Maintenance and manipulation of somatosensory information in ventrolateral prefrontal cortex

Abstract

Neuroimaging studies of working memory (WM) suggest that prefrontal cortex may assist sustained maintenance, but also internal manipulation, of stimulus representations in lower‐level areas. A different line of research in the somatosensory domain indicates that neuronal activity in ventrolateral prefrontal cortex (VLPFC) may also represent specific memory contents in itself, however leaving open to what extent top‐down control on lower‐level areas is exerted, or how internal manipulation processes are implemented. We used functional imaging and connectivity analysis to study static maintenance and internal manipulation of tactile working memory contents after physically identical stimulation conditions, in human subjects. While both tasks recruited similar subareas in the inferior frontal gyrus (IFG) in VLPFC, static maintenance of the tactile information was additionally characterized by increased functional coupling between IFG and primary somatosensory cortex. Independently, during internal manipulation, a quantitative representation of the task‐relevant information was evident in IFG itself, even in the absence of physical stimulation. Together, these findings demonstrate the functional diversity of activity within VLPFC according to different working memory demands, and underline the role of IFG as a core region in sensory WM processing. Hum Brain Mapp 35:2412–2423, 2014. © 2013 Wiley Periodicals, Inc.

INTRODUCTION

Working memory (WM) refers to the temporary internal representation of information that no longer exists in the environment. Such representations can be subject to goal‐directed top‐down control, accomplishing sustained maintenance of the short‐lived information over longer periods of time, but also purposeful manipulation of the temporary memory content to support adaptive behavior [Baddeley, 1997; D'Esposito, 2007]. Across different WM tasks, materials, and sensory domains, central WM processes have been attributed to lateral prefrontal cortex (PFC), but the precise functional contribution of PFC delay activity to WM remains a matter of debate.

It has been proposed that lateral PFC may be organized in material‐specific manner, with maintenance of nonspatial information related in particular to ventral areas [e.g., Courtney, 2004; Levy and Goldman‐Rakic, 2000]. Other authors have emphasized a process‐specific organization, where ventral areas serve basic WM encoding and retrieval, while more dorsal regions support higher‐level monitoring and manipulation processes [e.g., Fletcher and Henson, 2001; Petrides et al., 1993]. Across studies, diverse results were obtained regarding the precise regional allocation of different WM functions to specific PFC subareas [e.g., D'Esposito et al., 1999; Glahn et al., 2002; Mohr et al., 2006; Roth et al., 2006; Veltman et al., 2003].

Another theoretical approach suggests that persistent WM activity in PFC may not directly reflect the storage of information, but rather relates to top‐down mechanisms that assist WM processing of sensory representations in earlier, postcentral cortical areas [e.g., Curtis and D'Esposito, 2003; Miller and D'Esposito, 2005]. Using functional neuroimaging and connectivity analyses in humans, the notion of a crosstalk between lateral PFC and lower‐level sensory areas during WM maintenance is receiving accumulating empirical support from studies in the visual and verbal domains [e.g., Fiebach et al., 2006; Rissman et al., 2004; Vuilleumier and Driver, 2007; Yoon et al., 2006; Zanto et al., 2011].

Complementary insights into prefrontal contributions to WM have been gained in the somatosensory domain, using vibrotactile stimuli. Thereby, electrophysiological recordings in monkeys [for review, see Romo and Salinas, 2003] and noninvasive studies in humans [e.g., fMRI: Kostopoulos et al., 2007; Preuschhof et al., 2006; EEG: Spitzer et al., 2010; TMS: Auksztulewicz et al., 2011] converge on a central involvement of ventrolateral prefrontal cortex (VLPFC) during WM processing of previous tactile information. In particular, as a seminal finding in vibrotactile WM tasks, delay activity in VLPFC directly reflected the task‐relevant stimulus attribute [e.g., Romo et al., 1999; Spitzer and Blankenburg, 2011] and can thus be regarded as a substrate of WM processing per se [Goldman‐Rakic, 1995]. These findings may establish the somatosensory modality as a very useful model for the investigation of basic WM operations, yet little is known about potential interactions between lateral PFC and lower‐level sensory areas, or how internal manipulation processes may be implemented, also in the domain of tactile memories.

Here, we introduce a paradigm demanding either static maintenance or continuous internal modification of previously presented tactile stimulus information throughout a retention interval, given physically identical stimulation conditions (cf., Fig. 1). Using fMRI, we sought to replicate previous activity patterns in VLPFC during tactile WM maintenance, and to further identify activity specifically related to internal WM manipulation. Moreover, we employed functional connectivity analysis and parametric modeling of the to‐be‐processed stimulus quantity, to examine the extent to which lateral PFC may exert top‐down control over lower‐level areas and/or engage in direct processing of the WM contents, during the different WM tasks.

Figure 1.

Tactile stimulation protocol and cognitive tasks. Left: Visual icons indicating the task condition for upcoming trials: maintenance (“pause”, upper), manipulation (“play”, middle), and control (lower). A: Schematic of tactile stimulation applied to the left index finger, for exemplary trials in the different task conditions. A high‐frequency carrier signal (100 Hz, for display purpose 10 Hz are plotted) was amplitude‐modulated by a 0.1 Hz sine function. The first stimulus segment (stim1, purple) was identical across conditions. In the “pause” task (red), subjects had to statically maintain the stop intensity of the first stimulus segment (dashed red) during the stimulus‐free delay. The “play” task (blue) required to mentally continue the systematic intensity fluctuation (dashed blue) of the first stimulus segment throughout the delay. In the “control” task, subjects were asked whether the second stimulus segment was in a rising phase or not, requiring no WM processing during the delay. Additional “catch” trials warranted that both stimulus segments were attentively encoded in the control task (see text for details). Examples for the second stimulus segment (stim2, right) are illustrated separately for the two delay period lengths (saturated colors: 5 s delay; desaturated colors: 2.5 s delay). The phase of the second stimulus segment varied according to task condition: The waves in red exemplify stimulus continuations that would be “correct” in the pause task, but “incorrect” in the play task. The waves in blue show the correct continuations in the play task. B: In all conditions, the starting phase of the entire stimulus sequence varied randomly from trial to trial. Shown are example trials where the starting phase is successively shifted in steps of π/2. See Materials and Methods for further details.

MATERIALS AND METHODS

Subjects

Fifteen healthy, right‐handed volunteers (22–29 years, 9 females, 6 males) participated in the experiment with written informed consent. The data of two participants (27 and 26 years, both male) were discarded due to technical problems. The study was approved by the Ethics Committee of the Charité University Hospital Berlin and corresponded to the Human Subjects Guidelines of the Declaration of Helsinki.

Stimuli and Cognitive Task

Tactile stimulation of the left index finger was delivered by a 16‐dot piezo electric Braille display (4x4 quadratic matrix, 2.5 mm spacing) controlled by a programmable stimulator (Piezostimulator, QuaeroSys, St. Johann, Germany). The pins of the Braille display were simultaneously driven by a constant 100 Hz sinusoidal carrier signal which was amplitude‐modulated by a 0.1 Hz sine function (Fig. 1A). The resulting stimulation was perceived as an ongoing vibration of periodically changing intensity that describes the cyclic shape of a sine wave.

A trial consisted of a 7.5 s segment of stimulation, which was followed by a stimulus‐free delay period of variable length (5 s or 2.5 s, randomly varied), then another 2.5 s segment of stimulation was presented (cf. Fig. 1A). The phase of the first stimulus segment was randomly shifted in steps of π/2 (i.e., ¼ cycle, corresponding to 2.5 s of the 0.1 Hz sine function) from trial to trial (Fig. 1A,B). The phase of the second stimulus segment differed according to experimental condition.

In the pause condition (maintenance task) the second stimulus segment started at the phase at which the first stimulus segment had stopped before the delay (Fig. 1, red). In this condition, participants were instructed to memorize the “position” (i.e., the phase) at which the first stimulus segment had stopped, and to decide whether the second stimulus segment corresponded to the continuation of the first segment (i.e., as if the stimulus cycle had paused during the delay; note that a simplified analysis of this task was recently used as a functional localizer in a related TMS study, cf. Auksztulewicz et al., 2011). In the play condition (manipulation task), the second stimulus segment started at the phase the first stimulus would have reached if the periodic stimulation had continued during the delay period (Fig. 1, blue). On these trials, subjects' task was to mentally continue the progression of the previous stimulus segment throughout the delay period, and to decide whether the second stimulus segment corresponded to the continuation of the extrapolated segment (i.e., whether the stimulus cycle had continued to “play” during the delay). Due to the variable length of the delay period, this task could not be solved on the basis of predictions derived from the first stimulus segment alone, but required continuous ‘online’ processing throughout the delay (cf. Fig. 1A). By shifting the phase of the entire stimulus sequences in steps of 2.5 s across trials (cf. Fig. 1A,B), it was warranted that the mean to‐be‐processed intensity in each task‐ and delay condition was identical, and constant over time (Note that for each of the functions in Figure 1, the average of the four phase‐shifted versions across trials is a flat line at 50% of the intensity range). For half of the trials in the play and the pause conditions, ‘incorrect’ stimulus sequences were devised by randomly shifting the phase of the second stimulus segment in steps of ± π/2 (i.e. ± ¼ cycles) relative to the ‘correct’ second segments shown in Figure 1.

In a third condition (perceptual control), the second stimulus segment was randomly chosen from any of the above conditions and subjects' task was to judge, independently from the first segment, whether the vibration intensity during the second segment was increasing (i.e., in a rising phase) or not (i.e., in a falling phase). Because this control task did not require attention to the first stimulus segment, additional “catch” trials were included in which subjects were asked immediately after the first stimulus segment whether this segment had stopped in a rising or falling phase. In this way, it was warranted that subjects attentively encoded both stimulus segments also in the control condition. Prior to each subject's scanning, several test runs in‐ and outside the scanner room were performed to familiarize the subject with the setup and to confirm proper mechanical functioning of the Braille stimulation device.

Design and Procedure

The experiment consisted of three sessions. Each session consisted of twelve blocks, and each block included four trials of the same task condition, resulting in a total of 144 trials per participant. Blocks of different conditions were equally distributed across sessions, and the order of the blocks within sessions was pseudo‐randomized with the restriction that no task repetitions occurred. In addition, the order of trials within each block was pseudo‐randomized with the constraint that each block contained two trials with a short delay period (2.5 s) and two trials with a long delay period (5 s). All possible combinations of the two delay periods with the four different starting phases of the first stimulus segment were used equally often in each task. “Correct” and “incorrect” continuations of the second stimulus segment were randomly distributed across trials. Note that on average, the presented stimulus segments were physically identical for all task conditions and delay period lengths (cf., Fig. 1A,B). Catch trials were presented in 12.5% of the perceptual control trials. These trials were pseudo‐randomly interspersed so that on average they occurred once in every second control task block, and were not included in the analysis.

Before each block, a visual icon (illustrated in Fig. 1, upper left) was presented for 3.5 s, which signaled the task condition for the forthcoming trials. After each block, a black screen was displayed for 16 s. Each trial started with the presentation of a white fixation cross centered on a black screen. After 3 s, the fixation cross disappeared briefly (0.25 s) and 0.5 s later the tactile stimulus sequence started as exemplified in Figure 1. The fixation cross remained on the screen until the end of the sequence. Subsequently, a question mark appeared and participants had to enter a response by pressing one of two buttons on an optical key pad using the right index or middle finger, respectively. The assignment of the two buttons to “yes”‐ and “no”‐responses was counterbalanced across subjects. On catch trials in the perceptual control condition, the question mark was presented immediately after the first stimulus segment. After a 2‐s response interval, the question mark disappeared and performance feedback was given by presenting either a plus or a minus sign (indicating whether the response was correct or not). Feedback was provided in order to promote subjects' motivation and correct realization of the respective cue conditions throughout the experiment. Because in all tasks the physical stimulus input was systematically varied from trial to trial (cf., Fig. 1), additional effects of the feedback per se on task performance (e.g., in terms of low‐level stimulus‐response reinforcement), were assumed to be negligible. After 0.75 s, the feedback signs disappeared and the next trial started.

fMRI Methods

Imaging was performed using a 1.5 T Magnetom Sonata MRI system (Siemens, Erlangen, Germany) with a standard head coil. In order to minimize head motion, participants' head was immobilized by means of a cushion. BOLD sensitive images were acquired using a T2*‐weighted echo‐planar imaging (EPI) sequence (TR = 2010 ms, TE = 40 ms, flip angle = 90°, matrix = 64×64, voxel size = 3×3×3 mm³, 36 slices, interleaved acquisition). For each participant, three successive runs of 626 volumes each were acquired.

The data were analyzed using SPM8 (Wellcome Department of Cognitive Neurology, London, UK). The first five scans in each session were discarded in order to allow longitudinal magnetization to reach equilibrium. Standard preprocessing steps were initially performed: images were realigned to the first image of the time series and corrected for movement‐induced image distortions (6‐parameter rigid body affine realignment). Then, images were normalized to the Montreal Neurological Institute (MNI) EPI‐template, resampled into 2 x 2 x 2 mm³ sized voxels, and spatially smoothed with a three‐dimensional Gaussian filter of 8 mm full‐width at half‐maximum (FWHM).

Conventional GLM Analysis

Voxelwise statistical analyses were calculated using the General Linear Model (GLM). Twelve regressors were included: (1–9) Boxcar regressors separately specifying the times of the first stimulus period, the delay period, and the second stimulus period, for each of the three task conditions (pause, play, and control), (10) compound regressor for error trials and catch trials, (11) visual cues pooled over conditions, and (12) button presses for all trials. To model task unrelated noise, six additional nuisance regressors of no interest were included, corresponding to the first eigenvariates extracted from any activation outside of the brain mask (thresholded at P < 0.05) (for a similar approach, see e.g. Wacker et al., 2011; Weissenbacher et al., 2009). The first eigenvariate thereby represents the principal component time series that accounts for most variance in the respective noise activity.

Predictors of the hemodynamic response were generated by convolution with a canonical hemodynamic response function (HRF). A temporal high pass filter with a period of 300 s was used in order to remove low‐frequency drifts. Contrast images of interest were calculated from the corresponding parameter estimates. Group analysis at the second level was then performed by entering the individual subjects' contrast images into a 3 x 3 within‐subjects ANOVA with the factors task (pause, play, and control) and trial phase (first stimulus segment, delay period, second stimulus segment). This allowed us to compute the effects of tactile stimulation, as well as differential effects between the different conditions for each trial phase, using contrast vectors to produce Statistical Parametric Maps (SPMs). The resulting SPMs were thresholded at P < 0.001 (uncorrected; voxel‐level) and clusters surpassing a threshold of P < 0.05 (FWE corrected; cluster‐level) were considered as significantly activated. The minimum cluster size (p<0.05, FWE) in the whole‐brain analysis was 176 voxels, at an estimated smoothness of 11x11x11 mm, FWHM. For illustrating purposes, functional activation maps thresholded at P < 0.005 (uncorrected) were superimposed on the anatomical MNI template as provided in SPM.

PPI Analysis

For analysis of hypothesized changes in connectivity between PFC and lower‐level sensory areas during tactile WM processing, we analyzed psycho‐physiological interactions [PPI; Friston et al., 1997] associated with the two memory tasks during the delay period. Within activated areas in PFC, as identified in the above GLM analysis (cf. Fig. 2B), spheres with a radius of 5 mm around the individual contrast maxima were defined as seed regions for extracting the first eigenvariate of the signal and creating the psychophysiological interaction term for the respective delay‐period contrast of interest (pause>control and play> control). These terms were then entered as regressors into the individual subject GLMs described above, and the resulting PPI contrast images were subjected to group‐level analysis using one‐sample t‐tests.

Figure 2.

A: Overall activation during the first stimulus segment (purple wave in Fig. 1) contrasted against Baseline. Significant clusters (P < 0.05, whole brain FWE) include contralateral SI, bilateral SII, bilateral IFG, and SMA. B: WM‐induced activations during the stimulus‐free delay period (all clusters P < 0.05, whole brain FWE). Upper: Static WM maintenance (pause task) activated the right IFG (BA 44). Middle: Dynamic WM manipulation showed similar activation in right IFG and additionally recruited the SMA. Additional less significant activations in left IPC were observed in both tasks (cf. Table 1; not visible in figure renderings). Lower: In direct comparison, dynamic manipulation (play) additionally recruited the SMA, and inferior subregions of IFG (BA 44/45). C: Psycho‐Physiological Interactions (PPI) with WM‐related delay activity in IFG. Left: Increased coupling between IFG and SI during static maintenance (pause task; P < 0.001, uncorrected; P < 0.05 FWE‐corrected for SI/SII search volume). Right: Spurious IFG‐SI coupling during dynamic WM manipulation (play task, P < 0.005, uncorrected).

Parametric Modeling of the to‐be‐Processed Stimulus Dynamics

To examine brain dynamics that may directly reflect processing of the continuously changing stimulus intensity (cf., Fig. 1), we modeled the trial‐specific envelope of the stimulus function (cf., cartoons in Fig. 3, orange lines) as parametric modulations of the conventional GLM's boxcar stimulus functions. Thereby, the stimulus‐free delay periods for all task conditions were modeled as the extrapolation of the first stimulus segment as was to‐be‐processed in the play task (dashed blue in Fig. 1). The envelope functions were zero‐centered, normalized, and convolved with the HRF. Note that the resulting continuous predictor of the hemodynamic response is only sparsely sampled by the fMRI time‐series at ∼0.5 Hz (TR =2.01 s, see fMRI methods), which however provides sufficient temporal resolution to sample the trial‐specific cyclic phase of the 0.1 Hz stimulus function (by Nyquist‐Shannon sampling theorem). Because across trials the stimulus sequence was temporally decorrelated (jittered) from image acquisition, sparse sampling was evenly distributed over the trial epoch. The continuous stimulus functions were added separately for each task condition (pause, play, and control) and task period (first stimulus segment, delay period, and second stimulus segment) in the design matrix, resulting in nine additional regressors. As an orthogonal extension to the conventional GLMs boxcar stimulus functions (which are identical across trials of the same length), the parametric regressors model residual variance attributable to the trial‐specific cyclic phase of the tactile stimulus, and thus isolate fMRI responses that covary with the slow fluctuations of the stimuli's intensity over time (cf. Fig. 1). Inspection of the individual design matrices in SPM confirmed that collinearities between the parametric and the conventional GLM regressors were low, on average ranging between 0.01 and 0.02 (0: orthogonal; 1: collinear) across subjects. Parametric contrast images were derived and subjected to group‐level analysis using a 3 x 3 (task condition x task period) within‐subjects design analogous to the conventional GLM group analysis described above.

Figure 3.

Parametric modulations by instantaneous stimulus intensity. A: Intensity changes during periods of physical stimulation (cartoon, yellow curve) parametrically modulate activity in SI (P < 0.05, whole brain FWE). B: Differential parametric modulation in periods/conditions requiring active online processing of instantaneous intensity, compared with the modulation in periods/conditions of passive, or no processing (cf., cartoon inset; see text for details) is found in IFG (P < 0.001, uncorrected; P < 0.05 FWE‐corrected for IFG search volume). Bar graph on bottom shows normalized contrast estimates for the parametric modulation in IFG, separately for each trial period and task condition. Error bars show 90% confidence intervals. Increased parametric modulation during active processing (play task) is in particular evident in the delay period, in which no physical stimulation was present. Note: For illustration purpose, display threshold is P < 0.005 (uncorrected) in all MRI renderings.

All reported coordinates correspond to the anatomical MNI space. The SPM anatomy toolbox [Eickhoff et al., 2005] was used to establish cytoarchitectonic reference where possible.

RESULTS

Behavioral Performance

The mean proportion of correct responses, collapsed across delay lengths, was 84.5% (SE = 0.013) in the pause task (maintenance), and 78.2% (SE = 0.024) in the play task (manipulation). A 2 x 2 repeated measures analysis of variance (ANOVA) with the factors of task (play vs. pause) and delay length (2.5 vs. 5 s) showed a main effect of task [F(1,12) = 7.26, P < 0.05], with higher performance in the pause‐ compared to the play task. No significant main effect of delay length [short: 83.3, long: 79.4%; F(1,12) =2.62, P > 0.10], and no interaction between the two factors was found [task x length; F(1,12)<1; P > 0.80]. Mean performance in the perceptual control task (83.7% correct; SE = 0.015) was not significantly different from the performance levels in either of the two WM tasks [both t(12)'s < 2; both Ps > 0.05] and‐expectedly‐did not differ significantly between the two delay period lengths [t(12) = 1.80, P > 0.05].

An additional 3x2 ANOVA, contrasting the performance levels in the first compared with the second half of each scanning session, indicated that performance levels remained stable over the time of the session [F(1,12) <1, P > 0.50], in all three task conditions [interaction task x time: F(2,24)<1 P > 0.90]. Further, no significant differences between the performance levels in the first compared with the last scanning session were found [all F's<2, P > 0.10].

Conventional fMRI Analysis

To identify brain areas involved in tactile stimulus processing, activation during the first stimulus segment (cf. Fig. 1, violet wave), pooled over conditions, was contrasted against the session mean (Fig. 2A). In line with previous findings [e.g., Blankenburg et al., 2006; Pleger et al., 2006], this contrast showed increased activation in contralateral primary somatosensory cortex (SI), bilateral secondary somatosensory cortex/parietal operculum (SII), the inferior frontal gyrus (IFG) bilaterally, and supplementary motor area (SMA, cf. Table 1). Importantly, differential contrasts between the play, pause, and control tasks showed no significant differences in activation during the first stimulus segment (all Ps > 0.005 uncorrected). Thus, it appears very unlikely that subsequent contrasts between the cognitive tasks in the delay period, after physically identical stimulation, might be attributable to a carry‐over of differential effects from the encoding period.

Table 1.

Cluster peak MNI coordinates of activations identified by conventional GLM analysis (cf., Fig. 2A,B)

Contrast	Region	Hemisphere	X	Y	Z	t‐value
Stimulus overall > BL	Primary Somatosensory Cortex	R	58	−14	48	5.52
	Secondary Somatosensory Cortex	R	54	−22	20	6.44
		L	−60	−20	20	5.86
	Inferior Frontal Gyrus	R	52	10	18	5.23
		L	−56	4	20	4.82
	Supplementary Motor Area	R/L	4	4	62	4.47
Delay pause > control	Inferior Frontal Gyrus	R	52	6	22	5.53
	Inferior Parietal Cortex	L	−52	−36	50	4.46
Delay play > control	Inferior Frontal Gyrus	R	54	8	20	6.88
	Supplementary Motor Area	R/L	4	6	66	5.49
	Inferior Parietal Cortex	L	−46	−42	48	5.58
Delay play > pause	Inferior Frontal Gyrus	R	50	12	0	4.63
	Supplementary Motor Area	R/L	−2	0	68	4.35

Next, we examined activity associated with stimulus maintenance by contrasting the delay period activity between the pause condition and the control condition (Fig. 2B upper). Significant activation was found in right IFG and‐to a lesser extent‐in left inferior parietal cortex (IPC, cf. Table 1). Analogously, delay period activity was contrasted between the play condition and the control condition (Fig. 2B middle). Online manipulation of the stimulus intensity recruited similar areas as maintenance, including strong activation of right IFG, and additionally engaged more inferior subregions of right IFG, as well as the SMA (cf. Table 1). This was confirmed by directly contrasting the play against the pause condition during the delay period (Fig. 2B, lower), showing increased activity in inferior subregions of right IFG, and in SMA, during the play task (cf. Table 1).

Note that short (2.5 s) and long (5 s) delay trials were pooled in the GLM analysis (cf. Materials and Methods). Supplementary control analysis indicated no significant differences in any of the above reported contrasts between the two delay period lengths (Supporting Information Fig. 1). The clusters of WM‐related activation reported in the conventional fMRI analysis (Fig. 2B) exceeded a whole‐brain significance level of P < 0.05 (FWE) and were located within the network of areas illustrated in Figure 2A. The subsequent connectivity‐ and parametric analyses were performed within this overall functional network.

Connectivity Analysis

For analysis of a potential coupling between IFG and earlier sensory areas during WM processing, we examined Psycho physiological Interactions (PPIs) using the areas of activity in IFG identified in the above analysis as seed regions (see Materials and Methods, PPI analysis). The region of interest for potential somatosensory target areas was defined by the extensive activation clusters including right SI and bilateral SII as identified in Figure 2A (thresholded at P < 0.005, uncorrected). Inspecting areas of increased coupling in the pause task compared to the control task, during the delay period (Fig. 2C, left), we found increased coupling with right SI (Brodmann area 1, x = 50, y = −14, z = 50; P < 0.05, FWE‐corrected for search region).

An analogous analysis of increased IFG coupling during the play task compared with the control task showed no significant result (all voxels P > 0.001, uncorrected). Only below the statistical threshold (Fig. 2C, right; display threshold P < 0.005), a tendency for coupling with SI became evident (x = 50, y = −16, z = 48). In direct comparison, for the peak voxel of the pause>control PPI contrast (Fig. 2C, left), which was also identified in the less significant play>control contrast (Fig. 2C right), the coupling with IFG was stronger during the pause task than during the play task (P < 0.05). We report this trend for completeness although the significance level was not adjusted for a whole brain analysis. Together, the PPI analysis demonstrates increased coupling between IFG and early somatosensory cortex (SI) during static WM maintenance (pause task), whereas only a slight indication of such coupling was evident during dynamic manipulation of the WM content (play task).

Parametric Modulations by the to‐be‐Processed WM Content

To examine brain dynamics that may directly reflect the task‐relevant WM content (i.e., the intensity changes over time, cf. Fig. 1), the trial‐specific envelopes of the cyclic intensity function (cf., cartoons in Fig. 3, orange lines) were modeled by continuous parametric regressors in the fMRI analysis (see Materials and Methods, Parametric modeling). As an orthogonal extension of the conventional GLM contrasts reported above (cf. Fig. 2A,B), the parametric analysis yields nonredundant information on a potential quantitative representation of the to‐be‐processed intensity values. We first inspected parametric modulations during the task phases in which stimulation was physically present compared to baseline (Fig. 3A). This contrast, collapsed across the different task conditions and across both stimulation periods, showed that the ongoing changes of physical stimulus intensity were reflected very reliably by parametric modulations of activity in SI (x = 42, y = −32, z = 64, P < 0.05, whole‐brain FWE).

In the next step, we sought to identify areas that may specifically reflect active compared with passive processing of the dynamic stimulus changes, according to task instructions and task period (see cartoon in Fig. 3B): It can be inferred that in the play task, active online processing of the cyclic stimulus dynamics was required during the first stimulus segment (as the basis for subsequent extrapolation), during the last stimulus segment (for verification of the manipulation result), and crucially, also during the delay period, in which no physical stimulus was present. In the pause task, in contrast, continuous online processing of the ongoing stimulus changes was encouraged in neither task phase, as only the stop intensity of the first segment had to be maintained for comparison against the start intensity of the second segment. Finally, in the control task, no active WM processing was required during the delay, but an evaluation of the intensity change throughout the last stimulus segment was necessary to perform the task. The first stimulus segment on control trials cannot be unequivocally classified according to the above scheme, because the task on catch trials (see Stimuli and Cognitive Task) required active change evaluation for the later portion of the first segment, but discouraged active processing of the longer preceding part.

The overall contrast between the parametric modulations during conditions and periods of “active” compared with “passive” processing (Fig. 3B, right), revealed increased parametric processing in IFG (Fig. 3B, right; x = 48, y = 14, z = 22, P < 0.001, uncorrected). Restricting this analysis to the areas of WM‐related activity in PFC as identified above (ROI defined by conjunction of Fig. 2B middle and lower, thresholded at P < 0.005, uncorrected) due to a strong a‐priori hypothesis, the increase of parametric processing in IFG exceeded FWE‐corrected significance (cluster‐level: P = 0.029, FWE; peak‐level: P = 0.018, FWE). An analogous parametric analysis, but modeling the delay period activity by the trial‐specific intensity levels that were to‐be‐maintained in the pause task (cf. dotted red in Fig. 1), yielded no significant results (all voxels P > 0.005, uncorrected). This negative finding may be attributable to reduced statistical power, because potential areas of active online processing of the static intensity values could have been inferred from the (pause) delay period only (cf. Fig. 1, see Discussion). In contrast, all task periods could be exploited for the presented analysis of dynamic intensity changes (cf. Fig. 3B).

While the above analysis indicates a role of IFG during active online processing of the stimulus dynamics in the play task, the question arises whether such parametric representation was driven mainly during actual stimulation, or may also be evident during merely internal processing of the dynamic WM content during the delay. The bar graph in Figure 3B (bottom) illustrates normalized contrast estimates for the parametric modulation in IFG, for each task period and condition. Evidently, while the parametric modulations by tendency exhibited the expected pattern (active conditions > passive conditions) across all three task phases, this pattern appeared most clearly during the delay period, with significantly enhanced modulation in the play task, compared with the remaining tasks (P < 0.005). This finding corroborates that IFG was directly engaged in internal computation of the dynamic stimulus information processed in working memory. It must be acknowledged that due to the limited temporal resolution of fMRI, spurious carry‐over between the similar effects observed in the different task periods cannot be fully ruled out in this analysis. However, the observation that the modulation enhancement was by tendency strongest in the delay period renders it unlikely that this effect was driven during the stimulation periods only. We note for completeness that across subjects, the delayed increase of parametric IFG modulation in the play task (cf. Fig. 3B; bottom, delay period) was correlated with individual behavioral performance in the play task (Spearmans Rho = 0.55), just below significance (P = 0.055). No evidence for correlations with performance in the pause or control tasks was found (both Rho's < 0.10).

DISCUSSION

Using fMRI, we examined brain activity during static WM maintenance of previously presented tactile input, and during dynamic internal manipulation of the to‐be‐maintained information. Delay processing in both tasks was primarily characterized by increased activity in overlapping areas in right VLPFC/IFG, with dynamic stimulus extrapolation recruiting additional IFG subregions. Static stimulus maintenance was further accompanied by increased functional coupling between IFG and SI, whereas such effect was barely evident during dynamic internal manipulation of the WM contents. Finally, a direct representation of the to‐be‐processed quantitative information in IFG was found, in terms of parametrically modulated delay activity, under conditions of continuous internal updating of the task‐relevant WM content.

The principled finding of VLPFC/IFG engagement during tactile WM processing was well expected on basis of a large body of related previous work [e.g., Fiehler et al., 2011; Kaas et al., 2007, 2013; Kostopoulos et al., 2007; Preuschhof et al., 2006; Romo et al., 1999; Spitzer et al., 2010]. Beyond this finding, however, conventional GLM analysis yielded no evidence for additional activation of more dorsal prefrontal areas during manipulation processing, which seems to conflict with the traditional view of a process‐specific allocation of maintenance/manipulation processes to ventral/dorsal subregions of lateral PFC [e.g., Fletcher and Henson, 2001; Petrides et al., 1993]. Compared to the pause task (maintenance), the play task (manipulation) was even characterized by additional activity in more inferior IFG subregions (but also in SMA, cf. Fig. 2B, lower). However, we are cautious in attributing this regional finding directly to the manipulation aspect of the task. The play task may have been generally more WM demanding (cf., behavioral results), and overall IFG activation may have been stronger (cf., Table 1), with additional neighboring areas surpassing the analysis threshold. Moreover, manipulation in the play task was based on the speed of the stimulus dynamics, thereby involving also processing of time. Notably, previous fMRI studies indicate a central role of inferior VLPFC (i.e., the frontal operculum)‐and SMA‐during time processing [Coull et al., 2008]. In this light, it appears likely that the recruitment of precisely the same areas in the play task may reflect additional temporal processing, rather than a specific regional signature of WM manipulation in the tactile domain. Still, the present results do not point to any engagement of dorsolateral PFC during active WM manipulation of somatosensory information. Rather, both tactile WM tasks were associated with similar regions of activation in VLPFC/IFG, consistent with a general role of this area in WM processing of nonspatial tactile information [e.g., Kostopoulos et al., 2007; Preuschhof et al., 2006; Romo et al., 1999; Spitzer et al., 2010; see also Courtney, 2004; Levy and Goldman‐Rakic, 2000].

As one central new finding, delay activity during static maintenance of previous tactile input was characterized by increased coupling of functional activity between IFG and primary somatosensory cortex. The notion of a crosstalk between lateral PFC and areas of lower‐level sensory information processing is well‐established in the WM literature [for review, see e.g., Curtis and D'Esposito, 2003; D'Esposito, 2007; Miller and D'Esposito, 2005] and is gaining increasing empirical support from human neuroimaging studies in the visual and verbal domains [e.g., Fiebach et al., 2006; Rissman et al., 2004; Vuilleumier and Driver, 2007; Yoon et al., 2006; Zanto et al., 2011; for review, see Gazzaley and Nobre, 2012]. In the tactile domain, to the best of our knowledge, only one previous imaging study has reported functional interactions between VLPFC and somatosensory areas (in particular SII) during tactile memory disambiguation at retrieval [Kostopoulos et al., 2007]. Here, we for the first time demonstrate increased coupling between IFG and SI associated with tactile WM maintenance, with SI being independently identified also as the central processing spot of the task‐relevant intensity information already during stimulus encoding (see below; cf. Fig. 3A). In the play task, which likely involved more abstract internal processing, with less attention devoted to the sensory periphery, only spurious IFG‐SI coupling was observed. Together, our connectivity results suggest that functional interactions between lateral PFC and domain‐specific lower‐level areas in particular characterize conditions of static WM maintenance, not only of visual and verbal material, but also of tactile memories.

Parametric fMRI analysis revealed that physical intensity changes during stimulation were mirrored by parametric modulations of BOLD activity in SI, but the active internal online processing of the dynamic stimulus information was selectively represented by enhanced parametric activity in IFG. Importantly, such parametric online processing in IFG was in particular evident during the stimulation‐free delay period in the play task, when subjects merely imagined the intensity fluctuations in WM. This observation may be directly linked to previous electrophysiological work demonstrating parametric WM coding of to‐be‐maintained vibrotactile information in VLPFC of monkeys [Barak et al., 2010; Romo et al., 1999] and humans [e.g., Spitzer et al., 2010; Spitzer and Blankenburg, 2012]. In particular, in human EEG recordings, temporary parametric modulations of upper beta oscillations, with a putative source in right IFG, were observed during purposeful internal updating and refreshment of the task‐relevant WM contents for sustained maintenance [Spitzer and Blankenburg, 2011; Spitzer et al., 2013]. A direct relation between prefrontal BOLD responses and beta oscillations is not yet established [e.g., Hanslmayr et al., 2011; Laufs et al., 2003; Logothetis, 2008; Whittingstall and Logothetis, 2009], and the limited temporal sampling of fMRI may render short‐lived parametric effects difficult to detect in a static maintenance task. However, the manipulation aspect of the present play task can be considered as a condition of continuous internal “updating” of the current WM content throughout the entire delay period. In this light, the observation of a reliable quantitative modulation of IFG BOLD responses in this task appears very consistent with the view that parametric IFG activity, on a large scale population level, may in particular reflect the top‐down controlled updating of analogue quantity information in WM [cf., Spitzer and Blankenburg, 2011, 2012; Spitzer et al., 2013].

In a majority of human neuroimaging studies in the visual and verbal domains, WM processing was operationalized in terms of top‐down selection, rehearsal, and/or manipulation of (sets of) distinct items, for subsequent identification [for reviews, see e.g., Bledowski et al., 2010; Nee et al., 2012]. In this study, a single continuous quantity (i.e., intensity) was to be processed for subsequent sensory matching, in the tradition of previous animal studies of sensory WM function (for review, see e.g., Pasternak and Greenlee, 2005). It may be argued on theoretical grounds that the former class of operations may address established “working” memory function [e.g., Baddeley, 1997; Bledowski et al., 2010], whereas the latter may reflect more basic mnemonic processes, also referred to as “short‐term memory” [STM; for review, see Aben et al., 2012]. On a neuronal level however, the present evidence that internal manipulation operations and control over lower‐level sensory areas can be mediated by the same PFC subareas that also host basic tactile STM function [e.g., Preuschhof et al., 2006; Romo et al., 1999; Spitzer et al., 2010] indicates a strong overlap and mutual relevance of these putatively separate concepts.

In summary, three main findings on the role of VLPFC during tactile memory processing can be derived from the present results. First, static maintenance and dynamic manipulation of tactile information engaged very similar areas in the right IFG, and there was no evidence for a process‐specific allocation of manipulation operations to more dorsal areas of lateral PFC. Second, WM‐related IFG activity during static maintenance was characterized by increased functional coupling with early somatosensory cortex. Finally, we report evidence for a parametric representation of the to‐be‐processed WM contents in IFG BOLD activity, corroborating that population activity in this area may directly reflect task‐relevant quantitative information, in particular during purposeful internal WM updating. Together, complementing previous evidence from monkeys and humans, our findings underpin the multi‐facetted role of human VLPFC/IFG as a core region in somatosensory WM processing.

Supporting information

Supplementary Information