A subsequent-memory effect in dorsolateral prefrontal cortex

A subsequent-memory effect in dorsolateral prefrontal cortex

Cognitive Brain Research 16 (2003) 162–166 www.elsevier.com / locate / cogbrainres Research report A subsequent-memory effect in dorsolateral prefro...

182KB Sizes 3 Downloads 86 Views

Cognitive Brain Research 16 (2003) 162–166 www.elsevier.com / locate / cogbrainres

Research report

A subsequent-memory effect in dorsolateral prefrontal cortex Bart Rypma a , *, Mark D’Esposito b a

b

Department of Psychology, Rutgers University, Smith Hall, 101 Warren Street, Newark, NJ 07102, USA Helen Wills Neuroscience Institute and Department of Psychology, University of California, Berkeley, CA, USA Accepted 26 September 2002

Abstract The importance of brain regions for long-term memory encoding has been examined by comparison of encoding-related neural activity on trials in which successful recollection subsequently occurred to the encoding-related activity on trials in which successful recollection did not occur. We applied similar analyses to event-related functional magnetic resonance imaging (fMRI) data to explore the relative roles of dorsolateral and ventrolateral prefrontal cortex (PFC) regions during specific components of a working-memory (WM) maintenance task. The results of this study indicated that increases in dorsolateral PFC activity during encoding was related to subsequent retrieval-success. These results lend support to the hypothesis that ventrolateral PFC mediates a limited-capacity WM buffer that supports rehearsal maintenance functions while dorsolateral PFC mediates WM organization functions that accommodate the capacity limits of WM maintenance.  2002 Published by Elsevier Science B.V. Theme: Neural basis of behaviour Topic: Learning and memory: systems and functions Keywords: Working memory; Prefrontal cortex; Functional magnetic resonance imaging; Subsequent-memory effect

1. Introduction Working memory (WM), the cognitive system that allows individuals to maintain information over brief time intervals, may be divided into separate processes such as those required for brief retention of information, and those required for allocating attention and coordinating information that is being temporarily maintained [4] Evidence has accumulated to support the notion that PFC may be organized to support different WM processes. Rypma et al. [25] and Rypma and D’Esposito [22,23] for instance, observed that, under low memory load conditions (two to three letters), activation in frontal regions was limited to left ventrolateral PFC (BAs 44, 45 and 47). Additional activation of dorsolateral PFC (BAs 9 and 46) was observed, only during encoding, under high memory demand conditions (six letters). These findings are important because they parallel observed behavioral dissociations between cognitive pro*Corresponding author. Tel.: 11-510-643-4416. E-mail address: [email protected] (B. Rypma).

cesses that mediate performance under low memory demand conditions and those that mediate performance under high memory demand conditions [24]. Specifically, research on the capacity limits of WM storage suggest that low memory load storage is mediated by a short-term buffer of limited capacity (perhaps 41 / 21 items, see Cowan [7]), whereas higher memory loads require additional processing to compress, or ‘chunk’ the information for efficient storage and retrieval, so as to accommodate the capacity limits of the short-term maintenance buffer [18,29,24]. Neuroimaging results showing differential activation patterns under low and high memory demands, and behavioral results suggesting dissociable memory processes under such memory-demands suggest a ‘memory organization hypothesis’ of PFC function in WM. This hypothesis states that ventrolateral PFC mediates the limited capacity short-term buffer of WM while dorsolateral PFC plays a role in the strategic organization, or ‘chunking,’ of information [7,12,18] that permit storage of large amounts of information [22,24]. The role of cortical regions in memory encoding have

0926-6410 / 02 / $ – see front matter  2002 Published by Elsevier Science B.V. doi:10.1016/S0926-6410(02)00247-1

B. Rypma, M. D’ Esposito / Cognitive Brain Research 16 (2003) 162–166

been explored in long-term episodic memory paradigms by comparing activation during encoding on trials in which the subjects later successfully retrieved the information to activation on trials in which later retrieval was unsuccessful [19,10,5,17]. These studies have yielded an important phenomenon, a ‘subsequent-memory effect,’ that has aided in understanding the necessary neural network for successful memory encoding. The ‘subsequent memory effect’ may be defined as significantly greater activation of a particular brain region during memory encoding for trials in which later retrieval was successful compared to that in trials in which later retrieval was unsuccessful. Behavioral evidence suggesting the presence of memory organization functions, and neuroimaging data suggesting a specific role for dorsolateral PFC in these memory organization functions led us to predict that a subsequent memory effect should be observed in dorsolateral PFC when subjects were encoding information with the intention of remembering it several seconds later. To test this prediction, we had subjects perform a WM maintenance task during fMRI scanning. In order to generate activation data on trials in which subjects responded incorrectly, we used a broader and more varied range of WM demands than in studies we have conducted previously. In the present study, subjects performed a delayed-response WM task in which they were required to maintain between one and eight letters over an unfilled delay interval. We used event-related fMRI analysis methods to examine neural activity separately in the encoding, maintenance and retrieval periods of the task.

2. Methods

2.1. Subjects Eight right-handed subjects (age range521–30; three men) were recruited from the medical and undergraduate campuses of the University of Pennsylvania. Subjects were excluded if they had any medical, neurological or psychiatric illness or if they were taking any type of prescription medication. All subjects gave informed consent.

acquisition to allow tissue to reach steady-state magnetization.

2.3. Behavioral task To start each trial, letter strings, ranging in length from 1 to 8, were presented simultaneously in pseudo-random order for 4 s followed by a 12-s unfilled delay. A probe letter then appeared for 2 s during which the subject pressed a button with their right thumb if the probe item was part of the memory set or with their left thumb if the probe item was not part of the memory set. Following these behavioral events there was a 16-s intertrial interval (ITI). The total time from trial onset to trial offset was 34 s (Fig. 1). All subjects completed eight runs of 10 trials each. This design allowed us to examine neural activity uniquely associated with the stimulus encoding, the delay, and the response periods. A total of 136 gradient-echo echoplanar images in time were obtained per slice in each 340-s run. Thus, a total of 1360 observations were obtained for each voxel in the brain for each subject, giving us considerable power to estimate effects within subjects. Subjects viewed a backlit projection screen from within the magnet bore through a mirror mounted on the head coil. Stimulus presentation and reaction time (RT) recording were handled by a Power Macintosh computer.

2.4. Data analysis Off-line data processing was performed on SUN Ultra workstations. After image reconstruction and prior to motion correction, data were sinc interpolated in time to correct for the fMRI acquisition sequence since hemodynamic responses were to be compared across slices that were obtained at different points in the acquisition sequence. The data were motion corrected using a slice-wise motion-compensation method to remove spatially coherent signal changes using a partial correlation method [31] and by applying a six parameter, rigid-body, least squares realignment routine [9]. The details of the event-related fMRI analysis used in this study are presented elsewhere

2.2. MRI technique Imaging was carried out on a 1.5T Signa scanner (GE Medical Systems) equipped with a fast gradient system for echo-planar imaging. A standard radiofrequency head coil was used with foam padding to comfortably restrict head motion. High resolution sagittal and axial T1-weighted images were obtained in every subject. A gradient echo, echoplanar sequence (TR52000 ms, TE550 ms) was used to acquire data sensitive to the BOLD signal. Resolution was 3.7533.75 mm in plane, and 5 mm between planes (21 axial slices were acquired). Twenty seconds of gradient and radiofrequency pulses preceded the actual data

163

Fig. 1. Trial sequence.

164

B. Rypma, M. D’ Esposito / Cognitive Brain Research 16 (2003) 162–166

[31]. Briefly, fMRI signal changes that occurred during particular temporal periods of the behavioral trials were modeled with covariates comprised of shifted, BOLD hemodynamic response functions (HRFs), the fMRI response resulting from a brief pulse of neural activity. Changes in BOLD signal associated with the encoding, delay, and response periods of the behavioral task were assessed with covariates that modeled the expected BOLD signal response in the event of an increase in neural activity (relative to the ITI) occurring in each of the task periods. Our rationale for deriving an HRF is explained elsewhere [1]. An HRF was derived from primary sensorimotor cortex in each subject in the following manner. Before performing the WM task described above, each subject performed a simple RT task in which a central white fixation cross changed briefly (500 ms) to a circle every 16 s cueing subjects to make a bilateral button press. Twenty such events occurred during the 320-s scan (160 images). All scanning parameters were identical to those used for the WM experiment. Because fMRI data are temporally autocorrelated under the null hypothesis [31], the data were analyzed using the modified general linear model for serially correlated error terms [30]. A time-domain representation of the expected 1 / f power structure and a filter that removes frequencies above 0.244 was placed within the K matrix. This filter was also applied to the fMRI time series to remove artifacts at the Nyquist frequency (0.25 Hz). Low frequency (sine and cosine) confounds up to 0.025 Hz and trial-effect covariates were included in our model, to account for frequency components and mean signal change, respectively, that were associated with each trial. Relationships with each task period and the ITI were assessed by contrasts (yielding t-statistics with |1195 df) involving the parameter estimates that corresponded to covariates that modeled each task period. Three covariates modeled each 4-s of the delay period. Given estimates of the temporal smoothness of the hemodynamic response, the covariate modeling the first 4-s of the delay period would be contaminated by hemodynamic activity from the encoding period. Thus, only the second 4-s interval (designated as delay period 1) and third 4-s interval of the delay period (designated as delay period 2) are considered in the analyses. Delay periods 1 and 2 were analyzed separately. To examine activity in specific regions of PFC, dorsolateral PFC regions-of-interest (ROIs) were drawn to include middle and superior frontal gyri (corresponding to BAs 9 and 46) according to the Talairach and Tournoux [27] atlas on standard T1 axial slices in the axial plane. A similar procedure was used to draw ventrolateral PFC ROIs to include inferior frontal gyrus corresponding to BAs 44, 45 and 47. These ROIs were then normalized to each subject’s T1 axial images using a 12-parameter affine transformation [9]. with a non-linear deformation routine [3]. Parameter estimates were derived from each voxel of

each ROI and averaged separately for dorsolateral and ventrolateral PFC ROIs for each subject. Random-effects tests were then used to compare activation (as measured by ROI-wise parameter estimates) for each task period in correct and incorrect trials. Note that our procedure does not entail creation of statistical maps and therefore does not depend on thresholds for determination of neural activity. Thus, we assessed cortical activation in each subject’s dorsolateral and ventrolateral PFC in the encoding, delay and response periods in each of the eight memory load conditions (i.e. one to eight letters) using the parameter estimates (non-thresholded) for the covariates that modeled each task-period, in each memory-load condition. Testing of a priori hypotheses involved planned comparisons of the activation differences between correct and incorrect trials in each task period.

3. Results

3.1. Behavioral performance RT increased with increasing memory-load, F(7,49)5 24.1, P,0.0001, MSe59069.6. The relationship between RT and memory load was well-described by a linearly increasing function (slope50.97, r 2 50.96), which was significant, t(7)513.3, P,0.0001, typical of performance on such tasks [26]. Performance accuracy was high across memory-load conditions (91.8%) but decreased with increasing memory load F(7, 49)57.6, P,0.0001, MSe5 67.3. This relationship was well-described by a linearly decreasing function (slope520.87, r 2 50.75), which was significant, t(7)524.26, P,0.003. These results did not differ with the presence or absence of the probe in the memory set (i.e. whether the correct response was ‘yes’ or ‘no’; ts,1).

3.2. FMRI signal 3.2.1. Correct vs. incorrect trials To assess our prediction of increased neural activity in dorsolateral PFC during encoding of correct trials, we used planned comparisons of average fMRI signal values for trials in which subjects responded correctly to those in which subjects responded incorrectly in each task period, in dorsolateral and ventrolateral PFC. Fig. 2 shows the mean parameter estimates from dorsolateral and ventrolateral PFC in each task period (averaged across memory loads) for correct and incorrect trials. In dorsolateral PFC, during the encoding period, reliable differences were observed between parameter estimates from correct trials and those from incorrect trials, using Bonferroni-corrected ([16]) parametric t-tests, t(7)52.63, P,0.03, and nonparametric paired sign tests (P50.008; [15]). No other effects were significant in dorsolateral PFC and no signifi-

B. Rypma, M. D’ Esposito / Cognitive Brain Research 16 (2003) 162–166

Fig. 2. Mean fMRI signal (parameter estimates) from dorsolateral and ventrolateral PFC during encoding, delay and retrieval periods for trials in which subjects responded correctly (white bars) and those in which subjects responded incorrectly (hatched bars).

cant effects were observed in ventrolateral PFC in both parametric and nonparametric tests.

4. Discussion In this study, we observed greater encoding-related neural activity in trials in which subjects responded correctly than in trials in which subjects responded incorrectly on a WM maintenance task. Significant activation differences between correct and incorrect trials did not occur in ventrolateral PFC but did occur in dorsolateral PFC. The relatively small number of participants in this study could raise concerns about statistical power in the present study. Two points are worth noting in this regard. First, the single largest effect size (1.14; [6]) was observed in dorsolateral PFC during encoding and this effect was more than twice that in any other task period or in ventrolateral PFC. Second, the reliable pattern, of positive activation for correct trials and negative activation for

165

incorrect trials, is qualitatively distinct from any other task period or prefrontal region. The meaning of the negative activation we observed on ‘incorrect’ trials is difficult to interpret for at least two reasons. First, whereas the positive activation on ‘correct’ trials (similar to RTs on correct trials) could reasonably be interpreted as reflecting a fairly circumscribed sequence of cognitive processes that are well studied and well understood, the negative activation on incorrect trials may result from a number of possible processes including attentional lapses or pursuit of nonoptimal strategies. Whatever psychological process negative activations may reflect, these results suggest that a considerable number of dorsolateral PFC neurons simultaneously reduce their activation (compared to baseline) during the encoding of to-be-remembered information on some trials [21]. This process appears to have a sufficiently deleterious effect on ‘downstream’ memory processes to impede successful retrieval of information. The finding, that positive dorsolateral PFC encoding activation is associated with retrieval success and negative dorsolateral PFC encoding activation is associated with retrieval failure suggests support for the hypothesis that dorsolateral PFC is critical for mediation of working memory processes [25], specifically, for initial acquisition of to-be-remembered information [22]. Rypma et al. [25] observed ventrolateral PFC activation during low (three letters) and high (six letters) memory demand conditions. Additional dorsolateral PFC activity was observed only in the high memory-demand conditions. Rypma and D’Esposito [22], using event-related fMRI, isolated this demand-differential activity to the encoding period. Based on these observations, Rypma and his colleagues have argued that dorsolateral PFC plays a critical role in encoding processes that permit efficient memory maintenance and retrieval. These results suggest further support for the notion that capacity-limited WM maintenance processes may be mediated by ventrolateral PFC whereas strategic organization processes may be mediated by dorsolateral PFC [22–24]. How might such strategic organization processes facilitate subsequent retrieval of stored information? A number of theoretical frameworks have been proposed that could provide an answer to this question. In Baddeley’s WM model [23], maintenance rehearsal is a process of perpetual reactivation of items currently stored in the maintenance buffer. In our view, strategic organization acts as a datacompression program to permit supracapacity information maintenance within this buffer and availability to retrieval mechanisms. Another framework, proposed by Cowan [7], emphasizes the role of attention in information maintenance. In this view, items in long-term memory may be activated based on the context of information that is currently within the focus of a capacity-limited attentional spotlight. Chunking of information then, serves to activate the long-term memory representations of items currently in

166

B. Rypma, M. D’ Esposito / Cognitive Brain Research 16 (2003) 162–166

the attentional spotlight, rendering them more accessible to later re-entry into the spotlight. Our conclusions must be considered tentative for a number of reasons. One is that dorsolateral PFC is a complex brain region whose functions extend beyond working memory [8]. Thus, observation of activity in this region may not be exclusively attributable to WM. One additional possibility, for instance, is that these differences in activation reflect the concurrent processing of emotion associated with the demands of higher memory loads (indeed, the behavioral results indicate that most of the error data were generated by these types of trials [14]). Another concern is that the paucity of incorrect trials in low memory-load conditions did not permit us to examine the subsequent memory effect in separate memory-load conditions. We are currently exploring methods to increase error rates in low memory-load conditions [11]. Nonetheless, the isolation of a subsequent-memory effect to dorsolateral PFC regions during memory encoding is consistent with the memory-organization hypothesis we outlined above. These results extend those of other studies in two important ways. First, they are consistent with an emerging set of results suggesting that behavioral differences play a critical role in determining the variability of PFC activation. Previous studies in our laboratory have indicated that response speed (as measured by reaction time) is one important behavioral determinant of PFC activation [22,23]. The present results suggest that another important determinant of PFC activation is response accuracy. Future research will be needed to determine more precisely the nature of the relationship between these behavioral measures and neural activity. Second, these results extend previous observations of subsequent memory effects during long-term episodic memory task performance. In those studies, greater activation was observed during encoding of items that were later remembered than those that were later forgotten in ventrolateral PFC [17] and medial temporal lobe regions [5,10,28]. There are known anatomical and functional connections between these structures and dorsolateral PFC [2,13]. Studies that systematically control short- and long-term memory task requirements will be required to better understand the interactions among these cortical structures and different kinds of memory tasks.

5. Conclusion In summary, the observation that increased dorsolateral PFC activation during memory encoding is associated with successful memory retrieval lends support to a growing body of evidence which suggests that PFC is organized in a dorsolateral / ventrolateral fashion according to the type of processing required in the task [20] and that individual subject performance is a critical determinant of PFC

activation [22,23]. It may be that capacity-limited WM maintenance is mediated by ventrolateral regions of PFC. Dorsolateral PFC regions may be additionally recruited to mediate the memory organization processes required for later successful retrieval of information from the limitedcapacity buffer.

References [1] G.K. Aguirre, E. Zarahn, M. D’Esposito, NeuroImage 8 (1998) 360–369. [2] P. Alvarez, L.R. Squire, Proc. Natl. Acad. Sci. USA 91 (1994) 7041–7045. [3] J. Ashburner, K. Friston, NeuroImage 3 (1996) S111. [4] A. Baddeley, Working Memory, Oxford University Press, New York, 1986. [5] J.B. Brewer, Z. Zhao, J.E. Desmond et al., Science 281 (1998) 1185–1187. [6] J. Cohen, Statistical Power Analysis For the Behavioral Sciences, Erlbaum, Hillsdale, NJ, 1998. [7] N. Cowan, Behav. Brain Sci. 24 (2001) 87–185. [8] M. D’Esposito, D. Ballard, G.K. Aguirre, E. Zarahn, NeuroImage 8 (1998) 274–282. [9] K.J. Friston, A.P. Holmes, K.J. Worsley et al., Hum. Brain Map. 2 (1998) 189–210. [10] J.D.E. Gabrieli, J.B. Brewer, J.E. Desmond, Science 276 (1997) 264–266. [11] S.E.B. Gibbs, E.H. Schumacher, M. D’Esposito, Proc. Soc. for Neurosci. (2002), in press. [12] M. Glanzer, M.J. Razel, Verb. Learn. Verb. Behav. 13 (1974) 114–131. [13] P.S. Goldman-Rakic, L.D. Selemon, M.L. Schwartz, Neuroscience 12 (1984) 719–743. [14] J.R. Gray, T.S. Braver, M.E. Raichle, Proc. Natl. Acad. Sci. USA 99 (2002) 4115–4120. [15] M. Hollander, D.A. Wolfe, NonParametric Statistical Methods, Wiley, New York, 1999. [16] S. Holm, Scand. J. Stat. 6 (1979) 65–70. [17] B.A. Kirchoff, A.D. Wagner, A. Maril et al., J. Neurosci. 1 (2000) 6175–6180. [18] G. Miller, Psychol. Rev. 63 (1956) 81–97. [19] L. Nyberg, R. Cabeza, E. Tulving, Psychon. Bull. Rev. 3 (1996) 135–138. [20] M. Petrides, Semin. Neurosci. 8 (1996) 57–63. [21] M.E. Raichle, Proc. Natl. Acad. Sci. USA 95 (1998) 765–772. [22] B. Rypma, M. D’Esposito, Proc. Natl. Acad. Sci. USA 96 (1999) 6558–6563. [23] B. Rypma, M. D’Esposito, Nature Neurosci. 3 (2000) 509–515. [24] B. Rypma, J.D.E. Gabrieli, Behav. Brain Sci. 24 (2001) 143. [25] B. Rypma, V. Prabhakaran, J.E. Desmond et al., NeuroImage 9 (1999) 216–226. [26] S. Sternberg, Science 153 (1966) 652–654. [27] J. Talairach, P. Tournoux, Brain Atlas, Thieme Medical, New York, 1988. [28] A.D. Wagner, D.L. Schacter, M. Rotte, Science 281 (1998) 1188– 1191. [29] N.C. Waugh, D.A. Norman, Primary memory, Psychol. Rev. 72 (1965) 89–104. [30] K.J. Worsley, K.J. Friston, NeuroImage 2 (1995) 173–182. [31] E. Zarahn, G.K. Aguirre, M. D’Esposito, NeuroImage 5 (1997) 179–195.