Cue- versus Probe-dependent Prefrontal Cortex Activity during Contextual Remembering

Transcription

1 Cue- versus Probe-dependent Prefrontal Cortex Activity during Contextual Remembering Ian G. Dobbins and Sanghoon Han Abstract & Functional neuroimaging comparisons of context and item memory frequently implicate the left prefrontal cortex (PFC) during the recovery of contextually specific memories. However, because cues and probes are often presented simultaneously, this activity could reflect operations involved in planning retrieval or instead reflect later operations dependent upon the memory probes themselves, such as evaluation of probe-evoked recollections. More importantly, planningrelated activity, wherein subjects reinstate details outlining the nature of desired remembrances, should occur in response to contextual memory cues even before retrieval probes are available. Using event-related functional magnetic resonance imaging, we tested this by dissociating cue- from probe-related activity during context memory for pictures. Cues forewarning contextual memory demands yielded more activity than those forewarning item memory in the left lateral precentral gyrus, midline superior frontal gyrus, and right frontopolar cortex. Thus, these anticipatory, cue-based activations indicated whether upcoming probe decisions would require contextually specific memories or not. In contrast, the left dorsolateral/ midventrolateral and anterior ventrolateral PFC areas did not show differential activity until the probes were actually presented, demonstrating greater activity for context than for item memory probes. Direct comparison of proximal left PFC regions demonstrated qualitatively different response profiles across cue versus probe periods for lateral precentral versus dorsolateral regions. These results potentially isolate contextual memory-planning-related processes from subsequent processes such as the evaluation of recollections, which are necessarily dependent on individual probe features. They also demonstrate that contextual remembering recruits multiple, functionally distinct PFC processes. & INTRODUCTION Duke University Memory tasks vary greatly in the amount of contextual recovery necessary for success. At one end of the spectrum, item recognition simply requires subjects to indicate whether retrieval probes seem novel or familiar, requiring minimal or no recollection of the contextual specifics of previous encounters ( Yonelinas, 2002; Mandler, 1980). In contrast, remembering for example, which of two persons gave one a particular gift (e.g., source or context memory), requires retrieving specific context information previously associated with the probe (i.e., the gift). In this case, one is not assessing the familiarity of the probe itself, but instead trying to recover associated contextual details signifying its origin. Cognitive research contrasting context and item memory converges upon the idea that these tasks likely involve different retrieval processes (e.g., Quamme, Frederick, Kroll, Yonelinas, & Dobbins, 2002; Yonelinas, 1999), different decisions mechanisms (Banks, 2000), or perhaps some combination of the two. Neuroimaging studies comparing context to item memory have consistently demonstrated increased activity in a highly left lateralized network (Dobbins, Foley, Schacter, & Wagner, 2002; Raye, Johnson, Mitchell, Reeder, & Greene, 2000; Rugg, Fletcher, Chua, & Dolan, 1999; Nolde, Johnson, & D Esposito, 1998), including the left prefrontal cortex (PFC), lateral parietal, midline precuneus, and retrosplenial areas. Within the PFC, previous research suggests that the anterior ventrolateral region underlies deliberative conceptual processing of items (Wagner, Paré-Blagoev, Clark, & Poldrack, 2001; Poldrack et al., 1999; Klein et al., 1997; Thompson-Schill, D Esposito, Aguirre, & Farah, 1997; Demb et al., 1995). During context memory, this anterior ventrolateral region is likely recruited when the targeted memories are closely linked to specific semantic attributes of the memory probes (Dobbins & Wagner, 2005). This suggests that the region enables selective consideration of memory-relevant, conceptual characteristics of the probes during contextual retrieval attempt (Dobbins et al., 2002). There is also evidence that the left dorsal/dorsolateral PFC plays a distinct role in context memory retrieval. For example, this region was exclusively active during context memory retrieval attempts in Dobbins et al. (2002), yet failed to demonstrate activity during either item memory attempts or during earlier semantic analysis of D 2006 Massachusetts Institute of Technology Journal of Cognitive Neuroscience 18:9, pp

2 the probes. This was taken as evidence that the region participates in monitoring or evaluating the sufficiency of recollections with respect to a memory goal. One potential shortcoming of previous context memory studies, however, is that the cues specifying the type of the context to be retrieved are typically presented simultaneously with the probes upon which memory is to operate. This is potentially problematic because it is generally accepted that contextual memory retrieval necessarily begins with a transformation of the retrieval cue into an online specification of the types of recollective content that will be required during subsequent memory search, and perhaps the behavioral relevance of each type of content (viz., retrieval planning or retrieval descriptions) (Schacter, Norman, & Koutstaal, 1998; Burgess & Shallice, 1996; Norman & Bobrow, 1979). Such planning potentially improves the efficiency of search and ensures that the behavioral significance of recovered recollections is rapidly appreciable. Thus, currently, it is not clear whether the increased left PFC activity observed during context memory signifies increased planning demands relative to item memory, or instead underlies contextual retrieval operations initiated upon, or tied to, the retrieval probes themselves. Here we test the idea that compared to item memory judgments, contextual memory retrieval recruits functionally separable cue- and probe-dependent processes. An example of the later would be evaluating the adequacy of recollections, which would necessarily depend upon first having access to the specific identities of the retrieval probes. Although functional magnetic resonance imaging (fmri) research has not distinguished cue-based planning from later probe-dependent operations, recent reports using event-related potential (ERP) methods support the distinction and broadly implicate PFC as a likely candidate. For example, Herron and Wilding (2004) demonstrated that context memory cues requiring the subject to remember tasks previously performed upon the probes evoked a differential response over left prefrontal regions compared to cues requiring memory for whether the items originated from the first or second lists encountered during prior encoding (see also Morcom & Rugg, 2002). Thus, left PFC regions may act in an anticipatory manner when subsequent retrieval requires the recovery of one s prior cognitive operations (Herron & Wilding, 2004). Because of the relatively slow, integrative nature of blood oxygenation level dependent compared to ERP responses, it is possible that these anticipatory responses have been conflated with later probe-dependent retrieval processes in prior fmri studies. However, although the ERP research examining cue effects during retrieval suggest differential cue-based responses in the PFC, the specific PFC locations underlying these differential responses remain unknown. Using event-related fmri, we explored these hypotheses by jittering retrieval cue onset (context or item) with respect to the subsequent appearance of identically constructed memory probes, and modeling cue versus probe onsets as separate events (Figure 1). In addition, we added the typical simultaneous condition where the cues and probes appear jointly. PFC regions critical for planning context retrieval should demonstrate increased activity for context compared to item memory cues before the presentation of retrieval probes. In contrast, regions involved in conceptual elaboration of the probes, or evaluating probe-specific recollections, should not demonstrate the context versus item advantage until the probes are actually present. METHODS Subjects Fifteen subjects (19 to 26 years old, 6 women) were included in the study. Informed consent was obtained in a manner approved by the Institutional Review Board of Duke University Medical Center. Materials The stimuli consisted of 504 color pictures of common objects on a white background (252 photo-quality pictures, 252 cartoon-quality pictures). Study Procedure Prior to scanning, all subjects performed a brief practice version of the task (15 min) in the laboratory. During this practice task, it was explained that they would perform ratings of pictorial stimuli and later that they would be asked to select an item from a memory set associated with a particular prior rating task, or instead be asked to select a new item from a set (see full description below). They were also explicitly told that each memory set contained a single item from each of the possible memory categories (i.e., one from each rating task and one new item). Following practice, subjects were taken to the scanner and participated in the experiment proper, which consisted of two study/test cycles. Given their prior experience with the practice version of the task, encoding during the scanning portion of the experiment can be considered intentional. During study, subjects were scanned while making either of two judgments for singly presented pictures ( Pleasant?, 1 = yes 2=no, or Realistic?, 1 = yes 2= no). Each judgment task had 84 trials with an additional 42 interspersed fixation baseline trials. During fixation trials, a cross was presented underneath the instruction Relax. Stimulus onset asynchrony for trials (SOA) was 4 sec (Figure 1A). The two encoding tasks constituted the contexts that would later be queried during subsequent context memory testing. For the pleasantness judgment, subjects simply indicated whether or not they 1440 Journal of Cognitive Neuroscience Volume 18, Number 9

3 Figure 1. Experimental paradigm. (A) During study phase, participants rated serially presented pictures on either one of two characteristics (pleasantness or realism). (B) During immediate testing memory discriminations were made among three horizontally arranged pictures (a triplet). For each triplet, one item was novel, and the remaining two were drawn from the two prior classification tasks. Cues specifying the type of retrieval task either appeared simultaneously with each triplet (simultaneous cueing condition) or anteceded the triplet by 6, 8, or 10 sec (delayed/jittered cueing condition). For context memory cues (Pleasantness Task? or Realism Task?), subjects indicated which of the three was associated with a particular prior classification task. For item memory cues (New Item?) they indicated which of the three was novel. felt the depicted item was pleasant. It was stressed that there was no correct or incorrect response to the query and that they should simply indicate their personal preference. For the realism judgment, subjects simply indicated whether they felt the item appeared photorealistic or not. The items themselves varied from photorealistic images to caricatures during both judgment tasks. During subsequent context memory trials, it was assumed that subjects would attempt to recover these prior cognitive operations to satisfy the retrieval demand. Testing immediately followed each study phase. During the simultaneous cueing condition, test probes consisted of a task cue (context or item) above a three-alternative forced-choice (3AFC) probe triplet. Each triplet contained an item from all three possible classes (novel, previously rated for pleasantness, and previously rated for realism; Figure 1B). The retrieval cue disappeared after 0.5 sec with the test probes remaining onscreen for an additional 2.5 sec after cue offset. This was followed by a 1-sec blank screen indicating the end of a trial. Across the test trials, the nature of the retrieval cue pseudorandomly varied ( New?, Pleasantness task?, or Realism task? ). During New? trials, subjects indicated which of the test probes was novel (item memory); during Pleasantness task? and Realism task? trials, subjects indicated which of the probes originated from that prior cognitive task (context memory). Responses were made using a left, middle, or right key press with the left hand. The location of the correct response was pseudorandomized. During the delayed/jittered cueing condition, the onset of retrieval cues (context or item) was jittered with respect to the subsequent appearance of probe triplets. The retrieval cues were presented for 0.5 sec, followed by a fixation screen for one of three durations (5.5, 7.5, or 9.5 sec). After the delay, the fixation screen was replaced with test probes for 3 sec (Figure 1B) and subjects were to respond before the stimuli disappeared. Again, there was a 1-sec blank period before the beginning of the next trial. Thus, for each test scan, we crossed memory type (context or item) with presentation condition (simultaneous or delayed). Within the delayed condition, the cue and probe events were modeled as separate events. The order of these retrieval event types, along with 28 additional fixation trials (25% of total trials), was determined by using an optimal sequencing program (Wager & Nichols, 2003). The entire experimental session contained two study/test cycles. In total, there were 56 context memory trials and 28 item memory trials in both Dobbins and Han 1441

4 the simultaneous and delayed presentation formats for each subject. fmri Acquisition Scanning was performed on a 4T General Electric (Waukesha, WI) scanner using a standard head coil. Functional data were acquired by using a spiral pulse sequence (TR = 2000, TE = 31 msec, 34 axial slices parallel to the AC PC plane with near-isotropic voxels of mm, no gap) designed to minimize susceptibility artifact (Guo & Song, 2003). Before functional data collection, four dummy volumes were discarded to allow for T1 equilibration. Participants head motion was minimized by using a vacuum cushion and a forehead strap. High-resolution T1-weighted anatomical images (3-D spoiled gradient recalled acquisition [SPGR]) were acquired for visualization. fmri Analyses Our analyses focused on the retrieval scans. Data were processed using statistical parametric mapping (SPM99: Slice acquisition timing was corrected by resampling all slices in time relative to the middle slice collected, followed by rigid body motion correction across all runs. Functional data were spatially normalized to a canonical echo-planar imaging template using a 12-parameter affine and nonlinear cosine transformation, and then spatially smoothed with an 8-mm gaussian kernel. Each scanning session was rescaled such that the mean global signal was 100 across the volumes. Statistically, subjects were treated as a random effect. For the analyses, volumes were treated as a temporally correlated time series and modeled by convolving a canonical hemodynamic response function (HRF) with a delta function marking each trial onset. In the case of delayed cueing, both the cue and the probe onsets were modeled as separate events, and similar techniques have been successful in isolating anticipatory responses during working memory tasks (Sakai & Passingham, 2003). We also included temporal and dispersion partial derivatives of the canonical HRF for all conditions in the model. The inclusion of these derivatives can reduce the residual error by capturing systematic delay or dispersion differences relative to the assumed canonical HRF. However, the examination of these derivatives in isolation is problematic because their interpretation rests on fit of the canonical form and because their range of accommodation is fairly restricted (Henson, Price, Rugg, Turner, & Friston, 2002). Given this, our analyses were restricted to regions demonstrating a significant canonical response, and the inclusion of the time and dispersion derivatives primarily served to potentially improve sensitivity by potentially capturing small systematic deviations from the canonical form and preventing these from contributing to the residual error (Henson, Rugg, & Friston, 2001). It is also important to note that within the linear-convolution model, increases in the duration of neural activity of up to approximately 4 sec are reflected primarily by changes in the amplitude of the measured canonical response; hence, the canonical model is capable of capturing moderately extended neural responses (Henson, 2004). This approach strikes a balance between assuming activations are purely event related and canonical, versus modeling activation as a necessarily temporally extended process, and this seemed reasonable given previous findings contrasting load versus delay activity during working memory ( Jha & McCarthy, 2000). The resulting functions were used as covariates in a general linear model, along with a basis set of cosine functions that were used to high-pass filter the data and a covariate representing session effects. The least squares parameter estimates of the best-fitting canonical HRF for each condition of interest were used in pairwise contrasts and stored as a separate image for each subject. These images were then tested against the null hypothesis of no difference between contrast conditions using one-tailed t tests. Activations were considered significant and potentially further scrutinized using region of interest (ROI) procedures if they consisted of 24 or more resampled voxels (3 mm isotropic) and exceeded an alpha threshold of.003 for simple contrasts. This cluster extent threshold was determined to yield a mapwise alpha level of.05 through Monte Carlo simulation (Slotnick, Moo, Segal, & Hart, 2003). Functional ROIs were extracted from the maps satisfying the above statistical criteria using peristimulus time averaging based on the model for each included voxel. Percent signal averages were obtained for the significant voxels within an 8-mm radius of each of the SPM99- identified maxima and further analyzed using off-line statistical software. The implicit baseline response of each voxel reflects the constant term in the least squares linear model describing that voxel s response during the session, and the activity of each voxel was scaled such that the constant was equivalent to 100 for each session. Thus, departures from this value reflect percentage changes relative to a baseline or constant activation level for the sessions. RESULTS Behavioral Data An analysis of variance (ANOVA) contrasting onset condition (simultaneous vs. delayed) and retrieval cue type (Pleasant task item? [P], Realism task item? [R], or Novel item? [N]) was conducted on mean accuracy scores (Table 1). This analysis yielded a main effect of retrieval cue type, F(2,28) = 30.99, p <.001, and an interaction between onset condition and retrieval cue type, 1442 Journal of Cognitive Neuroscience Volume 18, Number 9

5 Table 1. Behavioral Data for Simultaneous and Delayed Test Conditions Retrieval Task Cue Onset Pleasantness Realism Novelty Accuracy Simultaneous.68 (14).59 (.13).86 (.09) Delayed.71 (.12).58 (.16).79 (.14) RT Simultaneous 1822 (257) 1960 (245) 1738 (174) Delayed 2152 (161) 2226 (233) 2137 (152) Accuracy data corresponds to mean three-alternative forced-choice percent correct. RT data represent means of the median correct RT for each subject for each condition. Values in parentheses indicate one standard deviation of the mean. F(2,28) = 4.03, p <.05. The main effect of retrieval cue type resulted because accuracy for novel item detection was superior to context memory for pleasantness ratings that was in turn superior to context memory for realism ratings ( p <.05; Tukey s HSD). More importantly, there was no main effect of onset condition on accuracy ( p >.25) indicating that the cue onset manipulation neither generally facilitated nor impaired memory accuracy. Furthermore, although there was an apparent interaction between onset and retrieval cue, pairwise comparisons of performance for each cue type across simultaneous versus delayed presentation conditions revealed no significant differences (all ps >.1; Tukey s HSD). That is, the onset manipulation did not differentially affect any particular retrieval condition. Given this and the failure to observe a main effect for the cue onset condition, it appears that the onset manipulation did not significantly alter retrieval accuracy. An analogous ANOVA of median reaction time (RT) for correct trials yielded a main effect of cue onset, F(1,14) = 61.20, p <.001, and retrieval cue type, F(2,28) = 5.02, p <.05, with no interaction between the two (Table 1). The main effect of cue type indicates that subjects were generally slower during the delayed than simultaneous cue conditions during correct responding. With respect to the main effect of retrieval cue type, post hoc comparisons demonstrated that context memory for realism ratings was significantly slower than novel item detection (Tukey s HSD); however, no other pairwise comparisons were significant. Overall, the behavioral data suggest that the delayed cueing neither impaired nor improved retrieval accuracy in a general fashion. However, perhaps somewhat surprisingly, subjects were generally slower to respond during the delayed condition. The reason for this slowing is unclear. It may be a function of the jittering of the cue and probe onsets that could result in some degree of startle. That is, the unpredictable probe onsets, which occurred anywhere from 5.5 to 9.5 sec following the cue, may have yield some small degree of startle or otherwise slowed initiation of retrieval attempts. Alternately, subjects may have chosen to spend longer on the retrieval decision when given the additional time afforded by the advanced cue. However, it is important to note that spending longer on retrieval decisions does not guarantee more accurate responding, particularly if evidence accrual asymptotes early. Thus, although subjects may have chosen to spend longer on the decision, it did not improve their performance relative to the simultaneous cue/probe condition. Finally, it could also be the case that subjects forgot the cue on some small proportion of trials, necessitating recovery of the actual task to be performed before the actual memory retrieval attempt. Although this can never be ruled out in any paradigm using a delay between cues and probes, such forgetting would not systematically differ across the context and item cue types but instead would generally elevate RT. Regardless of the actual cause of the relative slowing between simultaneous and delayed conditions, it is important to note that none of the imaging results or conclusions noted below rests upon a contrast across the simultaneous and delayed cueing trials. Instead the key contrasts are between context and item memory trials within either the simultaneous or delayed cueing conditions, and given this, a general difference in RT between these cueing conditions is likely inconsequential. Functional Imaging Data Simultaneous Cue Probe Onset The first analysis focused on the simultaneous presentation condition to ensure consistency with prior findings. Because task outcome (correct vs. incorrect) has been shown not to alter responses in the PFC ROIs during context memory judgments (Kahn, Davachi, & Wagner, 2004; Dobbins, Rice, Wagner, & Schacter, 2003), we initially collapsed across this factor to maximize power, although we specifically investigate context retrieval outcome below to verify that PFC context memory activations were insensitive to trial outcomes. Replicating previous research, when context and item recognition trials were directly compared during simultaneous presentation, the expected left-lateralized network demonstrated greater activity during the former (Figure 2A). This network included left dorsal/dorsolateral, ventrolateral, frontopolar, and medial PFC regions. In addition, left lateral parietal, retrosplenial, and precuneus areas were also more active during attempted context compared to item retrieval. The reverse comparison, contrasting item with context memory, revealed greater activity selectively in the right hemisphere in the putamen, posterior ventrolateral PFC, and superior parietal area (not shown). Consistent with previous findings (Kahn et al., 2004; Dobbins et al., 2002), the left PFC regions recruited during context memory were not recruited to a greater Dobbins and Han 1443

6 Figure 2. (A) A contrast of source and item memory during the simultaneous cue probe presentation implicated expected left prefrontal, midline, and parietal regions previously observed in contrasts of context and item memory. Results are overlaid upon a canonical brain and thresholded at.003, 24 voxels. (B) Regions in the left dorsal and lateral PFC did not correlate with RT during context memory judgments, even at a relaxed threshold of.01, 24 voxels. In contrast, dorsal premotor regions (BA 6) demonstrated a linear relationship with RT during the context memory judgments. (C) Targeted contrasts of cue- (dark) and probe-dependent (light) processes implicated different areas of PFC and other regions. The findings in the left PFC demonstrate that prior contrasts of context and item memory likely included regions with different functional properties that can be seen by contrasting this overlay and that depicted in (A). The axial slices depict the same activations with the white lines in the sagittal view showing the relative locations of the axial slices. extent when retrieval was unsuccessful, and this argues against a general difficulty or effort explanation. More specifically, when incorrect versus correct context memory trials were directly compared, even at a liberal threshold (.01 threshold with 24-voxel extent), none of the context-memory-sensitive left PFC regions noted above were implicated, although subjects tended toward slower responding during incorrect trials, t(13) = 1.93, p =.075. A similar analysis was not performed on item memory trials because incorrect responses were too infrequent. In addition to the null effect of trial outcome, the same left-lateralized PFC regions are implicated during context memory even when the comparison/ reference memory task is clearly more behaviorally demanding (Simons, Gilbert, Owen, Fletcher, & Burgess, 2005; Dobbins et al., 2003), and activations in similar, although somewhat more dorsal and anterior regions, have been shown to not track large differences in accuracy or RT across episodic judgment tasks (Dobbins & Han, 2005). However, to more fully rule out a difficulty account of the lateral PFC activity, we also conducted an additional parametric fmri analysis that tested whether RTs modulated activity during the memory judgments in these left dorsal and lateral PFC regions. The model included regressors for each simultaneous retrieval condition (context and item), plus a parametric regressor for the conditions designed to capture regional responses that covaried linearly with RT. Figure 2B demonstrates that although dorsal premotor regions (Brodmann s area [BA] 6) demonstrated a positive linear relationship with RT during context memory judgments, regions along the left dorsolateral and lateral PFC surface were not implicated, and this remained the case even when the threshold was relaxed considerably (.01, 24-voxel extent). There was no parametric relationship evident during item memory judgments. In summary, the comparison of context and item recognition during standard simultaneous presentation of the cues and probes demonstrated increased activation in left dorsal and lateral PFC regions previously implicated using similar contrasts. This increased response is unlikely the result of simple differences in success rates or RTs across the tasks, as these regions 1444 Journal of Cognitive Neuroscience Volume 18, Number 9

7 were not implicated when contrasting unsuccessful versus successful context memory trials, nor did activity in these regions appear to show a reliable relationship with RT during either retrieval task. The insensitivity of these regions to simple differences in RT or accuracy is consistent with prior research. Isolated Cue-related Responses To examine whether significant anticipatory differences arise for context versus item task cues, we contrasted cue-related activity during the isolated cueing period of the delayed condition. The SPM analysis demonstrated that compared to item memory cues, context memory cues elicited increased activation in the PFC and posterior visual association areas, including increases along the left precentral (BA 44/6), medial superior frontal (BA 6/32), and right frontopolar PFC (BA 10) areas (Figure 3A; Table 2). Visual inspection of the hemodynamic responses extracted from ROIs suggested that this differential context advantage was either absent or muted during the subsequent probe presentation. This was confirmed statistically by using the mean response from 4 to 8 sec for each condition. These activity estimates were entered into ANOVAs with factors of Memory Task (context or item memory) and Trial Phase (cue or probe). Significant interactions were obtained in the right frontopolar PFC (BA 10), F(1,14) = 15.15, p <.005 (Figure 3E); left superior/medial frontal gyrus (BA 6/32), F(1,14) = 7.78, p <.05 (Figure 3D); left lateral precentral gyrus (BA 6), F(1,14) = 29.18, p <.001 (Figure 3C); and posterior visual association cortex (BA 18), F(1,14) = 24.91, p <.001 (Figure 3B). Thus, the increased response to context versus item cues did not translate into a similarly increased response to context versus item probes in these regions. Although these regions demonstrated an increased response to the context versus item cues that was largely restricted to the cue phase, the areas were not simply inactive (with the exception of the right frontopolar region) during the later probe presentation. Consistent with this, three of the four ANOVAs above also yielded significant main effects of Trial Phase, indicating greater overall activity during the probe than cue phases in the Figure 3. (A) Regions demonstrating significant increases in response to isolated context compared to item memory cues. Activation is overlaid on a canonical brain and thresholded at.003, 24 voxels. (B E) Extracted hemodynamic responses from functional ROIs drawn from the map generated in (A). Solid lines demonstrate the context and item responses during the isolated cue period, whereas the dashed lines show the responses from the same regions during the subsequent probe presentations. The y axes indicate percent signal change and the x axes show approximate poststimulus onset out to 20 sec. The small inset on the right of each plotted time course shows the location and size of the ROI used for each extraction. Dobbins and Han 1445

8 Table 2. Regions Demonstrating Significant Increases in Response to Isolated Context versus Item Memory Cues Region Lat. BA x y z Vox. Z Score Left posterior PFC Precentral gyrus L L IFG/Insula L Superior/Medial PFC Pre-SMA L 32/ R 32/ Frontopolar MFG/SFG R R Parietal Postcentral gyrus R 5/ Occipital Pole/Medial occipital Cuneus L R R 19/ R Occipital/Inferotemporal Lingual gyrus L L 19/ R 19/ R R 19/ Lingual/Fusiform R 18/ Fusiform R Middle occipital gyrus R Posterior temporal MTG L Anterior temporal STG L 22/ Listed regions are SPM maxima containing at least 24 significant voxels within an 8-mm-radius sphere. Lat. = laterality; Vox. = number of significant voxels; IFG = inferior frontal gyrus; SMA = supplementary motor area; MFG = middle frontal gyrus; SFG = superior frontal gyrus; MTG = middle temporal gyrus; STG = superior temporal gyrus. Hemodynamic responses for entries in bold are shown in Figure 3. left superior/medial frontal gyrus (BA 6/32), F(1,14) = 7.27, p <.05 (Figure 3D); left lateral precentral gyrus (BA 6), F(1,14) = 5.59, p <.05 (Figure 3C); and posterior visual association cortex (BA 18), F(1,14) = 8.50, p <.05 (Figure 3B). The only exception to this pattern was the right frontopolar region ( p >.08). This finding suggests that although these regions are recruited fairly selectively during the cue period, with the exception of the right frontopolar region, they later participate in a less specific or more generalized fashion during subsequent presentation of the probes for both item and context memory trials. The implication of this for the functional characterizations will be addressed in the discussion. Isolated Probe-related Responses Although posterior precentral regions were implicated in context cue based responding, much of the differential left PFC activity observed during simultaneous presentation trials (Figure 2A) was clearly not evoked by the context memory cues in isolation. Given this, we reasoned that probe-dependent processes must also play a key role in the large left PFC responses typically observed when contrasting context and item memory. Furthermore, because probe-dependent processes should be present during both the simultaneous condition and during the isolated presentation of probes in the delayed condition, we used the conjunction of these two contrasts to target probe-related processing. Thus, probe-dependent processes selective for context memory were identified by determining the overlap between the regions demonstrating greater context versus item responses during simultaneous cue probe presentations, and also greater context versus item trial responses during the isolated probe presentations of the delayed presentation condition (.01 each effect; nominal joint probability,.0001; minimum extent, 24 voxels) (Figure 4 and Table 3). This analysis implicated regions not observed during the previous cue-based analysis including the left anterior ventrolateral PFC, left posterior/middle ventrolateral PFC, left dorsal/dorsolateral PFC, midline precuneus and retrosplenial areas, and left lateral parietal cortex (Figure 4A; Table 3). To further verify the selectivity of these responses to probe-dependent processes, we directly contrasted extracted ROI responses of these regions across the isolated cue and probe phases of the delayed condition. Again, the mean response from 4 to 8 sec was used as the dependent variable in ANOVAs with factors of Memory Task (context or item memory) and Trial Phase (cue or probe). If the context memory responses are primarily restricted to probe operations, then significant interactions should be observed across the trial periods. Consistent with this, the analyses demonstrated significant interactions in the left dorsal/ dorsolateral PFC (BA 9), F(1,14) = 13.73, p <.005 (Fig Journal of Cognitive Neuroscience Volume 18, Number 9

9 Figure 4. (A) Regions demonstrating probe-dependent responses. Map was derived from the conjunction of context greater than item memory responses in both the isolated probe and simultaneous cue probe presentation conditions under the prediction that probe-related processes should be present in both contrasts (.01 each contrast; nominal joint,.0001; voxel extent, 24 voxels). (B E) Extracted hemodynamic responses from ROIs contrasting cue and probe activity during the delayed-onset presentations. Conventions are the same as those in Figure 3. ure 4E); left lateral parietal cortex (BA 40/7), F(1,14) = 9.08, p <.01 (Figure 4D); left medial precuneus (BA 7), F(1,14) = 4.93, p <.05 (Figure 4B); and left anterior ventrolateral PFC (BA 10), F(1,14) = 6.27, p <.05. These interactions demonstrate that unlike the regions identified in the cue-based analyses, the regions here tended to demonstrate a larger context versus item response difference during the probe instead of cue phases. However, similar to the cue-based analyses, there was a general tendency for net activity to increase when transiting from the cue to probe phases as demonstrated by a significant main effect of Trial Phase in all the regions reported above ( ps <.05). Interactions across Left PFC Regions Overall, the analyses above suggest that whereas some PFC regions show preferential context memory responses during the cue phases, others show preferential context memory responses during the later probe phase of the trial. When considered jointly, these regions largely coincide with what was observed during the more typical paradigm in which the cues and probes are shown simultaneously. This can be seen by comparing the map demonstrating both cue- and probeselective responses (Figure 2C) with the map obtained from contrasting the context and item trials during simultaneous presentation (Figure 2A). Notably, it appears that very proximal regions in the left PFC show qualitatively different responses across the cue and probe phases of the delayed trials. More specifically, the overlay in Figure 2C suggests that the posterior precentral PFC region was preferentially sensitive to context versus item cue differences, whereas the more anterior dorsolateral/midventrolateral region was preferentially sensitive to context versus item probes. We statistically confirmed this by directly comparing the mean response across two proximal ROIs from these regions by using an ANOVA with factors of Region (precentral or dorsolateral), Trial Phase (cue or probe), and Memory Task (context or item). This analysis yielded a significant three-way interaction, F(1,14) = 49.69, p <.001. As can be seen in Figure 5 (and consistent with the two-way interactions for these ROIs reported earlier), this interaction resulted because the differential context versus item memory response for the precentral region was restricted to the retrieval cues, whereas the differential response of the dorsolateral region was instead Dobbins and Han 1447

10 Table 3. Probe-dependent Responses Showing Increases for Source versus Item Trials (Conjunction of Source Greater than Item Responses for Both Isolated Probe and Simultaneous Cue Probe Presentation Conditions) Region Lat. BA x y z Vox. Z Score Left posterior PFC MFG L L L Left anterior ventral PFC MFG L MFG/IFG L L Parietal SPL L 40/ Precuneus L 31/ L L 31/ See Table 1 for abbreviations. SPL = superior parietal lobule. Hemodynamic responses for entries in bold are shown in Figure 4. Z score is the average for the isolated probe and simultaneous cue probe conditions. Hemodynamic responses for entries in bold are shown in Figure 4. restricted to the memory probes. The only other significant effects of the analysis were a main effect of Trial Phase, F(1,14) = 6.78, p <.05, indicating greater net activity during the probe compared to cue phase, and a main effect of Memory Task, F(1,14) = 6.30, p <.05, indicating greater net activity during the context as opposed to the item memory tasks. Interpretations of the absolute size of the responses (as opposed to differences in the pattern of response) across regions are not valid because of potential differences in vasculature. DISCUSSION The goals of the present study were to (1) identify regions potentially involved in cue-initiated planning aspects of contextual retrieval, (2) isolate these from areas involved in probe or probe-dependent processes, and (3) determine if these areas were coincident with ones identified in the more typical simultaneous presentation of cues and probes. Cue-related Processing Contrasting context and item cues presented in isolation revealed that numerous PFC and posterior regions demonstrated increased responses that anticipated subsequent context versus item retrieval demands (Table 2). In particular, regions along the left precentral gyrus overlapped with those implicated during the contrast of context and item trials when cues and probes were presented simultaneously, and this area has frequently been implicated in other studies of context memory (Kahn et al., 2004; Dobbins et al., 2002; Rugg et al., 1999; Nolde et al., 1998). We take this activity as indicative of the need to plan or specify the nature of future retrieval operations. More specifically, during context memory trials, subjects must use the retrieval cue to help specify the types of target content to be recovered and behavioral relevance of this Figure 5. Demonstration of qualitatively different response profiles of proximal BA 9 and BA 6 lateral PFC regions. Box plots show mean response 4 8 sec post onset for the conditions of interest. Box is one standard error of the between-subjects mean; box plus whiskers equals two standard errors. The axial overlay demonstrates the proximity of the two ROIs used in the ANOVA. Dark gray in the box plot and overlay correspond to BA 6, and light gray correspond to BA 9. During the cue event, the more posterior BA 6 region was differentially sensitive to the context versus item distinction, but the more anterior BA 9 region was not [left of box plot; F(1,14) = 36.61, p <.001]. In contrast, during the probe event, the BA 9 region was sensitive to the context item distinction but the BA 6 region was not [right of box plot; F(1,14) = 14.54, p <.005]. Overall, this resulted in a significant three-way interaction across Region (BA 9 vs. BA 6), Trial Phase (cue vs. probe), and Retrieval Task (context or item memory), F(1,14) = 49.69, p <.001. DLPFC = dorsolateral prefrontal cortex Journal of Cognitive Neuroscience Volume 18, Number 9

11 to-be-remembered content. Such contextual planning demands are largely absent when instead the cue notifies the subject that simple familiarity/novelty will be sufficient, and indeed it is often assumed that subjects can register items as novel or familiar with minimal reference to the specifics of prior contexts. It is important to note that the proposed planning process is clearly not unique to episodic retrieval. For example, Sakai and Passingham (2003) noted similar anticipatory responses in left lateral precentral regions when instructional cues forewarned an upcoming verbal working memory task, but not an upcoming spatial working memory task. This anticipatory response did not simply return to baseline when the actual verbal working memory task began, but instead increased during the actual execution of the verbal working memory task. Somewhat similarly here, the precentral region showed an anticipatory response selective for contextual memory cues, and then its response was generally elevated during the subsequent appearance of probes regardless of the retrieval task (Figure 3C). The frequent association of precentral gyrus with verbal working memory tasks (Cabeza & Nyberg, 2000) suggests that in the current paradigm, the requirement to plan contextual memory retrieval involved greater verbal or conceptual working memory demands than imposed during isolated item memory cueing. In contrast, both the context and item recognition tasks would be expected to engage working memory once the probes are presented, because in both cases the subjects need to decode and maintain the identity of the items to carry out the memory tasks. From this perspective, context and item memory tasks differ in the need to recruit working memory when translating the cues into retrieval plans, but less so with respect to the processing of the probes themselves. If correct, then one would expect that in the standard paradigm, when cues and probes are presented simultaneously, the activity of precentral regions would reflect an amalgam of these separate working memory demands. This is consistent with the current findings in that during simultaneous presentation, the precentral region shown in Figure 5 was significantly more active during context memory (M =.26) than during item memory (M =.14), t(14) = 2.23, p <. 05, although the latter was still significantly above baseline (95% confidence interval, ). This suggests that this often observed pattern of precentral response (context memory > item memory > baseline) in the literature is at least partially the result of an amalgam of separate working memory demands for cue and probe processing, with context memory cue processing placing a greater demand on the region than item memory cue processing. Probe-related Processing In contrast to the anticipatory responses during the cueing phase, a network of left-hemisphere regions required the presence of retrieval probes to evince increased activity during context compared to item trials (Table 3). These included the left anterior ventrolateral (BA 47/10), posterior dorsal/dorsolateral PFC (BA 9), and midventrolateral PFC (BA 45), along with the left lateral superior parietal lobule (BA 7) and midline precuneus (BA 7). Activation in the left anterior ventrolateral PFC has been previously associated with conceptual or semantic item processing ( Wagner et al., 2001; Poldrack et al., 1999; Thompson-Schill et al., 1997) and during contextual retrieval has been argued to reflect the need for such processing when the retrieval task is closely linked to the conceptual attributes of the retrieval probes (Dobbins & Wagner, 2005; Dobbins et al., 2002). Because the context memory task in the current paradigm specifically requires recovering previous decisions about the probes properties, increased activity in this region is consistent with this hypothesis. In contrast, the posterior dorsal/dorsolateral PFC has been proposed as supporting recollective monitoring, which is the active screening or selection of recollections for current task relevance (Simons & Spiers, 2003; Dobbins et al., 2002; Fletcher & Henson, 2001). The abutting more ventrolateral region has somewhat similarly been suggested to underlie general selection mechanisms (Gold & Buckner, 2002; Thompson-Schill et al., 1997), which would presumably also be necessary when selecting or evaluating probe-dependent contextual recollections. Consistent with these characterizations, the current data demonstrated a selective increase in this region (and lateral parietal cortex) in response to context but not item recognition that was contingent upon the presence of the probes. That this differential recruitment was not observed during the cueing period more firmly establishes the role of this region during active retrieval processing dependent on probe identity. Again, as with the precentral and anterior ventrolateral regions, the above characterization does not assume that the dorsal/dorsolateral region is exclusively dedicated to long-term episodic retrieval, and indeed the selection or evaluation of recovered episodic content clearly represents only a subset of the evaluative tasks that recruit this region. As an illustration, Raye, Johnson, Mitchell, Reeder, and Greene (2002) have identified a similar dorsal region as involved in the refreshing of previously presented words in tasks dependent upon short-term as opposed to long-term memory. For example, in an fmri study comparing the young and elderly, young, but not elderly, subjects demonstrated a selectively elevated response when cued to think of words appearing only 550 msec earlier, compared to simply reading presentations of repeated or new words ( Johnson, Mitchell, Raye, & Greene, 2004). The current data and those of Johnson et al. (2004) generally point to the left dorsal/ dorsolateral PFC as critical during selection or evaluation recovered conceptual content with respect to a Dobbins and Han 1449

12 response goal, regardless of whether it originates from long-term or short-term stores (see also Rowe, Toni, Josephs, Frackowiak, & Passingham, 2000). Left PFC Regional Interaction In the current study, proximal regions in the left PFC appeared to show qualitative differences in their sensitivity to the context versus item memory distinction, depending upon whether subjects were in the cue or probe phase of the delayed trials. This was confirmed with ANOVA and yields the first evidence, to our knowledge, that these areas serve fundamentally different roles during context memory recovery. When combined with previous reports suggesting a unique role for anterior ventrolateral PFC (Dobbins & Wagner, 2005; Dobbins et al., 2002; Wagner et al., 2001), this suggests that at least three functionally separable left PFC regions contribute to the volitional recovery of contextually specific remembrances, with different roles for anterior ventrolateral, dorsal/dorsolateral, and precentral regions. Although functional imaging cannot establish the necessity of these regions per se, recent neuropsychological data further support the idea that context memory requires the coordination/contribution of multiple, functionally specialized PFC regions (Alexander, Stuss, & Fansabedian, 2003). Right Frontopolar Activity One unexpected region that was recruited during the context memory versus item memory cue periods was the right frontopolar cortex (Figure 3E, solid lines). As noted earlier, context memory effects are usually exclusively left lateralized, and, indeed, during the simultaneous cue probe presentation condition, there was no evidence for a differential recruitment of this, or any other right PFC region, even with a very liberal threshold (.01, 24-voxel extent). Thus, the data suggest an anticipatory function or role for the region during the isolated context memory cueing that was apparently masked when cues and probes were presented simultaneously and was also not present during the isolated probe presentations of the delayed condition (Figure 3E). The notion of an anticipatory role for frontopolar regions is not new (see also Buckner et al., 1998), and this assertion is also consistent with the need to sometimes model late-onset regressors to capture frontopolar activity (Henson, Rugg, Shallice, & Dolan, 2000; Buckner et al., 1998; Schacter, Buckner, Koutstaal, Dale, & Rosen, 1997). Furthermore, an anticipatory role would also explain why comparisons of more versus less demanding retrieval blocks ( Velanova et al., 2003) would capture frontopolar activity. In short, we suggest that the region may participate in operations executed in anticipation of contextual retrieval and perhaps other demanding cognitive tasks (cf. Braver, Reynolds, & Donaldson, 2003) where appropriate responding is highly contextually specific. When several demanding retrieval trials are predictably grouped together it would be possible to model this anticipatory activity as a blocked or tonic response. Furthermore, because the response is anticipatory, it would also be possible to capture it by modeling late event-related regressors for each trial, although the activity would in fact be indicative of processing in preparation for the upcoming trial. That said, labeling the response as anticipatory still leaves the cognitive operations it may support largely unspecified. Several high-level functional accounts of the frontopolar PFC have recently been proposed (Koechlin, Ody, & Kouneiher, 2003; Braver & Bongiolatti, 2002; Christoff et al., 2001). Although we are unable to review these accounts here, we note that a successful account of the current data would need to explain why the anticipatory response of the region was disrupted by presentation of the probes. One recent hypothesis potentially consistent with this pattern holds that frontopolar regions are critical in gating or regulating the relative influence of internally versus externally generated information (Burgess, Simons, Dumontheil, & Gilbert, in press). From within this framework, the right frontopolar response may indicate inwardly directed attentional processing, perhaps as subjects consider previous outcomes or strategies when planning the upcoming attempt. However, if the activity reflected an internally directed process, it would necessarily have to be suspended when the retrieval probes appeared so that stimulusdriven retrieval products could be generated and/or evaluated. This would then explain the absence of a response during the probe phase of the delayed trials. Regardless of the validity of this tentative hypothesis, the current data nevertheless firmly demonstrate that the right frontopolar region can selectively respond in an anticipatory manner depending upon the expected demands of upcoming retrieval, and, similarly, that it can respond in an event-related manner and therefore need not necessarily reflect tonically maintained cognitive operations (cf. Braver et al., 2003; Velanova et al., 2003). This latter observation potentially critically bears upon the characterization of this region as supporting retrieval mode (Lepage, Ghaffar, Nyberg, & Tulving, 2000), which is a hypothetical neurocognitive set or state in which one processes currently presented probes as potential cues to prior experiences, while the descriptions of those target experiences are statically held in mind. Although somewhat similar to the planning process hypothesized here, one key point of departure is that retrieval mode is assumed not to differ across different types of episodic retrieval tasks (Rugg & Wilding, 2000). That is, provided one is asking episodic questions, the exact format and nature of the questions/cues is held not to affect activity underlying retrieval mode Journal of Cognitive Neuroscience Volume 18, Number 9