Abstract

Autobiographical memory (AM) provides the opportunity to study interactions among brain areas that support the search for a specific episodic memory (construction), and the later experience of mentally reliving it (elaboration). While the hippocampus supports both construction and elaboration, it is unclear how hippocampal–neocortical connectivity differs between these stages, and how this connectivity involves the anterior and posterior segments of the hippocampus, as these have been considered to support the retrieval of general concepts and recollection processes, respectively. We acquired fMRI data in 18 healthy participants during an AM retrieval task in which participants were asked to access a specific AM (construction) and then to recollect it by recovering as many episodic details as possible (elaboration). Using multivariate analytic techniques, we examined changes in functional and effective connectivity of hippocampal–neocortical interactions during these phases of AM retrieval. We found that the left anterior hippocampus interacted with frontal areas during construction and bilateral posterior hippocampi with visual perceptual areas during elaboration, indicating key roles for both hippocampi in coordinating transient neocortical networks at both AM stages. Our findings demonstrate the importance of direct interrogation of hippocampal–neocortical interactions to better illuminate the neural dynamics underlying complex cognitive tasks such as AM retrieval.

Introduction

Autobiographical memory (AM) retrieval involves a variety of mental processes, which include searching for and accessing specific life episodes (the construction stage), and recollecting them by reassembling vivid episodic details (the elaboration stage). These cognitive processes are supported by a number of neocortical areas as well as by the hippocampus, which might play a key role in initiating and coordinating their engagement. To date, however, there is little experimental evidence regarding the nature of this hippocampal–neocortical interplay during the different stages of AM retrieval. Here, we examine functional and effective connectivity of hippocampal–neocortical networks during AM construction and AM elaboration.

In an influential model, Conway (2009) organizes AM into nested hierarchical levels of self-knowledge, from the most conceptual and general knowledge of life periods or events (e.g., my teenage years) to the most concrete and contextually specific life episodes (e.g., the song playing during my first kiss). Memories of specific life events are unique because they contain concrete and vivid episodic elements (Conway and Pleydell-Pearce 2000; Conway 2009), which allow us to recollect past events in a vivid manner (Tulving 2002). According to this model, when an event-specific AM is prompted with a thematic cue, such as “kiss” or “party” (a typical experimental setting of AM retrieval), the first level of entry into AM knowledge is typically through memory for repeated or extended life events. In this case, one usually accesses episodic elements only after accessing more general and abstract personal knowledge. In other words, retrieving a specific autobiographical episode typically involves searching through general knowledge (akin to the construction stage), and then reliving the episode vividly by accessing episodic elements (the elaboration stage). In this sense, AM retrieval can be seen as a particular instance of cued contextually rich episodic memory retrieval that offers the unique opportunity to study construction and elaboration processes.

Although there is very little information about how medial temporal and neocortical regions interact with one another in the different stages of AM retrieval, there are some data regarding the general regions involved in this two-process model. Based on an early study using electroencephalography, Conway speculated that a frontotemporal network would support construction and a more diffuse occipitotemporoparietal network would support the multimodal re-experience of episodic elements underlying elaboration (Conway et al. 2001). Although these predictions are generally supported by neuroimaging studies showing the importance of the frontal lobes to memory retrieval processes (St. Jacques et al. 2011; Addis et al. 2012) and the importance of the visual perceptual areas to visual imagery (Rubin and Greenberg 1998; Greenberg et al. 2005), research focusing on the differentiation between construction and elaboration of AMs did not show the clear anterior–posterior dissociation that Conway predicted more than 10 years ago (Holland et al. 2001; Addis et al. 2007; Rabin and Rosenbaum 2012 but see Daselaar et al. 2008). However, none of these previous studies focused specifically on hippocampal–neocortical connectivity, although St Jacques et al. (2011) used independent component analysis (ICA) to identify networks involved in different phases of AM retrieval.

With the current work, we explored the possibility that the hippocampus might be the hub that coordinates these anterior and posterior networks during AM retrieval. This proposal is based on functional magnetic resonance imaging (fMRI) studies showing sustained hippocampal activation during AM retrieval, indicating a pivotal role for the hippocampus in both construction and elaboration (Addis et al. 2007; St Jacques et al. 2011). Also, neuroanatomical studies show that the hippocampus is perfectly situated between and strongly connected to frontal cortices and visual perceptual areas, that is, all areas supposedly involved in AM construction and elaboration (Lavenex and Amaral 2000; Lavenex et al. 2002). Lastly, a connectivity-based neuroimaging study from our group showed that hippocampal–neocortical networks support encoding and retrieval of episodic memories (McCormick et al. 2010), thus strengthening the idea that the hippocampus forms transient links between neocortical brain regions that are necessary for contextually rich memory processing (Teyler and DiScenna 1986; Buzsaki 1996; Teyler and Rudy 2007; Rugg et al. 2008). Here, we tested the hypothesis that hippocampal–neocortical connectivity shifts from an early hippocampal–frontal network at construction to a late hippocampal–occipitoparietal network at elaboration during AM retrieval.

Growing evidence suggests that the hippocampus can be segregated into functionally distinct segments along its long axis (Fanselow and Dong 2010; Poppenk and Moscovitch 2011; Viard et al. 2012; Bonnici et al. 2013; Poppenk et al. 2013). Together with a recent review aimed to map different cognitive systems onto anterior and posterior memory networks (Ranganath and Ritchey 2012), this formulation suggests that the anterior part of the hippocampus, possibly in interaction with frontal areas, might support cue-specific retrieval attempts and activation of general concepts, functions that are needed for AM construction. In contrast, the posterior hippocampus might support recovery of detailed perceptually rich episodic memories through communication with visual perceptual areas, suggesting a potential role in AM elaboration. However, experimental evidence for this dissociation remains thin, particularly in the context of AM, and, so in the current study, we specifically assessed anterior and posterior hippocampal connectivity.

We aimed to map both functional and effective connectivity among brain regions that support construction and elaboration of AM retrieval. We were specifically interested in characterizing how frontotemporal and occipitotemporoparietal regions interact at each phase with the anterior and posterior hippocampus, which we hypothesized to play functionally distinct roles in AM retrieval.

Materials and Methods

Participants

Eighteen healthy controls (7 women, 11 men, 15 right-handed, 3 left-handed) with no history of neurological or psychiatric disorders participated in this study and gave written informed consent in accordance with a research protocol approved by the University Health Network Ethics Board. Age ranged from 22 to 61 with an average of 36.17 ± 12.3. This broad age range was chosen because we recruited participants age-matched to a patient group whose data are not part of the current analysis.

Experimental Procedure

Immediately before the scanning procedure, the experimental task was explained to the participants and they completed 6 practice trails of each condition. The contents of all retrieved AMs were then probed to confirm that participants understood the instructions (i.e., the generated events were specific in time and place).

The experimental task consisted of 2 different conditions, AM retrieval and math, with 22 trials of each condition. The duration of each trial was 16.5 s and trials were presented in a randomized fashion, with a jittered interstimulus interval (ISI) of 0.5, 1.0, or 1.5 s. In the AM retrieval condition, participants were presented with an event cue such as “a party.” The participants were asked to recall a specific event from their personal past that was coherent with the cue and to press a response button once they retrieved a specific incident. During the remaining seconds of the trial, they were instructed to elaborate the details of that particular memory, and to relive the event in their mind's eye. The button press indicated the end of the construction phase, and the beginning of the elaboration phase (see Addis et al. 2007 for a similar paradigm). In the math condition, participants were asked to solve a simple math problem, for example, 19 + 4, and press a response button once they had the solution in mind. During the remaining seconds of the trial, participants were instructed to add 3s to the solution (e.g., 23 + 3 + 3 ⋯ ). At the end of each trial, participants were asked to rate the memory as vivid or faint and the math problem as easy or hard. Owing to technical difficulties, the responses of 4 participants were not recorded during the task; however, in a debriefing after the scan, they confirmed that they followed the instructions throughout the task and we therefore included their fMRI data in the analyses.

Data Acquisition

Anatomical and functional data were acquired on a 3-T Sigma MR System (GE Medical Systems, Milwaukee). Anatomical scans, for co-registration of functional data, were acquired first (T1-weighted sequence, 120 slices, FOV = 220 mm, slice thickness = 1 mm, 0 gap, 256 × 256 matrix, resulting in a voxel size of 0.9 × 0.9 × 1.0). The functional data were acquired in an interleaved order (EPI, TR = 2 s; 30–32 slices to cover the whole brain, FOV = 240 mm, slice thickness = 5 mm, 0 gap, 64 × 64 matrix, resulting in a voxel size of 3.75 × 3.75 × 5.0). Functional images were taken in an oblique orientation with each slice being perpendicular to the long axis of the hippocampus. We acquired 2 functional sessions with 190 frames each. The first 3 frames were dropped for signal equilibrium. The fMRI protocol also included 2 other experimental tasks that are not part of the current analysis.

Data Analysis

All preprocessing of imaging data was performed using SPM8 (Statistical Parametric Mapping 8; Wellcome Department of Imaging Neuroscience, London). Functional images were co-registered to the subject's anatomical image, and then realigned and unwarped. The subject's anatomical image was segmented and spatially normalized to the T1-weighted Montreal Neurological Institute (MNI) template and the normalization parameters were then written to the functional data. Finally, fMRI data were smoothed using a Gaussian kernel of 8 mm full-width half-maximum (FWHM). SPM motion parameters were inspected for outliers (motion > 2 mm in any direction) but no subjects had to be excluded from the analysis.

Spatiotemporal Partial Least Squares

A spatiotemporal Partial Least Squares (ST-PLS) analysis was conducted in order to examine the time course of hippocampal activity during AM retrieval. PLS is a multivariate, correlational technique that allows the analyses of associations between brain activity and experimental conditions without assumptions about the shape and time course of the hemodynamic response function (McIntosh et al. 2004; McIntosh and Lobaugh 2004; Krishnan et al. 2011). Therefore, ST-PLS allows the investigation of changes in event-related brain activity for each TR of a trial event. In the current analysis, we examined time windows of 8 one-TR frames (i.e., lags) for all AM and math trials.

Detailed descriptions of PLS can be found elsewhere (Krishnan et al. 2011). In brief, PLS uses singular value decomposition (SVD) to extract ranked latent variables (LVs) from the covariance matrix of brain activity and conditions. These LVs express patterns of brain activity associated with each condition. When applying PLS to event-related fMRI data, patterns of brain activity are calculated for each lag, providing a time course of activity associated with the experimental conditions. Statistical significance of the LVs was assessed using permutation testing. In this procedure, each subject's data were randomly reassigned (without replacement) to different experimental conditions, and a null distribution was derived from multiple permutated solutions. In the current experiment, we used 500 permutations and considered LV as significant if P < 0.05. Further, we assessed the reliability of each voxel that contributed to a specific LV's activity pattern using a bootstrapped estimation of the standard error (i.e., bootstrap ratio, BSR). For each bootstrapped solution (100 in total), subjects were sampled randomly with replacement and a new analysis was performed. In the current study, we considered clusters of 10 or more voxels with BSRs > 3.00 (roughly equal to a P < 0.003) to represent reliable patterns of activation.

Seed Partial Least Squares

In order to examine whether hippocampal–neocortical interactions change between AM construction and elaboration, we conducted a seed PLS analysis. Seed PLS examines the relationship between a target region (seed voxel) and signal intensities in all other brain voxels as a function of the experimental conditions over time (Krishnan et al. 2011). The general difference to ST-PLS is that, in seed PLS, the covariance matrix used for SVD stems from correlation values between the seed voxel and all other voxels for each experimental condition (i.e., rather than brain activity values per condition).

We selected a seed in the left anterior hippocampus (ant lHC, MNI: −20 −10 −22) because this region differentiated AM retrieval from the math task reliably (as indicated by its high BSR) from lag 2 onward and showed sustained activation throughout AM retrieval (see Fig. 1, and Supplementary Table 1 for hippocampal peak voxels). Of note, although we were interested in a possible anterior–posterior hippocampal dissociation, we chose a single seed for the seed PLS analysis in order to differentiate functional connectivity between 2 time points (reflecting construction and elaboration) and not between 2 hippocampal regions. We therefore extracted signal intensities from the same seed, the left anterior hippocampus, at an early time point (lag 2) of AM retrieval, at which participants were searching for a specific AM (construction), and a later time point (lag 6), at which participants continued recovery and mentally “replayed” episodic details of the event (elaboration). These lags were selected as they fell comfortably within the timeframes of the construction and the elaboration phases for most items and most subjects, as indicated by participants button presses (see Behavioral results). We used a nonrotated version of seed PLS which allowed us to prespecify a contrast (as opposed to mean-centered ST-PLS, which identifies contrasts among conditions in a data-driven manner) between hippocampal functional connectivity at lag 2 and lag 6. This contrast differentiates brain regions whose activity correlates strongly with hippocampal activity at lag 2 from brain regions whose activity correlates strongly with hippocampal activity at lag 6. This pattern does not exclude the possibility that some brain regions correlate with hippocampal activity at both lag 2 and lag 6; however, the correlation at 1 time point has to be stronger than at the other time point. Again, we used 500 permutations to assess significance of the LV and 100 bootstrap samples to assess reliable voxels. We considered clusters of 5 or more voxels with a BSR > 2.00 (roughly equal to a P < 0.04). Of note, seed PLS calculates correlations between seed activity at prespecified time points (here lags 2 and 6), and activity in all remaining brain voxels at each time point within a trial's time window (here 8 lags). In our case, only voxels whose activity at lag 2 correlated with lag 2 hippocampal activity and voxels whose activity at lag 6 correlated with lag 6 hippocampal activity were meaningful to our hypothesis. We felt that this limitation justified the choice of a lower statistical threshold than that adopted for the ST-PLS analysis.

Figure 1.

Temporal characteristics of hippocampal activation during AM retrieval. (A) Hippocampal activation (slice number: y = −10) in red across 7 lags (=14 s) that showed a greater bootstrap ratio (BSR) than 3.0 and a cluster size of >10 contiguous voxels. Activation is displayed on a T1-weighted MRI. (B) Percent signal change for the left (lHC, MNI: −20 −10 −22; peak in lag 2) and right (rHC, MNI: 26 −16 −18; peak in lag 3) hippocampus during AM retrieval (memory) and the control task (math).

Figure 1.

Temporal characteristics of hippocampal activation during AM retrieval. (A) Hippocampal activation (slice number: y = −10) in red across 7 lags (=14 s) that showed a greater bootstrap ratio (BSR) than 3.0 and a cluster size of >10 contiguous voxels. Activation is displayed on a T1-weighted MRI. (B) Percent signal change for the left (lHC, MNI: −20 −10 −22; peak in lag 2) and right (rHC, MNI: 26 −16 −18; peak in lag 3) hippocampus during AM retrieval (memory) and the control task (math).

Structural Equation Modeling

While our preceding seed PLS analysis was restricted to functional connectivity from a single seed, the left anterior hippocampus, here we aimed to explicitly examine connectivity from anterior and posterior regions in the hippocampus with a specific emphasis on the different phases of AM retrieval. To do this, we used structural equation modeling (SEM; LISREL 8.80, Student Edition, Scientific Software, Inc., Mooresville, IN, USA), which examines interregional correlations and anatomical pathways among selected brain areas as the input to compute path coefficients (see McIntosh and Gonzalez-Lima (1994) for detailed description of SEM for neuroimaging data). These path coefficients provide information about the strength and directionality of influences between 2 areas and can, in distinction from symmetrical correlation analysis such as seed PLS, differ between 2 connected regions as a function of other influences in the model (McIntosh and Protzner 2012).

Region Selection

Our region selection for the SEM model was based on the highest bootstrap ratios and cluster size of the seed PLS analysis as well as functional relevance to AM retrieval (Greenberg and Rubin 2003; Conway 2009). Eleven voxels from regions whose activity co-varied with the left anterior HC were included in this model (abbreviations and MNI coordinates in brackets): bilateral anterior HC (ant lHC = −20 −10 −22; ant rHC = 28 −8 −16), bilateral posterior HC (post lHC = −24 −38 −2; post rHC = 26 −38 −2), left dorsomedial prefrontal cortex (ldmPFC = −6 50 44), left ventrolateral prefrontal cortex (lvlPFC = −50 36 −10), left medial prefrontal cortex (lmPFC = −4 56 −14), bilateral middle occipital cortices (lmidOcc = −22 −72 48; rmidOcc = 30 −76 36), left lingual gyrus (lLingual = −24 −72 2), and right fusiform gyrus (rFusiform = 22 −78 36).

Model Construction

An anatomical model of multisynaptic connections between these regions was derived from known primate neuroanatomy (Suzuki and Amaral 1994; Lavenex et al. 2002; Kondo et al. 2005; Fanselow and Dong 2010; Ranganath and Ritchey 2012). As we were especially interested in the role of the hippocampus during AM construction and elaboration, we only included hippocampal–neocortical connections and neglected corticocortical connections. We then constructed a functional model for AM retrieval during construction and elaboration. For each individual, the voxel signal intensities were extracted from each chosen region from lag 2 (construction) and lag 6 (elaboration) using the PLSgui.

Path Analysis

For the path analysis, the matrix of correlations between the extracted signal intensities was used to calculate path coefficients, which represent the magnitude of influence of each directional path. Using a stacked-model approach, we tested for differences in effective connectivity between AM construction and elaboration (McIntosh and Gonzalez-Lima 1994). This approach uses 2 models, a null and an alternative model. In the null model, path coefficients were set to be equal, whereas in the alternative model, path coefficients were free to vary. In order to test whether the null or alternative model provides a better fit to the data, we conducted a goodness-of-fit χ2 test for each model. Smaller χ2 values indicate better model fit. We therefore used the χ2 difference test to evaluate whether the difference between both χ2 values was significant. In our case, the alternative model had a significantly lower χ2 value than the null model, and thus provided a better fit to the data. To determine which connections contributed to the better fit of the alternative model, individual paths were allowed to vary in a stepwise manner and χ2 values were re-calculated each time. These χ2 values were then evaluated against the χ2 value of the null model using the χ2 difference test. Connections that increased the model fit (i.e., lowered the χ2 value significantly) were considered significant and were left to vary in the model. Because the order in which the connections vary could influence the results, we allowed connections to vary in 4 different orders (i.e., from anterior to posterior, from posterior to anterior, from the left to the right hemisphere, and from the right to the left hemisphere). The order that resulted in the smallest remaining χ2 distance (which was the model that began by anterior connections to vary) was chosen as the best fit. Models with stability indices below or equal 1 were considered sufficiently stable (Kline 2005).

Results

Behavioural Results

Participants indicated on average after 2.9 s (±0.7 s) that they had successfully retrieved a specific AM and would begin the elaboration phase, whereas they responded on average 3.9 s (±1.1 s) to indicate that they solved the math problem and would start adding 3s to the solution (t26 = 2.92, P < 0.01). Of the 22 AM trials, participants rated 14 (±4.4) as vivid and 16.7 (±3.9) of the 22 math trials as easy (t26 = 1.72, P > 0.05).

Early Hippocampal Activation

To examine the time course of hippocampal activity during AM retrieval, we conducted a ST-PLS. PLS identified 1 significant pattern, separating AM and math trials (LV 1, P < 0.001). The pattern describing the AM retrieval comprised all of the core and most of the secondary AM regions identified by a meta-analysis (Svoboda et al. 2006). Please see Supplementary Tables 1 and 2a and b for a list of hippocampal peaks and other activated brain regions. Of special interest, we found that the left anterior hippocampus (MNI, −20 −10 −22) was activated one lag earlier than the right hippocampus during AM trials (see Fig. 1). For the remaining lags, both hippocampi showed greater activation during AM than math trials, replicating previous findings of sustained hippocampal activation throughout AM retrieval (Addis et al. 2007; St Jacques et al. 2011).

Functional Hippocampal–Neocortical Connectivity During AM Retrieval

In this analysis, we examined functional connectivity from the same region, the left anterior hippocampus, at two different time points during AM retrieval that corresponded to the construction and elaboration stages. This analysis revealed one significant pattern that separated construction and elaboration (LV 1, P = 0.03; see Fig. 2 and Supplementary Table 3 for all correlated brain regions). During construction, activation of the left anterior hippocampus was correlated with a small set of brain regions (in total 2313 voxels), including the right anterior hippocampus and mainly left frontotemporal regions. During elaboration, widespread brain activation (in total 26 683 voxels) correlated with activity in the left anterior hippocampus. These regions included visual perception cortices, auditory association cortices, and motor cortices. Of note, we did not see this pattern in a direct univariate contrast of mean activity (rather than connectivity) of early/construction versus late/elaboration phases (see Supplementary Fig. 1).

Figure 2.

Functional hippocampal–neocortical networks during AM retrieval. Functional connectivity between the hippocampal seed (lHC, MNI: −20 −10 −22) and other brain regions during construction (= lag 2) and elaboration (= lag 6) displayed on a rendered T1-weigthed MRI. BSR > 2.0 and clusters of more than 5 contiguous voxels were considered significant.

Figure 2.

Functional hippocampal–neocortical networks during AM retrieval. Functional connectivity between the hippocampal seed (lHC, MNI: −20 −10 −22) and other brain regions during construction (= lag 2) and elaboration (= lag 6) displayed on a rendered T1-weigthed MRI. BSR > 2.0 and clusters of more than 5 contiguous voxels were considered significant.

These findings show that the same hippocampal region communicates with different cortical regions during construction and elaboration. We then used effective connectivity to determine whether the directionality of hippocampal–neocortical interactions differs during the 2 phases.

Directed Hippocampal–Neocortical Connectivity During AM Retrieval

One of our main research goals was to assess the role of the hippocampus as a central hub within the AM retrieval network. Further, we aimed to assess interactions between the anterior and posterior hippocampus and the rest of the AM retrieval network during construction and elaboration, respectively. We therefore built a model of the AM retrieval network based on brain regions found to be involved in construction or elaboration in our seed PLS analysis (see Fig. 3 for all regions included in the SEM analysis and their correlation values to the hippocampal seed), which included the anterior and posterior hippocampi bilaterally, and we examined the directionality of these connections during each AM retrieval phase.

Figure 3.

Location of SEM nodes and correlation to the seed. In addition to the seed voxel in the ant lHC, 10 regions were included in the following SEM analysis. For clarity, individual clusters are displayed separately on a T1-weighted MRI. The functional connectivity (correlation coefficient, r) between the seed and the peak voxel of the cluster during construction (Con; lag 2) and elaboration (Ela; lag 6) is shown underneath each region.

Figure 3.

Location of SEM nodes and correlation to the seed. In addition to the seed voxel in the ant lHC, 10 regions were included in the following SEM analysis. For clarity, individual clusters are displayed separately on a T1-weighted MRI. The functional connectivity (correlation coefficient, r) between the seed and the peak voxel of the cluster during construction (Con; lag 2) and elaboration (Ela; lag 6) is shown underneath each region.

We found that the model in which path coefficients were free to vary was a better fit to our data than the model in which path coefficients were fixed (CHIdiff = 44.72, df = 22, P = 0.003, all stability indices ≤1), indicating that effective connectivity differed significantly between construction and elaboration (see Fig. 4). We found 6 connections that differed between the 2 AM phases (path coefficients in brackets): During construction, the left anterior hippocampus had a greater positive influence on the dmPFC (ant lHC to ldmPFC, construction: 0.7, elaboration: 0.06) and the right anterior hippocampus (ant lHC to ant rHC, construction: 0.6, elaboration: 0.35) than during elaboration. Further, both left and right anterior hippocampi had a positive influence on activity in the posterior part of the hippocampi, while this influence was slightly negative during elaboration (ant lHC to post lHC, construction: 0.12, elaboration: −0.1; ant rHC to post rHC construction: 0.21, elaboration: −0.02).

Figure 4.

Effective hippocampal–neocortical networks during AM retrieval. Red arrows represent positive (solid) or negative (dashed) effective connections that differed between construction and elaboration, white arrows represent anatomical connections that were included in the model but did not differ between both AM retrieval processes. The thickness of the arrows indicates the strength of the influence from one region to the other (path coefficient).

Figure 4.

Effective hippocampal–neocortical networks during AM retrieval. Red arrows represent positive (solid) or negative (dashed) effective connections that differed between construction and elaboration, white arrows represent anatomical connections that were included in the model but did not differ between both AM retrieval processes. The thickness of the arrows indicates the strength of the influence from one region to the other (path coefficient).

During elaboration, the left posterior hippocampus had a greater influence on the left middle occipital gyrus (post lHC to lmidOcc, construction: 0.05, elaboration: 0.49), while the right posterior hippocampus had a greater influence on the right fusiform gyrus than during construction (post rHC to rFusiform, construction: −0.02, elaboration: 0.29). These findings indicate key roles for both hippocampi in the directed mediation of hippocampal–neocortical interactions during AM retrieval that shift from a hippocampal–frontal network at construction to a hippocampal–temporo-occipital network at elaboration.

Discussion

We report that the hippocampus is critically involved in coordinating different networks for constructing and elaborating autobiographical memories. Using multivariate connectivity analyses, our data provide direct and compelling experimental evidence that construction is supported by a frontotemporal network and elaboration by a widespread temporoparietooccipital network and further that these networks interact with anterior and posterior hippocampal regions, respectively. Importantly, these networks were not readily identified by voxel-based analyses, indicating that network approaches are particularly suited to reveal interactions among regions involved in complex cognitive processes such as AM retrieval. Recently, St Jacques et al. (2011) used another multivariate technique, ICA, and extracted four different functional networks underlying AM construction and elaboration, including MTL, medial PFC, frontoparietal, and cingulo-operculum networks. Although they did not impose a specific contrast between networks supporting construction and elaboration, they found that the frontoparietal and cingulo-operculum networks peaked during construction and declined during elaboration whereas the MTL and medial PFC networks showed sustained activation throughout the AM retrieval. Thus, our findings are complementary and we show additionally those hippocampal–neocortical networks (identified by direct contrast) that are uniquely associated with elaboration.

AM Construction: Hippocampal–Frontotemporal Connectivity

As Conway predicted more than 10 years ago (Conway and Pleydell-Pearce 2000), we found that AM construction was supported by a small network comprising the left hippocampus and mainly frontotemporal cortices, including both medial and lateral regions of the frontal and temporal lobes. Our results are in agreement with other neuroimaging studies showing frontotemporal involvement during construction (Daselaar et al. 2008; Holland et al. 2011; St Jacques et al. 2011; Rabin and Rosenbaum 2012). However, we expand this knowledge by showing that these regions form a transient network linked to the left anterior hippocampus and furthermore that the left anterior hippocampus effectively influences activity in the dmPFC during AM construction. Although the dmPFC has been shown to be a consistent part of both the network engaged during AM retrieval (Svoboda et al. 2006) and the resting state default mode network (Buckner et al. 2008), several studies have shown it can segregate from the hippocampal network in both settings (Andrews-Hanna et al. 2010; St Jacques et al. 2011). Nonetheless, we show here that the hippocampus can form transient links to this brain region during specific task circumstances. While the specific contribution of the dmPFC to autobiographical retrieval is unclear, it has been shown to be engaged in self-guided retrieval of semantic information (Binder et al. 2009), which is consistent with Conway's idea of AM construction. Of interest, our SEM finding that hippocampal activation influences the activation pattern in dmPFC, rather than the reverse, is new and interesting as it suggests a strong “bottom-up” direction of influence in construction versus elaboration. However, it might be that this direction of influence holds true only for readily accessible AMs. Using a similar effective connectivity analysis (dynamic causal modeling), St Jacques et al. (2011) showed that the MTL influenced the PFC during AM construction only in cases where the AM was retrieved quickly but not for AMs that required more effort to bring to mind. The relatively short mean retrieval time of our participants (our study: 2.9 ± 0.7 s, St Jacques et al.: 6.5 ± 2.11 s) indicates that the AMs in our study were also easily accessible which might have contributed to this result.

In contrast, the engagement of medial parts of the prefrontal cortex has been attributed to self-referential processes (Howe and Courage 1997; Conway and Pleydell-Pearce 2000). It is therefore not surprising that in our data functional and effective connectivity between the hippocampus and mPFC did not differentiate between construction and elaboration, since both processes rely on self-referential information (Svoboda et al. 2006).

AM Elaboration: Hippocampal–Temporoparietooccipital Connectivity

Again, as Conway predicted, we found that AM elaboration was supported by a widespread network comprising bilateral anterior and posterior regions of the hippocampi, as well as temporal, parietal and occipital cortices (Conway and Pleydell-Pearce 2000). Strikingly, our effective connectivity analysis revealed that both posterior hippocampi actively influenced activity in visual perceptual areas. Whereas the retrieval of visual, emotional and auditory information is considered a crucial part of vivid AM elaboration and should be supported by a widespread network comprising most primary and associative cortices (Greenberg and Rubin 2003; Svoboda et al. 2006; Conway 2009), there is surprisingly little experimental evidence that such regions actually contribute to AM elaboration. In fact, the majority of studies that examined the spatiotemporal characteristics of AM retrieval did not observe widespread activation in visual perceptual areas during elaboration (Addis, McIntosh, et al. 2004; Addis et al. 2007; Holland et al. 2011; Rabin and Rosenbaum 2012). Only 1 study reported greater activation in visual cortices during elaboration than construction (Daselaar et al. 2008) but even that study did not show the widespread posterior activation predicted by various AM theories (Brewer 1986; Conway and Pleydell-Pearce 2000; Greenberg and Rubin 2003; Conway 2009). A reason for this discrepancy might be that the above-mentioned studies used mean-based activation to extract differences between either AM construction and elaboration, or AM retrieval and some control task (see St Jacques et al. 2011 for an exception). We reason that we were able to reveal the full extent of the elaboration network by using connectivity rather than mean-based measurements. That is, typical task-based analyses using t-tests focus solely on whether the mean activation in a given brain region differs between 2 experimental conditions, while connectivity analyses evaluate how 2 brain regions co-vary in different experimental conditions. In fact, it is entirely possible that the experimentally induced difference in regional mean activity (i.e., main effect) is not significant, but that the change in the relationship between 2 regions due to experimental manipulation (i.e., interaction effect) assessed by connectivity analyses is significant. This finding would be missed if the covariance of the 2 brain regions was not evaluated (McIntosh and Gonzalez-Lima 1994). Our data support this notion, in that a direct contrast between AM construction and elaboration that was based on mean activation showed very little specific engagement of posterior regions during elaboration, whereas the covariance-based seed PLS revealed a widespread elaboration network (see Supplementary Fig. 1). Although St Jacques et al. (2011) also used a covariance-based method, ICA, they did not find a posterior network that peaked during elaboration. One potential reason for this discrepancy is that we specified a direct contrast in our nonrotated seed PLS between hippocampal connectivity during construction versus elaboration. This aspect forced the analysis to assess differences between the 2 conditions which may have been missed by a data-driven ICA analysis.

The Role of the Hippocampus in AM Retrieval

The functional importance of the hippocampus in AM retrieval, and generally contextually rich episodic memory retrieval, is well established (Svoboda et al. 2006; Ranganath 2010). Consistent with other studies, we found sustained hippocampal activation throughout AM retrieval trials (Addis et al. 2007; but see Daselaar et al. 2008; St Jacques et al. 2011). However, in this article, we extend beyond previous findings by demonstrating how the hippocampus plays a crucial role in coordinating construction and elaboration networks. Although other studies have shown that the hippocampus can form parts of different memory networks during encoding versus retrieval (McCormick et al. 2010), and during episodic versus semantic retrieval (Rajah and McIntosh 2005), to our knowledge, this is the first demonstration of the hippocampus connecting with separate sets of brain regions during different stages of the same AM trials.

Further, in our data, the left anterior hippocampus was functionally connected to the right anterior hippocampus during construction and to bilateral posterior hippocampi during elaboration. Whereas involvement of the anterior hippocampus has been attributed to access of specific AMs and gist-like episodic memories, involvement of the posterior hippocampus has been linked to recollection, memory for complex scenes and spatial navigation (Poppenk and Moscovitch 2011; Viard et al. 2012). For example, Poppenk et al. (2013) propose that, in AM retrieval, the role of the anterior hippocampus might be to link together principal actors, actions, and general settings, and the role of the posterior hippocampus might be to retrieve the exact spatiotemporal context and episodic details around this event. Although these theories emphasize the anatomical and functional dissociation between anterior and posterior parts of the hippocampus, there are major anatomical and functional connections between these segments, which might facilitate integration of different information (Patel et al. 2012; Sloviter and Lomo 2012). For example, Sloviter and Lomo (2012) reviews that the longitudinal associational axon projections of the mossy fibers in the dentate gyrus of the rat spans around 6.6 mm (roughly 80% of the total length of the rat hippocampus). Further, a recent lesion study in rats pointed toward the intermediate hippocampus as an important site for the integration of behavioral–control processes (i.e., associated with anterior hippocampal function) and place codes (i.e., associated with posterior hippocampal function, Bast et al. 2009). Our SEM results revealed that both anterior segments positively influenced posterior segments of the hippocampi during construction to a greater extent than during elaboration. In our experiment, AMs were brought voluntarily to mind using event cues. We therefore speculate that during construction, gist-like information indexed by the anterior part of the hippocampus might have initiated the search for visual perceptual information indexed by the posterior hippocampus during elaboration. In sum, our data support the functional dissociation of anterior and posterior segments of the hippocampus but also emphasize its functional integration in vivid AM retrieval.

Limitations and Future Directions

A limitation of our study is the lack of information on the remoteness of the individual AMs, especially in the context of the wide age range of our participants. The debate of hippocampal involvement in remote memory has been addressed by various studies and the evidence so far indicates that the hippocampus remains involved in remote AMs, as long as they remain vivid and detail-rich (Addis, Moscovitch, et al. 2004; Moscovitch et al. 2005; Sheldon and Levine 2013). In the current study, most of the AMs were reported as vivid. We therefore expect that our results are more likely to have been influenced by the vivid memories rather than the faint ones. Of note, we are currently exploring these issues in a separate paper by comparing AM construction and elaboration networks of healthy controls to patients with medial temporal lobe damage who are known to be impaired in retrieval of vivid AMs (St-Laurent et al. 2009, 2011).

Further, we would like to acknowledge that our effective connectivity results represent a “working model,” recognizing the truth of the assertion that essentially all models are wrong but some are useful (Box and Draper 1987). Decisions about regions to include and how they are connected had to be made and thus including other neocortical regions or corticocortical connections might have yielded different results. However, we believe that examining patterns and directions of connectivity will help to both elaborate and constrain biologically plausible models and to generate new hypotheses about the nature of how brain regions interact during complex cognitive tasks.

Conclusions

As initially proposed by Conway, we found evidence that AM construction and elaboration are supported by different sets of brain regions. We add that the hippocampus plays a key role in coordinating these flexible, transient neocortical networks during AM retrieval. Further, our data revealed a novel and interesting intrahippocampal dialog between different stages of AM retrieval. We were able to illustrate this complex network interplay with the use of functional and effective connectivity analyses that assess the relationship between brain areas instead of local mean activation levels. These findings support the idea that in order to understand the full function of a brain region it is important to examine its neuronal context, that is, what other regions it interacts with under different experimental conditions (McIntosh 2000).

Supplementary Material

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

Funding

This work was supported by grants from the Canadian Institutes of Health Research (CIHR, grant number: 97891, to T.V. and M.P.M.) and the James S. McDonnell Foundation (JSMF #:22002055, to M.P.M.). C.M. holds a scholarship from the German Research Foundation (DFG #: MC-244/1-1).

Notes

The authors thank Keith Ta, Eugen Hlasny, Melanie Cohn, Irene Giannoylis, and Areeba Adnan for their help to collect the fMRI data. Conflict of Interest: None declared.

References

Addis
DR
Knapp
K
Roberts
RP
Schacter
DL
.
2012
.
Routes to the past: neural substrates of direct and generative autobiographical memory retrieval
.
Neuroimage
 .
59
:
2908
2922
.
Addis
DR
McIntosh
AR
Moscovitch
M
Crawley
AP
McAndrews
MP
.
2004
.
Characterizing spatial and temporal features of autobiographical memory retrieval networks: a partial least squares approach
.
Neuroimage
 .
23
:
1460
1471
.
Addis
DR
Moscovitch
M
Crawley
AP
McAndrews
MP
.
2004
.
Qualities of autobiographical memory modulate hippocampal activation during retrieval: preliminary findings of an fMRI study
.
Brain Cogn
 .
54
:
145
147
.
Addis
DR
Wong
AT
Schacter
DL
.
2007
.
Remembering the past and imagining the future: common and distinct neural substrates during event construction and elaboration
.
Neuropsychologia
 .
45
:
1363
1377
.
Andrews-Hanna
JR
Reidler
JS
Sepulcre
J
Poulin
R
Buckner
RL
.
2010
.
Functional-anatomic fractionation of the brain's default network
.
Neuron
 .
65
:
550
562
.
Bast
T
Wilson
IA
Witter
MP
Morris
RG
.
2009
.
From rapid place learning to behavioral performance: a key role for the intermediate hippocampus
.
PLoS Biol
 .
7
:
e1000089
.
Binder
JR
Desai
RH
Graves
WW
Conant
LL
.
2009
.
Where is the semantic system? A critical review and meta-analysis of 120 functional neuroimaging studies
.
Cereb Cortex
 .
19
:
2767
2796
.
Bonnici
HM
Chadwick
MJ
Maguire
EA
.
2013
.
Representations of recent and remote autobiographical memories in hippocampal subfields
.
Hippocampus
 .
.
Box
G
Draper
N
.
1987
.
Empirical model-building and response surfaces
.
New York
:
Wiley
. p.
424
.
Brewer
WF
.
1986
.
What is autobiographical memory?
In:
Rubin
DC
, editor.
Autobiographical memory
 .
New York
:
Cambridge University Press
.
Buckner
RL
Andrews-Hanna
JR
Schacter
DL
.
2008
.
The brain's default network: anatomy, function, and relevance to disease
.
Ann N Y Acad Sci
 .
1124
:
1
38
.
Buzsaki
G
.
1996
.
The hippocampo-neocortical dialogue
.
Cereb Cortex
 .
6
:
81
92
.
Conway
MA
.
2009
.
Episodic memories
.
Neuropsychologia
 .
47
:
2305
2313
.
Conway
MA
Pleydell-Pearce
CW
.
2000
.
The construction of autobiographical memories in the self-memory system
.
Psychol Rev
 .
107
:
261
288
.
Conway
MA
Pleydell-Pearce
CW
Whitecross
SE
.
2001
.
The neuroanatomy of autobiographical memory: a slow cortical potential study of autobiographical memory retrieval
.
J Mem Lang
 .
45
:
493
524
.
Daselaar
SM
Rice
HJ
Greenberg
DL
Cabeza
R
LaBar
KS
Rubin
DC
.
2008
.
The spatiotemporal dynamics of autobiographical memory: neural correlates of recall, emotional intensity, and reliving
.
Cereb Cortex
 .
18
:
217
229
.
Fanselow
MS
Dong
HW
.
2010
.
Are the dorsal and ventral hippocampus functionally distinct structures?
Neuron
 .
65
:
7
19
.
Greenberg
DL
Eacott
MJ
Brechin
D
Rubin
DC
.
2005
.
Visual memory loss and autobiographical amnesia: a case study
.
Neuropsychologia
 .
43
:
1493
1502
.
Greenberg
DL
Rubin
DC
.
2003
.
The neuropsychology of autobiographical memory
.
Cortex
 .
39
:
687
728
.
Holland
AC
Addis
DR
Kensinger
EA
.
2011
.
The neural correlates of specific versus general autobiographical memory construction and elaboration
.
Neuropsychologia
 .
49
:
3164
3177
.
Howe
ML
Courage
ML
.
1997
.
The emergence and early development of autobiographical memory
.
Psychol Rev
 .
104
:
499
523
.
Kline
RB
.
2005
.
Principles and practice of structural equation modeling
 .
New York
(
NY
):
The Guilford Press
.
Kondo
H
Saleem
KS
Price
JL
.
2005
.
Differential connections of the perirhinal and parahippocampal cortex with the orbital and medial prefrontal networks in macaque monkeys
.
J Comp Neurol
 .
493
:
479
509
.
Krishnan
A
Williams
LJ
McIntosh
AR
Abdi
H
.
2011
.
Partial Least Squares (PLS) methods for neuroimaging: a tutorial and review
.
Neuroimage
 .
56
:
455
475
.
Lavenex
P
Amaral
DG
.
2000
.
Hippocampal-neocortical interaction: a hierarchy of associativity
.
Hippocampus
 .
10
:
420
430
.
Lavenex
P
Suzuki
WA
Amaral
DG
.
2002
.
Perirhinal and parahippocampal cortices of the macaque monkey: projections to the neocortex
.
J Comp Neurol
 .
447
:
394
420
.
McCormick
C
Moscovitch
M
Protzner
AB
Huber
CG
McAndrews
MP
.
2010
.
Hippocampal-neocortical networks differ during encoding and retrieval of relational memory: functional and effective connectivity analyses
.
Neuropsychologia
 .
48
:
3272
3281
.
McIntosh
AR
.
2000
.
Towards a network theory of cognition
.
Neural Netw
 .
13
:
861
870
.
McIntosh
AR
Chau
WK
Protzner
AB
.
2004
.
Spatiotemporal analysis of event-related fMRI data using partial least squares
.
Neuroimage
 .
23
:
764
775
.
McIntosh
AR
Gonzalez-Lima
F
.
1994
.
Structural equation modeling and its application to network analysis in functional brain imaging
.
Hum Brain Mapp
 .
2
:
2
22
.
McIntosh
AR
Lobaugh
NJ
.
2004
.
Partial least squares analysis of neuroimaging data: applications and advances
.
Neuroimage
 .
23
(Suppl 1)
:
S250
S263
.
McIntosh
AR
Protzner
AB
.
2012
.
Structural equation models of imaging data
. In:
Hoyle
R
, editor.
Handbook of structural equation modeling
 .
New York
:
Guilford Press
.
Moscovitch
M
Rosenbaum
RS
Gilboa
A
Addis
DR
Westmacott
R
Grady
C
McAndrews
MP
Levine
B
Black
S
Winocur
G
et al
2005
.
Functional neuroanatomy of remote episodic, semantic and spatial memory: a unified account based on multiple trace theory
.
J Anat
 .
207
:
35
66
.
Patel
J
Fujisawa
S
Berenyi
A
Royer
S
Buzsaki
G
.
2012
.
Traveling theta waves along the entire septotemporal axis of the hippocampus
.
Neuron
 .
75
:
410
417
.
Poppenk
J
Evensmoen
HR
Moscovitch
M
Nadel
L
.
2013
.
Long-axis specialization of the human hippocampus
.
Trends Cogn Sci
 .
17
:
230
240
.
Poppenk
J
Moscovitch
M
.
2011
.
A hippocampal marker of recollection memory ability among healthy young adults: contributions of posterior and anterior segments
.
Neuron
 .
72
:
931
937
.
Rabin
JS
Rosenbaum
RS
.
2012
.
Familiarity modulates the functional relationship between theory of mind and autobiographical memory
.
Neuroimage
 .
62
:
520
529
.
Rajah
MN
McIntosh
AR
.
2005
.
Overlap in the functional neural systems involved in semantic and episodic memory retrieval
.
J Cogn Neurosci
 .
17
:
470
482
.
Ranganath
C
.
2010
.
A unified framework for the functional organization of the medial temporal lobes and the phenomenology of episodic memory
.
Hippocampus
 .
20
:
1263
1290
.
Ranganath
C
Ritchey
M
.
2012
.
Two cortical systems for memory-guided behaviour
.
Nat Rev Neurosci
 .
13
:
713
726
.
Rubin
DC
Greenberg
DL
.
1998
.
Visual memory-deficit amnesia: a distinct amnesic presentation and etiology
.
Proc Natl Acad Sci USA
 .
95
:
5413
5416
.
Rugg
MD
Johnson
JD
Park
H
Uncapher
MR
.
2008
.
Encoding-retrieval overlap in human episodic memory: a functional neuroimaging perspective
.
Prog Brain Res
 .
169
:
339
352
.
Sheldon
S
Levine
B
.
2013
.
Same as it ever was: vividness modulates the similarities and differences between the neural networks that support retrieving remote and recent autobiographical memories
.
Neuroimage
 .
83C
:
880
891
.
Sloviter
RS
Lomo
T
.
2012
.
Updating the lamellar hypothesis of hippocampal organization
.
Front Neural Circuits
 .
6
:
102
.
St Jacques
PL
Kragel
PA
Rubin
DC
.
2011
.
Dynamic neural networks supporting memory retrieval
.
Neuroimage
 .
57
:
608
616
.
St-Laurent
M
Moscovitch
M
Levine
B
McAndrews
MP
.
2009
.
Determinants of autobiographical memory in patients with unilateral temporal lobe epilepsy or excisions
.
Neuropsychologia
 .
47
:
2211
2221
.
St-Laurent
M
Moscovitch
M
Tau
M
McAndrews
MP
.
2011
.
The temporal unraveling of autobiographical memory narratives in patients with temporal lobe epilepsy or excisions
.
Hippocampus
 .
21
:
409
421
.
Suzuki
WA
Amaral
DG
.
1994
.
Perirhinal and parahippocampal cortices of the macaque monkey: cortical afferents
.
J Comp Neurol
 .
350
:
497
533
.
Svoboda
E
McKinnon
MC
Levine
B
.
2006
.
The functional neuroanatomy of autobiographical memory: a meta-analysis
.
Neuropsychologia
 .
44
:
2189
2208
.
Teyler
TJ
DiScenna
P
.
1986
.
The hippocampal memory indexing theory
.
Behav Neurosci
 .
100
:
147
154
.
Teyler
TJ
Rudy
JW
.
2007
.
The hippocampal indexing theory and episodic memory: updating the index
.
Hippocampus
 .
17
:
1158
1169
.
Tulving
E
.
2002
.
Episodic memory: from mind to brain
.
Annu Rev Psychol
 .
53
:
1
25
.
Viard
A
Desgranges
B
Eustache
F
Piolino
P
.
2012
.
Factors affecting medial temporal lobe engagement for past and future episodic events: an ALE meta-analysis of neuroimaging studies
.
Brain Cogn
 .
80
:
111
125
.