Reviews and perspectivesEvent-related fMRI studies of episodic encoding and retrieval: Meta-analyses using activation likelihood estimation
Introduction
For decades, studies of brain-damaged humans and experimentally lesioned animals have provided the bulk of evidence regarding the brain bases of episodic memory (i.e., conscious memory for personally experienced events within a particular spatio-temporal context; Tulving, 1985). More recently, the advent of non-invasive functional neuroimaging has enabled researchers to examine normal memory processes in the healthy brain. Functional neuroimaging can address questions that the lesion method cannot, for example, the examination of similarities and differences between encoding and retrieval stages and the identification of wide-scale networks of regions that interact to support memory. Since the early 1990s, there has been a dramatic increase in the number of functional neuroimaging studies of episodic memory. The first used positron emission tomography (PET) and block-design functional magnetic resonance imaging (fMRI), and had relatively coarse spatial and temporal resolution (and many did not cover the whole brain). Nevertheless, reviews and meta-analyses of these early findings spurred the generation of novel hypotheses and new directions for research. Current studies generally image the Blood Oxygenation Level Dependent (BOLD) signal from the whole brain using fMRI, and have much greater spatial and temporal resolution than the first generation of studies. In fact, temporal resolution is now sufficient to analyze performance trial-by-trial in an event-related design. However, despite this large number of studies, relatively few reviews and meta-analyses have been conducted, and these have typically been selective. Therefore, this appears to be a particularly important time to perform a comprehensive, quantitative meta-analysis of the burgeoning event-related fMRI (ER-fMRI) literature on episodic memory.
Currently, researchers typically employ variations of two standard paradigms to examine episodic memory in the scanner. First, in the successful encoding method, also referred to as the subsequent memory or difference due to memory paradigm (Dm; Brewer, Zhao, Desmond, Glover, & Gabrieli, 1998; Paller, Kutas, & Mayes, 1987; Sanquist, Rohrbaugh, Syndulko, & Lindsley, 1980; Wagner et al., 1998), brain activity is measured when participants are initially exposed to memoranda, and, following a yes–no recognition memory test, average activations are compared between items that were encoded more successfully (i.e., elicited hits on the subsequent recognition test) vs. less successfully (i.e., misses). Second, in a successful retrieval analysis (e.g., Buckner et al., 1998), participants initially study a list of items and are imaged during the memory test. Activations then are compared between trials in which the person successfully endorsed a studied item (i.e., hits) vs. successfully rejected a new one (i.e., correct rejections; CRs). In both of the aforementioned methods, a difference in neural activity associated with each kind of response indicates that the two kinds of responses (i.e., hits vs. misses, or hits vs. CRs) are supported by at least partially non-overlapping neural substrates. Researchers can assume that the reason for a person's remembering some items and not others is due to some combination of the characteristics of each item, and each participant's prior experience, traits, and state at the time of encoding and retrieving each item (Gabrieli, 2001).
Following Scoville and Milner's (1957) report on amnesic patient HM, human and animal lesion work focused intensively on the role of the hippocampus and surrounding medial-temporal lobe (MTL) in memory. One of the more surprising findings from early functional neuroimaging reviews, however, was that medial-temporal activations were far from reliable, but activations in other regions were ubiquitous. These other regions included prefrontal and parietal cortices, the cingulate gyrus, the retrosplenial region, and the cerebellum (e.g., Buckner, 1996; Cabeza & Nyberg, 2000; Fletcher & Henson, 2001; Lepage, Habib, & Tulving, 1998; Nyberg, Cabeza, & Tulving, 1996; Schacter & Wagner, 1999).
Current models generally agree that the MTL supports the creation and possibly also retrieval of distributed memory traces that consist of ensembles of MTL and neocortical neurons (e.g., Alvarez & Squire, 1994; Moscovitch, 1992, Moscovitch et al., 2005), but there are still open questions concerning the functional neuroimaging evidence. For example, how reliable is the activity in the medial-temporal region? Early PET and block-design fMRI studies often failed to detect medial-temporal activation, especially during retrieval, and, in a qualitative review, Henson (2005) noted that current studies also often fail to do so. Second, is memory-related activation concentrated in the hippocampus proper, or distributed more broadly across the medial-temporal lobe, including the parahippocampal gyrus (i.e., perirhinal, entorhinal, and parahippocampal cortices)? This is a question that has received considerable interest in the context of dual-process models of memory (outlined below). Finally, various claims have been advanced regarding the locus of activation along the long axis of the hippocampus during encoding vs. retrieval (e.g., Greicius et al., 2003, Henson, 2005, Lepage et al., 1998, Ludowig et al., 2008; Parsons, Haut, Lemieux, Moran, & Leach, 2006; Schacter & Wagner, 1999). Would a meta-analysis reveal that activation is concentrated more in one segment than another of the hippocampus?
Additional brain systems are proposed to support or modify the basic operations of the MTL. For example, there is evidence for a major contribution of prefrontal cortex (PFC) to successful encoding. Specifically, encoding-related activity in the ventrolateral PFC (VLPFC) has been attributed to selection, maintenance, and control of incoming information, whereas activity in the dorsolateral PFC (DLPFC) is thought to support organization and associative encoding (for reviews, see Blumenfeld & Ranganath, 2007; Paller & Wagner, 2002; Simons & Spiers, 2003). At retrieval, sub-regions of prefrontal cortex have been implicated in various functions, including setting of a retrieval mode, specification of retrieval cues, and post-retrieval monitoring and verification (e.g., Burgess, Dumontheil, & Gilbert, 2007; Dobbins & Han, 2006; Fletcher & Henson, 2001; Moscovitch and Winocur, 2002, Petrides, 2002, Rugg et al., 1998, Shallice, 2002; Simons & Spiers, 2003). There appears to be at least a rough consensus among these models that setting of retrieval mode and specification of cues are more dependent on VLPFC, whereas monitoring and verification are more dependent on DLPFC. Furthermore, according to the hemispheric encoding/retrieval asymmetry (HERA) model (Habib, Nyberg, & Tulving, 2003; Tulving, Kapur, Craik, Moscovitch, & Houle, 1994), left PFC activation should be greater for encoding than retrieval, whereas right PFC activation should be greater for retrieval than encoding.
Posterior parietal cortex may also play a role in episodic memory processes. Although the contribution of parietal regions to successful encoding has received relatively little attention (but see Uncapher, Otten, & Rugg, 2006), possible contributions to retrieval have garnered considerable interest. Retrieval success effects in posterior cortex were described as early as the mid-1990s (e.g., Kapur et al., 1995; Rugg, Fletcher, Frith, Frackowiak, & Dolan, 1996), but explicit theories about parietal involvement in episodic memory are only now being formulated (e.g., Cabeza, 2008; Cabeza, Ciaramelli, Olson, & Moscovitch, 2008; Ciaramelli, Grady, & Moscovitch, 2008; Vilberg & Rugg, 2008; Wagner, Shannon, Kahn, & Buckner, 2005). According to one view, retrieval-related activity in superior parietal regions tracks the task relevance or salience of cues, whereas activity in inferior parietal cortex is involved in recollection (Vilberg & Rugg, 2008). A related view is that superior parietal cortex facilitates top-down attentional control during retrieval, whereas inferior parietal cortex mediates bottom-up attentional processes (Cabeza, 2008, Cabeza et al., 2008, Ciaramelli et al., 2008).
The potential roles of other regions, including the cingulate gyrus and the cerebellum, are less well-specified, but because the first generation of functional neuroimaging studies tended to identify them as being involved in encoding and/or retrieval, interest in their specific contributions is growing.
A major development in the cognitive neuroscience of memory over the last decade has been the ascent of dual-process models positing that retrieval of episodic memories can be accompanied by a vivid sense of re-experiencing (“recollection”) or by a sense of familiarity. The view that these subjective experiences reflect the operation of two independent memory processes (e.g., Atkinson & Juola, 1974; Jacoby, 1991, Mandler, 1980, Tulving, 1985, Yonelinas, 1994) has received strong support from behavioural-experimental studies and from neuropsychological dissociations (see Yonelinas, 2002, for review). A growing literature has investigated the neural substrates of recollection and familiarity in healthy individuals using neuroimaging. This research has yielded additional support for the dual-process view (for recent reviews, see Diana, Yonelinas, & Ranganath, 2007; Eichenbaum, Yonelinas, & Ranganath, 2007; Mayes, Montaldi, & Migo, 2007; Skinner & Femandes, 2007; Vilberg & Rugg, 2008). With respect to neural correlates, interest has focused mainly on the contributions of medial-temporal sub-regions to the two processes. For example, recently some researchers (Diana et al., 2007, Eichenbaum et al., 2007; see also Bowles et al., 2007; Brown & Aggleton, 2001; Davachi, 2006) have suggested a correspondence between perirhinal cortex and item memory (familiarity), parahippocampal cortex and context memory (familiarity and recollection), and hippocampus and item-context associations (recollection; but see Rutishauser, Schuman, & Mamelak, 2008; Squire, Wixted, & Clark, 2007). In addition, various frontal and parietal sub-regions have been asserted to be differentially involved in recollection and familiarity (for reviews, see Skinner & Femandes, 2007; Vilberg & Rugg, 2008), though there is still considerable controversy about the precise contributions of these regions. One possible reason for the controversy is that two classes of experimental paradigms, objective vs. subjective, have been used to separate recollection and familiarity.
Objective recollection paradigms (also referred to as “relational-recognition tests”; Eichenbaum et al., 2007) include direct tests of memory for associations or contextual features. Tests of source memory are the most commonly used objective recollection tests. Associative recognition and the process-dissociation procedure also fall into the objective recollection category, but we will not discuss them further because relatively few neuroimaging studies have employed this paradigm.
In the class of subjective-recollection paradigms, the most widely used is the Remember–Know procedure (Tulving, 1985). In this procedure, participants are asked to indicate, for each item they classify as “old” on a recognition test, whether their memory is vivid and rich in contextual detail (“Remember”), or whether it is based on a non-specific sense of familiarity (“Know”). Another approach involves confidence judgments, which can be used to plot receiver operating characteristics (ROCs). The shape of the ROC is then analyzed to provide information about the contributions of recollection and familiarity to recognition performance (e.g., Yonelinas, 1994).
Behavioural studies of objective and subjective recollection have often provided converging results despite the differences in methodology and in the subjective experiences to which they give rise (Yonelinas, 2002; see also Davidson, Anaki, Saint-Cyr, Chow, & Moscovitch, 2006). Nevertheless, the two types of recollection have been dissociated in frontal (e.g., Ciaramelli & Ghetti, 2007; Duarte, Ranganath, & Knight, 2005; Levine, Freedman, Dawson, Black, & Stuss, 1999) and parietal lesion patients (e.g., Davidson et al., 2008, Simons et al., 2008), as well as in older adults (e.g., Duarte, Henson, & Graham, 2008). Duarte et al. (2008) described diverging fMRI activation patterns for objective and subjective recollection, but the focus of their analyses was on identifying age differences within each type of recollection, rather than on the main effect of recollection type. Indeed, to our knowledge, no previous studies have formally compared the neural correlates of objective and subjective recollection.
Several reviews and meta-analyses of ER-fMRI studies of memory have appeared recently, but these have either been limited to a small subset of data (Naghavi & Nyberg, 2005), specific brain regions (e.g., Ciaramelli et al., 2008, Henson, 2005; Vilberg & Rugg, 2008; Wagner et al., 2005), or the recollection-familiarity distinction (e.g., Eichenbaum et al., 2007, Mayes et al., 2007; Skinner & Femandes, 2007). Accordingly, we conducted a more comprehensive review of ER-fMRI studies of episodic memory.
Rather than performing a qualitative review, we aimed to complete the first quantitative meta-analysis of this literature. The benefits of a meta-analysis, in which data from the extant literature are formally integrated into an overall statistical analysis, are many. In particular, this method reduces the bias that can enter more qualitative reviews, and it minimizes the limitations inherent to individual studies that can make it relatively difficult to compare one with another: low statistical power stemming from the small number of participants in each study, variability in the labeling of neuroanatomical regions (e.g., Laird et al., 2005), and idiosyncratic fMRI methods varying from laboratory to laboratory (including differences in image acquisition, smoothing and other pre-processing steps, and statistical analyses, among other factors).
We used the activation likelihood estimation (ALE) method (Turkeltaub, Eden, Jones, & Zeffiro, 2002), in which, for a unidirectional contrast of interest (e.g., Remember–Know), each activation focus reported in the literature is modeled as the peak of a 3D Gaussian probability distribution. The ALE statistic, calculated as the sum of these probabilities across studies, indicates the likelihood that each voxel is active in the task. The ALE method is thus fully automated and quantitative, facilitating statistical inference via thresholding of the concordance maps (Chein, Fissell, Jacobs, & Fiez, 2002; Turkeltaub et al., 2002).
Our first objective was to synthesize results from the large number of studies that have reported successful-encoding and successful-retrieval contrasts in order to provide standard maps for both contrasts and to allow insight into neuroanatomical similarities and differences between encoding and retrieval networks. As stated above, we expected wide ranging activation across cortical and subcortical regions, but focused our attention on three key areas. First, early studies, and even many current ones, often failed to detect medial-temporal activation (see Henson, 2005). Thus, we were curious as to whether a meta-analysis would yield significant medial-temporal concordance, and, if so, whether it would be concentrated in hippocampus proper or would be more widely dispersed. Comparing encoding to retrieval analyses, we sought to determine whether medial-temporal activity would be concentrated in the anterior, middle, or posterior section of hippocampus, because various analyses have made conflicting claims about the function of the different regions (e.g., Greicius et al., 2003, Henson, 2005, Lepage et al., 1998, Ludowig et al., 2008, Parsons et al., 2006; Schacter & Wagner, 1999).
Second, although we expected widespread concordance in PFC for both encoding and retrieval, we were interested in determining potential differences in the relative involvement of PFC sub-regions during encoding compared to retrieval, as well as differences in the lateralization of PFC activation across studies using verbal and nonverbal materials, as predicted by the HERA model (e.g., Habib et al., 2003, Tulving et al., 1994). Third, although lateral and medial parietal activations during retrieval have been reported widely and several hypotheses as to their functional significance have been proposed (e.g., Ciaramelli et al., 2008; Vilberg & Rugg, 2008; Wagner et al., 2005), we sought to determine whether the same parietal regions were reliably activated in studies of encoding.
Our second objective was to meta-analyze the ER-fMRI literature on recollection, distinguishing between objective and subjective measures. As mentioned previously, the neural correlates of these measures have not been formally compared (but see Duarte et al., 2008). As with the more general analyses of encoding and retrieval, brain regions of particular interest included medial-temporal, frontal, and parietal cortex.
Within the medial-temporal region, we sought to establish whether hippocampal activation, thought to be a hallmark of recollective processing (for reviews, see Diana et al., 2007, Eichenbaum et al., 2007; Skinner & Femandes, 2007), would be seen for both objective and subjective recollection. For example, Duarte et al. (2008) reported posterior hippocampal activation, across younger and older participant groups, for subjective but not objective recollection.
Similarly, within PFC, we were interested in identifying sub-regions sensitive to objective and subjective recollection. We hypothesized that DLPFC and VLPFC regions involved in cognitive control processes would be less strongly activated in subjective than in objective recollection. “Remember” responses, for example, can be based on any contextual information that is vividly re-experienced by the participant (Duarte et al., 2008), without necessarily taxing strategic search and post-retrieval monitoring processes (Moscovitch & Winocur, 2002). In contrast, source memory tasks require participants to search for the information specified by the experimenter, and to monitor information retrieved from MTL structures (Dobbins & Han, 2006; Fletcher & Henson, 2001; Simons & Spiers, 2003). Indeed, patients with prefrontal lesions (Ciaramelli & Ghetti, 2007; Duarte et al., 2005) and older adults (Duarte et al., 2008) have been shown to be disproportionally impaired on tests of objective, but not subjective, recollection.
With respect to parietal cortex, it has been proposed that inferior lateral parietal cortex supports mnemonic decisions involving bottom-up attentional capture by the result of a memory search, whereas superior lateral parietal cortex and the intraparietal sulcus track top-down attention during retrieval (Cabeza, 2008, Ciaramelli et al., 2008). Both objective and subjective recollection may engage bottom-up attentional capture by retrieved memory contents, although Duarte et al. (2008) found inferior parietal activation, across younger and older adults, for subjective recollection only. The strategic, top-down component, on the other hand, should be taxed more heavily by objective recollection, which involves search for a specific, experimenter-provided contextual detail (e.g., the spatial location in which an item was presented during encoding). We thus predicted greater involvement of superior parietal cortex in objective recollection, compared to subjective recollection.
Section snippets
Study selection
We conducted a Pubmed (www.pubmed.org) query using the keyword search “(memory OR recognition OR recall) AND fMRI.” We selected studies that were published in 2007 or earlier, used event-related fMRI, reported standard-space stereotactic coordinates of whole-brain activation maxima for at least one of the contrasts of interest (see below), and used a univariate fMRI analysis approach with uniform significance and cluster size thresholds applied throughout the brain. We excluded data from
Encoding success
Table 5 and Fig. 1 show the results for the ALE analysis of encoding success studies. Although 12 of the 21 significant clusters were located in the right hemisphere, the greatest concordance in terms of size of region and peak ALE value was observed in the left hemisphere, most notably in dorsolateral and ventrolateral PFC. Other major left-hemisphere clusters were present in parahippocampal gyrus/anterior hippocampus, in fusiform gyrus and neighboring occipitotemporal areas, and in
Discussion
The ALE meta-analyses revealed robust whole-brain spatial activation patterns related to episodic memory processes, including successful encoding, successful retrieval, and recollection (objective and subjective). In each case, we observed concordant activations across studies that differed in many methodological aspects (experimental paradigms, stimulus types, imaging protocols, etc.). As expected, medial-temporal, prefrontal, and parietal regions contributed the greatest number of significant
Conclusions
Consistent with previous reviews and meta-analyses (e.g., Ciaramelli et al., 2008; Gilbert et al., 2006, Henson, 2005, Naghavi and Nyberg, 2005; Skinner & Femandes, 2007; Vilberg & Rugg, 2008; Wagner et al., 2005), our meta-analysis revealed wide swaths of memory-related activity across temporal, frontal, parietal, and other regions of the brain. Because of the focus of our hypotheses, and the large number of individual results, we concentrated on medial-temporal, prefrontal, and parietal
Acknowledgments
This research was initiated when authors JS and PSRD were at the Rotman Research Institute, and continued while the latter was at the University of Alberta. We thank Elisa Ciaramelli for helpful comments. We are grateful to the Faculty of Arts at Ryerson University and the Faculty of Science at the University of Alberta for grant support.
References (155)
- et al.
Dorsolateral prefrontal cortex involvement in memory post-retrieval monitoring revealed in both item and associative recognition tests
NeuroImage
(2005) - et al.
Parietal contributions to recollection: Electrophysiological evidence from aging and patients with parietal lesions
Neuropsychologia
(2008) - et al.
Left ventrolateral prefrontal cortex and the cognitive control of memory
Neuropsychologia
(2007) - et al.
Hippocampal-prefrontal encoding activation predicts whether words can be successfully recalled or only recognized
Behavioural Brain Research
(2006) - et al.
Functional-anatomic study of episodic retrieval: II. Selective averaging of event-related fMRI trials to test the retrieval success hypothesis
NeuroImage
(1998) - et al.
The gateway hypothesis of rostral prefrontal cortex (area 10) function
Trends in Cognitive Sciences
(2007) Role of parietal regions in episodic memory retrieval: The dual attentional processes hypothesis
Neuropsychologia
(2008)- et al.
Word frequency and subsequent memory effects studied using event-related fMRI
NeuroImage
(2003) - et al.
Functional heterogeneity within Broca's area during verbal working memory
Physiology and Behavior
(2002) - et al.
What are confabulators’ memories made of? A study of subjective and objective measures of recollection in confabulation
Neuropsychologia
(2007)