapping: the analysis of functional brainimages. Amsterdam, Boston: Elsevier/Academic Press.

Romo R, Brody CD, Hernández A, Lemus L. 1999. Neuronal correlatesof parametric working memory in the prefrontal cortex. Nature.399:470–473.

Romo R, Salinas E. 2003. Flutter discrimination: neural codes, percep-tion, memory and decision making. Nat Rev Neurosci. 4:203–218.

Roux F, Wibral M, Mohr HM, Singer W, Uhlhaas PJ. 2012. Gamma-band activity in human prefrontal cortex codes for the number ofrelevant items maintained in working memory. J Neurosci.32:12411–12420.

Sauseng P, Klimesch W, Heise KF, Gruber WR, Holz E, Karim AA,Glennon M, Gerloff C, Birbaumer N, Hummel FC. 2009. Brain

oscillatory substrates of visual short-term memory capacity. CurrBiol. 19:1846–1852.

Spitzer B, Blankenburg F. 2011. Stimulus-dependent EEG activity re-flects internal updating of tactile working memory in humans.Proc Natl Acad Sci USA. 108:8444–8449.

Spitzer B, Blankenburg F. 2012. Supramodal parametric workingmemory processing in humans. J Neurosci. 32:3287–3295.

Spitzer B, Wacker E, Blankenburg F. 2010. Oscillatory correlates ofvibrotactile frequency processing in human working memory. JNeurosci. 30:4496–4502.

Tallon-Baudry C, Bertrand O, Peronnet F, Pernier J. 1998. Inducedγ-band activity during the delay of a visual short-term memorytask in humans. J Neurosci. 18:4244–4254.

Tuladhar AM, Ter Huurne N, Schoffelen J-M, Maris E, Oostenveld R,Jensen O. 2007. Parieto-occipital sources account for the increasein alpha activity with working memory load. Hum Brain Mapp.28:785–792.

van Ede F, Jensen O, Maris E. 2010. Tactile expectation modulatespre-stimulus beta-band oscillations in human sensorimotor cortex.Neuroimage. 51:867–876.

Verguts T. 2007. How to compare two quantities? A computationalmodel of flutter discrimination. J Cogn Neurosci. 19:409–419.

Voytek B, Davis M, Yago E, Barceló F, Vogel EK, Knight RT. 2010.Dynamic neuroplasticity after human prefrontal cortex damage.Neuron. 68:401–408.

Wager TD, Smith EE. 2003. Neuroimaging studies of workingmemory: a meta-analysis. Cogn Affect Behav Neurosci. 3:255–274.

Wang L, Li X, Hsiao SS, Bodner M, Lenz F, Zhou Y-D. 2012. Persistentneuronal firing in primary somatosensory cortex in the absenceof working memory of trial-specific features of the samplestimuli in a haptic working memory task. J Cogn Neurosci.24:664–676.

Weber EH. 1834. De Pulsu, resorptione, auditu et tactu: Annotationesanatomicae et physiologicae. Leipzig: C.F. Koehler.

Worsley KJ, Marrett S, Neelin P, Vandal AC, Friston KJ, Evans AC.1996. A unified statistical approach for determining significantsignals in images of cerebral activation. Hum Brain Mapp.4:58–73.

8 WM Coding of Analog Stimulus Properties • Spitzer et al.

at University of C

hicago on May 14, 2014

http://cercor.oxfordjournals.org/D

ownloaded from