|Frequency-specific network connectivity increases underlie accurate spatiotemporal memory retrieval.|
|Jump to Full Text|
|PMID: 23354333 Owner: NLM Status: MEDLINE|
|The medial temporal lobes, prefrontal cortex and parts of parietal cortex form the neural underpinnings of episodic memory, which includes remembering both where and when an event occurred. However, the manner in which these three regions interact during retrieval of spatial and temporal context remains untested. We employed simultaneous electrocorticographical recordings across multilobular regions in patients undergoing seizure monitoring while they retrieved spatial and temporal context associated with an episode, and we used phase synchronization as a measure of network connectivity. Successful memory retrieval was characterized by greater global connectivity compared with incorrect retrieval, with the medial temporal lobe acting as a hub for these interactions. Spatial versus temporal context retrieval resulted in prominent differences in both the spectral and temporal patterns of network interactions. These results emphasize dynamic network interactions as being central to episodic memory retrieval, providing insight into how multiple contexts underlying a single event can be recreated in the same network.|
|Andrew J Watrous; Nitin Tandon; Chris R Conner; Thomas Pieters; Arne D Ekstrom|
Related Documents :
|18339943 - Synaptic theory of working memory.
25010653 - Short-term induction of assimilation and accommodation.
22439943 - The norwegian standardization of the matrics (measurement and treatment research to imp...
|Type: Journal Article; Research Support, N.I.H., Extramural; Research Support, Non-U.S. Gov't Date: 2013-01-27|
|Title: Nature neuroscience Volume: 16 ISSN: 1546-1726 ISO Abbreviation: Nat. Neurosci. Publication Date: 2013 Mar|
|Created Date: 2013-02-25 Completed Date: 2013-04-09 Revised Date: 2014-05-15|
Medline Journal Info:
|Nlm Unique ID: 9809671 Medline TA: Nat Neurosci Country: United States|
|Languages: eng Pagination: 349-56 Citation Subset: IM|
|APA/MLA Format Download EndNote Download BibTex|
Brain / physiology*, physiopathology
Epilepsy / physiopathology
Memory / physiology*
Mental Recall / physiology
Nerve Net / physiology*, physiopathology
Space Perception / physiology*
|R01 NS076856/NS/NINDS NIH HHS; R01NS076856/NS/NINDS NIH HHS|
Nat Neurosci. 2013 Mar;16(3):257-8
Neurosurgery. 2013 Dec;73(6):N16-7 [PMID: 24257340 ]
Journal ID (nlm-journal-id): 9809671
Journal ID (pubmed-jr-id): 21092
Journal ID (nlm-ta): Nat Neurosci
Journal ID (iso-abbrev): Nat. Neurosci.
nihms-submitted publication date: Day: 13 Month: 2 Year: 2013
Electronic publication date: Day: 27 Month: 1 Year: 2013
Print publication date: Month: 3 Year: 2013
pmc-release publication date: Day: 01 Month: 9 Year: 2013
Volume: 16 Issue: 3
First Page: 349 Last Page: 356
PubMed Id: 23354333
|Frequency–specific network connectivity increases underlie accurate spatiotemporal memory retrieval|
|Andrew J. Watrous12|
|Arne D. Ekstrom125|
1Neuroscience Graduate Group, University of California, Davis
2Center for Neuroscience, University of California, Davis, 1544 Newton Court, Davis, CA
3Department of Neurosurgery, University of Texas Medical School, Houston, Texas
4Memorial Hermann Hospital – Texas Medical Center, Houston, Texas
5Department of Psychology, University of California, Davis, CA
|Correspondence: CORRESPONDING AUTHOR: Dr. Arne Ekstrom, PhD, Center For Neuroscience, University of California, Davis, 1544 Newton Court, Davis, CA 95618, Phone: 530.757.8850, firstname.lastname@example.org
Successfully recalling an event in our everyday life depends critically on retrieving the context associated with it. A memory, such as what we ate for dinner last night, is more vividly recreated when we can remember contextual details, such as the location of the restaurant1. Neuroimaging and patient lesion studies strongly support the contributions of specific brain regions to episodic memory retrieval, emphasizing selective roles for the medial temporal lobe (MTL)1–7, prefrontal cortex2, 3, 6, 8, 9, and parts of parietal cortex2, 10–12 in this process. Distributed, coordinated activity across brain regions is also critical to memory retrieval13–18, with synchronized activity in the local field potential (LFP) implicated in coordinating this process19. Specifically, low–frequency (3–12 Hz) coordinated activity between hippocampus and prefrontal cortex in rodents is related to learning new rules during a spatial navigation task20 and can bias neocortical neuronal firing21, 22. The theoretical perspectives emerging from such work appear to provide conflicting accounts of whether successful memory retrieval is primarily mediated by a specific brain region (Figure 1A, upper), is best characterized by changes in functional connectivity between multiple brain regions (Figure 1A, lower), or is some combination of the two.
Recalling a prior experience often involves retrieving multiple, disparate2 types of context, e.g., not only where we ate dinner, but when it occurred relative to other events. Assuming that specific contextual information is represented by distinct cell assemblies within the same or overlapping brain regions13–16, how do different context representations emerge from within constituents of the same network? Although regional specificity may provide one possible account of this issue23, 24, another interesting proposal, the so–called spectral–fingerprint hypothesis, argues that different cognitive operations manifest as distinct, frequency–specific patterns of interregional phase synchronization in large–scale networks25, 26. These frequency–specific phase interactions are a strong candidate mechanism for coordinating distributed cell assemblies in parallel27–29, known as “frequency multiplexing,” and may underlie the rapid retrieval of contextual information specific to a particular experience (Figure 1B). While previous studies suggest the importance of low–frequency band power modulations in the MTL using resting MEG (between 4–6Hz)30 and coherence during human free recall tasks using intracranial EEG (at ~3 and 8Hz)31,32, how and in what manner frequency multiplexing occurs between brain regions involved in retrieval of episodic memory is not known.
We set out to test two fundamental sets of theoretical perspectives on the neural basis of episodic memory retrieval. The first regards whether accurate episodic memory retrieval is characterized by a disproportionate contribution of a specific brain region and/or accomplished by changes in global connectivity across the network (Figure 1A). The second addresses how different contexts can be retrieved via interactions within the same set of brain regions (Figure 1B). Specifically, to address the second issue, we test the above two possibilities regarding regional vs. global changes in connectivity along with the spectral fingerprint hypothesis. To directly test these ideas, we employed electrocorticographical (ECoG) recordings in patients undergoing clinical monitoring, focusing on three areas strongly implicated in past literature as central to episodic memory: parahippocampal gyrus ([PHG], that serves as a “gateway” to the hippocampus5, 33, see Methods), parietal cortex, and lateral prefrontal cortex (Supplementary Figure 1). We did this by comparing low–frequency (1–10 Hz) phase synchronization between these key regions, which we term “the retrieval network,” as patients performed a task requiring both spatial layout and temporal order memory retrieval (see Methods and Figure 1C–E). By measuring whether retrieval manifested as changes in connectivity between specific regional “hubs,” changes in global connectivity, or was best characterized by specific changes in spectrotemporal dynamics, our approach allowed us to directly test these different theoretical models regarding the neural basis of episodic memory.
Evaluation of simultaneously recorded ECoG signals revealed prominent low frequency phase consistency between PHG and specific sub–regions of parietal and prefrontal cortex (Figure 2A,2B; Supplementary Figure 2 & 3) when patients accurately retrieved spatiotemporal contextual information. On individual electrode pairs, correct context retrieval was accompanied by increases in PHG–parietal and PHG–prefrontal pairwise phase consistency beginning at the onset of retrieval (Figure 2C,D, and Supplementary Figure 4; see Online Methods for Pairwise Phase Consistency [PPC] description). Similar findings were also evident at the population level: we found significantly (pfwe < .05) increased low frequency PPC during correct retrieval across the network, for instance, between PHG–parietal (PHG–IPL [Figure 2E]) and PHG–prefrontal (PHG–MFG [Figure 2F]) recordings (Supplementary Figure 5; see Methods for statistical approach). Control analyses examining individual electrode raw traces (Supplementary Figures 2 & 3) and regional power modulations (Supplementary Figure 6) revealed that oscillatory power typically increased following cue onset. This regional increase in oscillatory power, however, was not condition specific (Supplementary Figure 6), suggesting that our PPC changes could not be accounted for by regional power changes alone (i.e., compare Supplementary Figures 5 vs. 6).
To evaluate coherent phase interactions globally across the brain, we adopted a graph theoretic approach34. We treated each subregion (e.g., MFG, IPL, PHG) as a node in a network functionally connected via phase synchronization at a given time and frequency (Figure 3A–E; see caption and Online Methods for description of network construction). We found the greatest PHG connectivity in the low frequency (1–10Hz) band compared to other bands (Figure 4A and Supplementary Figure 7; χ2(3) =70.48,p < .0001), consistent with previous findings that have suggested MTL networks operate preferentially within the delta–theta band31, 35, 36. We therefore restricted subsequent analyses to the low–frequency band. Figure 4B shows the networks associated with correct and incorrect memory retrieval at 3,5, and 8 Hz. We found significantly more interregional pairs (henceforth referred to as network “edges”) that showed significant task–related low frequency PPC increases during correct retrieval than during incorrect retrieval at every frequency up to 9 Hz (Figure 4C lower panel, blue asterisks; all χ2(1) >12.93, all p < .0008). This finding held true when we compared the distribution of connectivity between conditions at each frequency band across subregions, (each χ2(6) > 98, p < .0001) and for each subregion across 1–9 Hz frequency bands (χ2(8) >100, p < .0001). It was also robust to comparisons between nearly all subregions and frequencies individually (Figure 4B,C; compare Figure 4C top panel to Figure 4C middle panel, binomial tests, p < .05). Thus, rather than discrete switches in regional specificity within the network during correct vs. incorrect retrieval, or changes in preferred frequency, our findings suggest instead that network connectivity increases globally during correct vs. incorrect retrieval processes across these low frequencies (Figure 4B, compare top panel [correct] vs. bottom panel [incorrect] and Figure 4C, lower panel, compare red vs. blue bars).
Despite overall increases in connectivity across the “retrieval network” during correct memory retrieval, it could be the case that specific subregions still acted as “hubs,” that is, some subregions showed greater levels of connectivity compared to other subregions. To address this, we measured node degree, the total number of connections a node has with all other nodes. Following previous work34, we determined whether there were hubs in the network based on the distribution of node degree being different than a uniform network; a hub was then defined to be the node showing the highest degree of connectivity. We found that node degree varied significantly compared to a uniformly distributed network and that PHG had the greatest number of connections compared to other nodes across several low frequency bands (Figure 4B,C upper panels, red asterisks: 1Hz χ2(6)=15.4, p < .02; 4Hz, 7–9Hz χ2(6)>18, p < .005 for all comparisons). This analysis demonstrates that PHG acts as a “hub” for interactions during accurate episodic memory retrieval. Notably, PHG node connectivity did not vary across frequencies (χ2(9)=6.9, p=.64), indicating that it acts as a hub for information transfer across all low frequency bands. Because the PHG serves as a gateway to the hippocampus33,5, these findings are consistent with many decades of work that point to the importance of the medial temporal lobes in coordinating episodic memory retrieval1–7.
Our behavioral paradigm involved retrieving both the location of an item within the spatial layout and its temporal order relative to other items (see Methods). We were therefore able to evaluate differences across identical patients, recording zones, and visually identical experimental paradigms (differing only on remembering spatial vs. temporal information) during retrieval of spatial or temporal contexts (Figure 1C–E). We then evaluated whether this led to changes in interregional connectivity, global connectivity, or frequency–specific differences in connectivity (Figure 1A–B). Our findings primarily support the third possibility. Phase coherent oscillations in the raw trace were visually evident in a lower frequency band during spatial retrieval trials (Figure 5A, upper) compared to temporal retrieval (Figure 5A, lower); this difference was also evident in comparisons at the patient (Figure 5B) and at the population level (Figure 5C; statistical methods identical to those in Figure 2E & 2F).
Using our graph–theoretic approach, we also found that these behavior–specific and frequency–dependent phase interactions were present across the network. Comparing node connectivity overall within the network, we observed significantly greater connectivity between nodes during spatial retrieval compared to temporal retrieval from 1–4 Hz (Figure 5D, all χ2(1) >14.61, all p < .0002; also see Figure 5G–J). In contrast, we observed significantly greater connectivity between nodes during temporal compared to spatial retrieval from 7–10 Hz (7–8Hz χ2(1) >16.5, p < .0001; 9–10Hz χ2(1) >4.1, p < .05). For spatial retrieval, 6 of 7 of nodes showed greater functional connectivity in the 1–4 Hz band compared to the 7–10 Hz band (Figure 5E; Chi–square test between bands; allχ2(1) >9.13, all p < .003). In contrast, during temporal retrieval, 6 of 7 nodes showed differential connectivity in the 7–10 Hz band (Figure 5F; all χ2(1) >11.3, all p < .001). We confirmed a lower preferred frequency during spatial retrieval both across subjects (Figure 5K, t(5) = 2.44, one–tailed p = .029) and across electrode pairs (t(664) = 4.43, p < .000001). These findings demonstrate that the “retrieval network” resonates at a lower frequency overall during spatial compared to temporal retrieval.
Despite these differences in the preferred frequency in which connectivity manifested, the PHG node again showed the highest degree of connectivity compared to any other node during both spatial and temporal retrieval (Figure 5E,F). During spatial retrieval, PHG showed the greatest connectivity in the 1–4 Hz band (Figure 5G) while during temporal retrieval, PHG showed the greatest connectivity in the 7–10 Hz band (Figure 5J). During spatial retrieval, we found that the PHG node connections were primarily clustered with SFG, MFG, and PCN (Figure 5G χ2(5)= 32.63, p < .00001) and there was significantly more PHG connectivity with these nodes in the 1–4 Hz band compared to the 7–10Hz band (χ2(5)= 208.65, p < .00001). In contrast, during temporal order retrieval, PHG connections were preferentially clustered with SFG, MFG, and IPL (Figure 5J; χ2(5)= 19.09, p < .002) and there was significantly more PHG connectivity with these nodes in the 7–10 Hz band compared to the 1–4Hz band (χ2(5)= 78.58, p < .00001). These differences amounted to a significant condition (spatial vs. temporal) by frequency (1–4 Hz vs. 7–10 Hz) interaction in PHG connectivity (Fisher’s Exact test, p < .00001). Follow–up analyses separately assessing the spatial correct vs. spatial incorrect and temporal correct vs. temporal incorrect networks showed similar results to those directly contrasting spatial correct vs. temporal correct (Supplementary Figure 8). These findings show that retrieving spatial layout vs. temporal order information from an episode was characterized primarily by frequency–specific interactions across the retrieval network. Just as in our earlier analyses, however, PHG had the highest degree of connectivity within the network compared to any other node, suggesting its central importance to both spatial and temporal retrieval.
We also evaluated whether spatial and temporal retrieval were characterized by changes in timing of connectivity within their preferred frequencies. For all pairwise combinations of regions within the network, we calculated the connectivity map for the two different conditions based on a matrix of spatial vs. temporal PPC difference scores at 2Hz (Figure 6A) and 8Hz (Figure 6B). During spatial retrieval, significantly more connectivity occurred early (0–1 second) while during temporal retrieval, significantly more connectivity occurred late (1–2 seconds) (Figure 6C–D; Fisher’s exact test, p<.0001). Visual inspection of the matrix of network connectivity suggested that spatial and temporal retrieval were also characterized by different clustering of connectivity across regional pairs (Figure 6A,B; Supplementary Movies S1 and S2). We confirmed this impression by computing the Phi correlation coefficient of binary connections over time (see Methods), with higher correlation coefficients indicating a more coherent pattern of activity across the entire network and lower correlation coefficients indicating a less coherent pattern of activity across the network. The distribution of correlation coefficients differed between spatial and temporal retrieval (Figure 6E, Kolmogorov–Smirnov test, p<.0001) and the connectivity in the 2Hz spatial network was significantly more clustered in time than the 8Hz temporal network (Figure 6F, t–test (189)= 5.77, p < .0001). Critically, though, 1) these two networks did not differ in their total number of connections (see Figure 5D,χ 2(1)= .81, p=.37), 2) we found similar results at adjacent frequencies (Supplementary Figure 9A), 3) and the findings could not be accounted for by the expected increase in temporal autocorrelation at lower frequencies (Supplementary Figure 9B). These results confirm that the connectivity during spatial and temporal retrieval, in addition to being characterized by different resonant frequencies, also displayed differences in the pattern of connectivity over time.
Finally, our results could not be accounted for by poor patient performance nor a difference in performance between spatial and temporal retrieval. Patients performed significantly above chance (t(5) = 3.58, p = .015) and showed similar reaction times in both the spatial and temporal tasks. Neither accuracy (t(5) = −.56, p = .59) nor reaction time (t(5) = −1.00, p =.35) differed between the spatial (71% accuracy; 7.57s median RT) and temporal (76% accuracy; 8.99s median RT) retrieval conditions. Our results were also robust at more stringent statistical thresholds (p<.001, Supplementary Figure 10).
In this work we sought to determine the functional network interactions amongst brain regions previously implicated in successful episodic memory retrieval and to determine how multiple contextual representations characterizing an episode were retrieved within this network. We tested three fundamental theoretical models to determine whether one, or a combination of these, best characterized the neural basis of episodic memory retrieval. These included the following: 1) a model in which a single brain region acted as a hub for mediating episodic memory retrieval, 2) a model in which interactions were distributed relatively equally across nodes yet changed for successful retrieval or retrieval of different contexts, 3) a model in which changes in spectrotemporal dynamics mediated memory retrieval (see Figure 1). Our first set of analyses compared correct vs. incorrect context retrieval, collapsing across spatial and temporal retrieval trials. We found that successful memory retrieval was characterized by increases in network functional connectivity via phase synchronization across the 1–10 Hz low frequency band. PHG electrodes showed the highest degree of inter–connectivity with other electrodes compared to any other subregion within the “retrieval network” and this functional connectivity preferentially occurred in the 1–10Hz band. The macro–electrodes located over the PHG likely captured signals from posterior parahippocampal, perirhinal, and entorhinal cortices (see Methods), all of which provide direct input to the hippocampus5, 33, 37, suggesting the likelihood that the hippocampus also participated in these interactions. Overall, while our findings of increased network connectivity during correct vs. incorrect context retrieval support the idea of global rather than regionally–specific changes in connectivity (Figure 1A, lower), the differentially enhanced PHG connectivity we observed supports the idea that the MTL acted as a hub in these interactions (Figure 1A, upper). Thus, our results support a hybrid of the two models put forth in Figure 1A, suggesting that successful memory retrieval is best characterized by an overall increase in interactions across key brain regions mediated primarily by the MTL. To our knowledge, this has not been demonstrated previously with invasive human recordings that can pinpoint and disambiguate activity simultaneously across multiple brain regions.
Our results also emphasize the importance of phase–synchronized oscillations between these disparate brain regions as important to memory retrieval. Together with studies examining episodic encoding38 and working memory39, 40, our results, which were obtained during retrieval, suggest that phase coding may be a mechanism involved in an array of human memory processes. Previous proposals19,25,41 suggest that synchronized oscillations may facilitate precisely–timed depolarizations between neurons in communicating brain regions. This idea is supported by the observation in humans 27,28, rats20,and monkeys29 that oscillations coordinate both local and distant neuronal activity. Furthermore, an influential theoretical model of memory retrieval, Multiple Trace Theory (MTT)17, argues that cortical areas represent features bound to experience–specific contextual representation in the hippocampus, whose successful interaction underlies memory retrieval. Thus, our results provide evidence for a possible neural mechanism underlying MTT, namely, coordinated low frequency oscillations between the medial temporal lobe and cortical areas.
Taken together, our results yield a potentially new perspective on previous findings regarding the importance of MTL, parietal cortex, and prefrontal cortex in mediating successful episodic memory retrieval. Several fMRI studies show increased parietal cortex activation during episodic retrieval tasks 1–3, 10, 11 yet lesions to this area in human patients do not consistently impair episodic memory10. Similarly, prefrontal cortex activation is often a hallmark of successful memory retrieval yet prefrontal cortex lesions produce nuanced impairments in episodic memory retrieval8. For example, prefrontal cortex lesions impact free but not cued recall and are thought to affect executive processes but not retrieval of memory traces specifically9. In contrast, fMRI studies consistently show hippocampal and parahippocampal activation during correct recollection of specific contextual details2, 3 and hippocampal lesions in particular profoundly impair recollection and episodic memory7, 42. Our results thus suggest that while lateral prefrontal and parietal areas communicate with the MTL during spatiotemporal retrieval, the MTL acts as a critical convergence “hub” during successful context retrieval, broadly consistent with MTT17. One possible interpretation of our results is that the MTL serves as the primary locus for indexing particular memory traces, with parietal and lateral prefrontal cortex interacting with the MTL to facilitate and augment memory trace retrieval17, 18.
Epilepsy is a disease marked by impaired episodic memory performance and increased synchronized activity of large numbers of neurons which may manifest as low frequency phase synchronization. Do these factors confound our results? We believe several considerations weigh against this possibility. First, electrodes showing ictal and interictal discharge were systematically removed from our analysis based on evaluation by our clinical team. All analyzed trials were visually inspected for artifacts related to epilepsy. Next, disease related low frequency phase synchronization is likely to impair brain function and may interfere with accurate memory retrieval. Therefore, we would expect more low frequency phase coherence during incorrect compared to correct retrieval if ictal discharge alone accounted for our results, which is inconsistent with our findings. Third, it seems unlikely that epileptic activity would manifest itself as task–related (spatial vs. temporal) differences. Finally, we utilized a metric of phase synchronization which excluded phase differences centered around 0 degrees (see Methods) to eliminate possible effects of volume conduction and any remaining epileptic synchronization. We note that this also excluded “true” zero–phase lag synchronization (increasing Type 2 error) but provides a more conservative estimate of interregional coupling. Although we observed increases in phase synchronization, this does not imply nor require direct anatomical connectivity between areas, and is instead taken primarily as a measure of functional connectivity. In summary, we have made every available effort to account for possible confounds while still capitalizing on the increased spatiotemporal resolution afforded by direct human brain recordings. Thus, we believe that our results, which provide novel insight into the dynamic networks underlying spatiotemporal episodic memory, would generalize to the population if these issues could be studied non–invasively.
Our behavioral paradigm also allowed us to compare retrieval of two different fundamental components of episodic memory, spatial and temporal context. Here, our results suggested that correct retrieval of the location of an item involved lower frequency interactions compared to correctly retrieving the order in which an item occurred relative to other items. While we again found that correct retrieval was characterized by higher degrees of connectivity compared to incorrect retrieval for the two processes separately (Supplementary Figure 8), the overall level of connectivity did not differ between spatial vs. temporal retrieval (Figure 5D). We also found differences in regional connectivity during spatial (PHG–SFG–MFG–Precuneus) vs. temporal (PHG–SFG–MFG–IPL) retrieval, with the critical difference being the subregions of parietal cortex engaged. Thus, our findings here primarily support the spectral fingerprint hypothesis25 (Figure 1B), with some regional specificity within parietal cortex (Figure 1A, upper) but no clear differences in overall connectivity (Figure 1A, lower). Because in both cases PHG acted as a hub for interactions, our results suggest that spatial vs. temporal retrieval involved similar degrees of connectivity with the MTL. In a previous fMRI study using a similar behavioral paradigm, we found similar degrees of hippocampal activation during spatial and temporal retrieval along with greater activation in precuneus during a contrast of spatial vs. temporal retrieval24 and, in another study, greater functional connectivity between hippocampus and precuneus during spatial retrieval 43. Other studies have also implicated parts of inferior posterior parietal cortex in temporal processing44. Notably, however, many of the same regions in lateral prefrontal, MTL, and parietal cortex remained significantly connected in both networks. Thus, the most salient differences between the networks for retrieval of spatial vs. temporal context, in contrast to our findings with correct vs. incorrect retrieval, was the spectrotemporal dynamics at which the two networks operated, again consistent primarily with the spectral fingerprint hypothesis (Figure 1B).
Our results thus provide a new basis for resolving the question of how multiple contexts underlying an episode can be stored and retrieved within the same network of brain regions. Behaviorally, a spatial layout can be treated as a map45 and therefore a coherent entity overall and, once well learned, can in principle be loaded and accessed quickly. We found that correct spatial retrieval was characterized by lower frequency interactions overall across the network along with early and prolonged increases in functional connectivity across the network compared to temporal order retrieval. In contrast, remembering temporal order information involves traveling back in time to retrieve different temporal contexts43,46,47, which would necessitate active comparison of each element within the sequence with the item to be compared. We found that temporal order retrieval was characterized by faster frequency interactions, a more delayed increase in network connectivity, and overall less coherence in time across the network compared to spatial retrieval. While somewhat speculative, these explanations provide possible insight into differences in how the brain is able to process aspects of space and time during memory retrieval. Specifically, our data lend support to the concept of spectrotemporal multiplexing25, 27, 29 as a means to store and retrieve spatial and temporal context information encoded amongst neurons in the same regions. Our results thus provide a possible mechanism by which spatial and temporal contextual information, thought to underlie episodic memory1,2, could be retrieved simultaneously. In our study, however, patients retrieved spatial and temporal context on separate trials, and it is not necessarily the case that the dynamics will be identical, for example, when spatial and temporal context are retrieved simultaneously, an issue future studies will need to address. Notably, a previous study during memory encoding38 identified frequency–specific differences in spike–field coherence in the low–frequency band that predicted later retrieval, raising the possibility that spectrotemporal multiplexing is a general feature in episodic memory processes.
By employing direct intracranial recordings and a graph theoretic approach, our study provides a new perspective on how the human brain processes episodic memories. Our data provide novel support for models that emphasize global network interactions and frequency–specific connectivity, rather than regionally–mediated activity alone, as central to how we recover spatial and temporal memories associated with recent experiences. Our results thus argue for the importance of carefully timed dynamics across multiple brain regions as critical to spatiotemporal memory retrieval.
Six adult patients with medically refractory epilepsy underwent electrocorticography (ECoG) to localize seizures and participated in the study after providing informed consent approved by the University of Texas Medical Center committee for the protection of human subjects. Electrophysiological methods and electrode localization were similar to that described previously 48. In brief, subdural circular platinum–iridium electrodes with a top hat design (4.5 mm overall diameter, 3 mm cortical contact, 10 mm interelectrode distance) were implanted and placed based solely on clinical considerations using standard techniques49. Electrode localization was verified by co–registering a post–operative CT image with a pre–operative MRI structural image. Lobar and gyral labels were assigned by an expert in human neuroanatomy (N.T.). Electrodes showing epileptiform activity based on neurologists evaluation were excluded from all analyses.
ECoG signals were sampled at 1000 Hz using Nihon Kohden NeuroFax software with a recording bandwidth from .15 to 300 Hz. Signals were referenced to a common average consisting of all non–ictal electrodes over lateral frontal and lateral temporal areas to minimize the impact of the referencing scheme on synchronization measures50. Recordings were then imported into Matlab (MathWorks; Natick, MA) for post–processing.
Patients played a virtual–taxi game similar to that used previously24 with a few exceptions that were necessary to accommodate the clinical environment. The virtual environment consisted of 5 stores spaced irregularly and arranged in a circle (Figure 1D). Patients performed multiple blocks of navigation in which they picked up a passenger in the center of the environment and delivered them to a specific store (Figure 1C). This involved freely navigating the environment with a fixed order of visits to stores. Both the spatial layout of the environment and the delivery order were independent of each other24 and maintained across blocks in order to subsequently test spatial layout and temporal order memory.
Following each block of navigation, patients performed interactive training emphasizing the spatial layout as a 2–D array and the temporal order of visits to stores as a 1–D sequence in time (Figure 1D). Interim testing was performed to emphasize the different characteristics of the spatial layout and temporal order of deliveries and to ensure that the patient had sufficient knowledge to accurately perform the final retrieval session. Interim testing involved the patient localizing each store onto a grid of the virtual environment viewed from an overview perspective or onto a timeline corresponding to the delivery order to stores (Figure 1D). The patient then viewed the correct answers displayed for 10 seconds separately for the spatial layout and the temporal sequence. The order of interactive training (spatial or temporal) was randomized following each block of navigation. Patients continued to perform multiple rounds of navigation and interactive training until they achieved either 100% accuracy on both spatial and temporal questions for two consecutive rounds or had completed eight rounds of navigation and interactive training.
During the final retrieval session (Figure 1E), which forms the basis of the results reported here, patients performed a block of spatial retrieval and a block of temporal retrieval trials, the order of which was counterbalanced across patients. The patient was provided with an image of the cue storefront along with two additional storefronts to choose from. The environment was arranged such that no two stores were ever the same distance apart. This generated 30 different possible unique trials for the spatial condition. During spatial retrieval trials, the patient was instructed to indicate which of the two stores was closer to the cue store in virtual space. For the temporal condition, we designed questions with the constraint that the choice stores either preceded or followed the reference store in time. This generated a total of 40 possible unique trials for the temporal condition. During temporal order retrieval trials, the patient indicated which of the two stores was closer to the cue store in delivery order. The number of trials analyzed for each patient in each condition is shown in Supplementary Table 2. One patient was excluded from the correct vs. incorrect analysis because they lacked sufficient incorrect responses. Responses for all portions of the paradigm described were subject paced to accommodate patient needs and the clinical testing environment; we therefore analyzed the fixed two second interval associated with cue onset during which subjects were instructed to begin retrieving from memory. We did this because we hypothesized that changes in activity related to retrieval would reliably occur immediately after cue onset, consistent with previous studies10. The onset of the each trial was jittered uniformly from 1 to 1.5 seconds following subject response.
All analyses used EEGLab51 and custom–written code in Matlab. Raw EEG signals from the spatial and temporal retrieval session were extracted for both correct and incorrect responses from 1 second before to 2.2 seconds following cue onset in order to remove edge effects associated with spectrotemporal decomposition. Our primary behavioral contrasts were correct (spatial correct and temporal correct) vs. incorrect (spatial incorrect and temporal incorrect) and spatial correct vs. temporal correct. Phase synchronization estimates were computed between each pairwise combination of PHG (consisting of Parahippocampal, Perirhinal, and Entorhinal cortices), prefrontal (Superior & Middle Frontal Gyrus, Pars Triangularis), and parietal electrodes (Superior & Inferior Parietal Lobule, Precuneus) for each condition and contrast. Abbreviations for these regions are PHG, SFG, MFG, IFG, SPL, IPL, and PCN, respectively, and total pair counts are shown in Table S1. Phase estimates were obtained using a Hanning tapered fixed window length fast–Fourier transform (FFT) at 27 timepoints using the EEGLab “newcrossf” function from 2 to 1943 ms relative to cue onset and at 10 logarithmically spaced frequencies from .97 to 9.76 Hz. Frequencies were subsequently rounded to the nearest whole number. For analyses assessing phase synchronization to 200 Hz (Figure 4A and Supplementary Figure 7), we sampled a total of 40 logarithmically spaced frequencies such that each of the four frequency bins (1–10 Hz; 11–40Hz; 40–100 Hz; 100–200Hz) contained 10 sampled frequencies.
We estimated phase synchronization between electrode pairs at the above times and frequencies using the pairwise phase consistency (PPC) index in order to address issues associated with differing trial numbers between conditions52. Briefly, the PPC index for an electrode pair was estimated by first computing the relative phase angle difference between signals on each trial. The cosine was then computed between all pairwise (i.e., between trials) combinations of relative phases, and the PPC was taken as the mean of these cosine values. PPC values range from −1 to 1, with positive values indicating phase synchronization. We accounted for potential volume conduction confounds, which may artificially inflate phase synchronization estimates, by removing relative phases at 0 degrees (±5 degrees) prior to calculating the PPC index. We note that using standard phase coherence estimates or PPC estimates that included zero–phase synchronization did not qualitatively change our primary findings.
We utilized graph theoretic measures34 to determine how functional interactions varied over time, frequency, and brain subregion. We considered a functional connection (i.e., edge) to exist between two subregions (“nodes”) if there was a significant (bootstrap corrected p<.05, one–tailed paired t–test in order to test directionality) difference between conditions across all electrode pairs. Specifically, we compared PPC for one condition against a different condition across interregional electrode pairs, pooled across patients, and thus the degrees of freedom were the total number of electrode pairs in a comparison. To account for issues related to multiple comparisons and to maintain a fixed Type 1 error rate of 5 percent, we estimated the distribution of t–values and shuffled the condition labels 1000 times before recomputing test statistics; observed t–values greater than the 95th percentile of this distribution were considered significant and were noted as significant connections (“edges”) in the network. This procedure was repeated separately at each time bin and at each frequency point (Figure 3). We then computed three basic measures of network connectivity. To assess the connectivity of a single node, we computed the total number of connections between a given node and all other connected nodes over time (0–2000 msec). We defined this to be node degree. To determine the relative strength of functional interactions between two nodes, (i.e. Figure 3C), we computed the total number of connections between those two nodes over time. Finally, we summed all connections in the network at a specific frequency over time, providing a frequency–specific measure of global functional interactions in the network. These measures were used for statistical testing (below), and were then converted to a percent connectivity measure. Percent connectivity was defined as the number of observed edges out of the total number of possible edges (accounting for the number of time, frequency, and interregional pairs sampled). Follow–up analyses using a more stringent bootstrap alpha of .001 (Supplementary Figure 10) were consistent with our primary findings.
In a random network, edges are uniformly distributed amongst nodes such that connectivity should not vary across regional pairs or frequencies34. Statistical analyses thus employed Chi Square tests to compare the distribution of edges and total number of connections for each node in the actual network against the null hypothesis that edges and number of connections were uniformly distributed.
To assess if edges were more clustered in time in the 2Hz Spatial or 8Hz Temporal networks, for each interregional pair, we extracted a binarized edge map, in which a 1 indicates a connected pair at a given time point and 0 indicates a lack of a connection. Binary vectors for each combination of interregional pairs (twenty interregional pairs chosen two at a time, n=190 pairs total) were then correlated using the Phi correlation coefficient, resulting in a distribution for each network. Differences between these distributions were assessed using a two–sample Kolmogorov–Smirnov test and additionally using paired t–tests. Follow–up analysis estimated the “chance” correlation by generating 1000 randomly seeded networks (matched for total number of edges in the 2Hz and 8Hz network using Monte Carlo simulations), re–calculating the mean Phi correlation across pairs, and extracting the 95th percentile of this distribution.
Statistical significance in our basic population analyses of task–related differences in PPC (i.e. Figures 2E,2F,5C; number of pairs for each comparison shown in Table S1) used family–wise error correction, with pfwe=.0553. This corresponded to 4 contiguous time–frequency points each individually significant with a one–tailed paired t–test at p < .005. We note that the most significant time–frequency points held up to thresholding without cluster correction to at least p<.00001.
Figure S1: Location of each electrode analyzed visualized on the lateral (upper panel), medial (middle panel), and ventral (lower panel) surface of a template brain. Each sphere represents an electrode and different subregions are shown as different colors, corresponding to PHG (Yellow), SFG (Light Blue), MFG (Purple), IFG (Dark Blue), SPL (Sea Green), PCN (Orange), and IPL (Olive Green). Arrows indicate electrodes used in Figures 2A and 2B (green), and 5A (red).
Figure S2. Example trials depicting raw traces for correct (A) and incorrect (B) trials from the pairs shown in Figure 2A.
Figure S3. Example trials depicting raw traces for correct (A) and incorrect (B) trials from the pairs shown in Figure 2B.
Figure S4. Absolute PPC estimates and individual trial phase angle differences for the electrode pairs shown in Figure 2A (S4A) and Figure 2B (S4B). A) Upper: Rose plots showing the distribution of 4Hz phase angle differences for correct and incorrect trials between the electrode pair (PHG & IPL) shown in Figure 2A. Rose plots are shown at time points (black asterisks; lower panel) corresponding to maximal differences between correct and incorrect PPC. Lower: Absolute PPC estimates for the pair across correct and incorrect trials. Whereas PPC increases following stimulus onset for correct trials, the pair shows a concomitant decrease in PPC for incorrect trials. B) Similar to A, except for the electrode pair (PHG & MFG) shown in Figure 2B.
Figure S5. 4Hz PPC estimates over time for correct (upper) and incorrect (middle) trials for all PHG & IPL (A) and PHG & MFG (B) electrode pairs (y-axis in upper and middle panel). The mean and SEM across electrode pairs is shown in the lower panel. Asterisks indicate significant differences between conditions using a bootstrap resampling procedure across electrode pairs (bootstrap alpha = .01, 1000 shuffles at each timepoint). To enhance clarity, the upper and middle panels have been individually scaled between zero and their maximum value, although many PPC estimates for incorrect trials were below zero and thus appear as dark blue.
Figure S6. Power estimates over time for correct (upper) and incorrect (middle) trials for all electrodes in MFG (A), IPL (B), and PHG (C). The mean and SEM across electrodes pairs is shown in the lower panel. Significant power effects were not observed in any region at the population level after correcting for multiple comparisons (paired t-test across electrodes at each time point, all p-values greater than Bonferroni corrected p <.05).
Figure S7. Node connectivity as a function of frequency (1–200HZ) and region for analyses comparing correct vs. incorrect. The percent total connectivity in each network is shown as a function of frequency below (asterisks indicate Bonferroni corrected p<.05).
Figure S8. Node connectivity as a function of frequency and region for analyses comparing spatial correct vs. spatial incorrect (A) and temporal correct vs. temporal correct (B). The percent total connectivity in each network is shown as a function of frequency below (asterisks indicate p<.05).
Figure S9. Network correlations over time. A) Mean network correlation over time derived for the spatial and temporal networks at each frequency. B) Mean network correlation over time collapsed across conditions. Note that the correlation does not decrease smoothly with frequency, as would be expected if temporal autocorrelation alone accounted for our results in Figure 6F. Error bars denote standard error of the mean across inter-regional pair combinations (n=190).
Figure S10. Node connectivity as a function of frequency and region for analyses comparing correct vs. incorrect (A) and spatial correct vs. temporal correct (B). Edges were defined using bootstrap alpha of .001. The percent total connectivity in each network is shown as a function of frequency below (asterisks indicate p<.05).
Table S1: Number of electrode pairs and subjects used for the Correct vs. Incorrect analysis and Spatial vs. Temporal analysis. Differences in included patients, and thus pair counts, occurred because Patient #2 lacked sufficient incorrect responses (2 total incorrect responses, see Table S2).
Table S2: Number of trials analyzed for each subject and each condition.
2 Click here for additional data file (NIHMS431152-supplement-2.mpg)
3 Click here for additional data file (NIHMS431152-supplement-3.mpg)
A.D.E., N.T., and A.J.W. designed the experiment. N.T., C.C., and T.P. collected the data. A.J.W. performed the data analysis. A.J.W., A.D.E., and N.T. wrote the manuscript.
This work was supported by the Sloan Foundation, Hellman Young Investigator Award, and NINDS RO1NS076856. We thank Charan Ranganath and Andrew Yonelinas, as well as their lab members, along with 3 anonymous reviewers for helpful comments on this manuscript.
|1.||Eichenbaum H,Yonelinas AP,Ranganath C. The medial temporal lobe and recognition memoryAnnu Rev Neurosci30123152Year: 200717417939|
|2.||Mitchell KJ,Johnson MK. Source monitoring 15 years later: what have we learned from fMRI about the neural mechanisms of source memory?Psychol Bull135638677Year: 200919586165|
|3.||Ranganath C,et al. Dissociable correlates of recollection and familiarity within the medial temporal lobesNeuropsychologia42213Year: 200414615072|
|4.||Scoville WB,Milner B. Loss of recent memory after bilateral hippocampal lesionsJ Neurol Neurosurg Psychiatry201121Year: 195713406589|
|5.||Squire LR,Stark CE,Clark RE. The medial temporal lobeAnnu Rev Neurosci27279306Year: 200415217334|
|6.||Simons JS,Spiers HJ. Prefrontal and medial temporal lobe interactions in long-term memoryNat Rev Neurosci4637648Year: 200312894239|
|7.||Vargha-Khadem F,et al. Differential effects of early hippocampal pathology on episodic and semantic memoryScience277376380Year: 19979219696|
|8.||Duarte A,Ranganath C,Knight RT. Effects of unilateral prefrontal lesions on familiarity, recollection, and source memoryJ Neurosci2583338337Year: 200516148241|
|9.||Blumenfeld RS,Ranganath C. Prefrontal cortex and long-term memory encoding: an integrative review of findings from neuropsychology and neuroimagingNeuroscientist13280291Year: 200717519370|
|10.||Vilberg KL,Rugg MD. Memory retrieval and the parietal cortex: a review of evidence from a dual-process perspectiveNeuropsychologia4617871799Year: 200818343462|
|11.||Spaniol J,et al. Event-related fMRI studies of episodic encoding and retrieval: meta-analyses using activation likelihood estimationNeuropsychologia4717651779Year: 200919428409|
|12.||Hutchinson JB,Uncapher MR,Wagner AD. Posterior parietal cortex and episodic retrieval: convergent and divergent effects of attention and memoryLearn Mem16343356Year: 200919470649|
|13.||Buzsaki G. The hippocampo-neocortical dialogueCereb Cortex68192Year: 19968670641|
|14.||Eichenbaum H. A cortical-hippocampal system for declarative memoryNat Rev Neurosci14150Year: 200011252767|
|15.||Norman KA,O’Reilly RC. Modeling hippocampal and neocortical contributions to recognition memory: a complementary-learning-systems approachPsychol Rev110611646Year: 200314599236|
|16.||McClelland JL,McNaughton BL,O’Reilly RC. Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memoryPsychol Rev102419457Year: 19957624455|
|17.||Nadel L,Moscovitch M. Memory consolidation, retrograde amnesia and the hippocampal complexCurr Opin Neurobiol7217227Year: 19979142752|
|18.||Teyler TJ,DiScenna P. The hippocampal memory indexing theoryBehav Neurosci100147154Year: 19863008780|
|19.||Fell J,Axmacher N. The role of phase synchronization in memory processesNat Rev Neurosci1210511821248789|
|20.||Benchenane K,et al. Coherent theta oscillations and reorganization of spike timing in the hippocampal-prefrontal network upon learningNeuron6692193620620877|
|21.||Siapas AG,Lubenov EV,Wilson MA. Prefrontal phase locking to hippocampal theta oscillationsNeuron46141151Year: 200515820700|
|22.||Sirota A,et al. Entrainment of neocortical neurons and gamma oscillations by the hippocampal theta rhythmNeuron60683697Year: 200819038224|
|23.||Spiers HJ,et al. Unilateral temporal lobectomy patients show lateralized topographical and episodic memory deficits in a virtual townBrain12424762489Year: 200111701601|
|24.||Ekstrom AD,Copara MS,Isham EA,Wang WC,Yonelinas AP. Dissociable networks involved in spatial and temporal order source retrievalNeuroimage561803181321334445|
|25.||Siegel M,Donner TH,Engel AK. Spectral fingerprints of large-scale neuronal interactionsNat Rev Neurosci1312113422233726|
|26.||Donner TH,Siegel M. A framework for local cortical oscillation patternsTrends Cogn Sci1519119921481630|
|27.||van der Meij R,Kahana M,Maris E. Phase-amplitude coupling in human electrocorticography is spatially distributed and phase diverseJ Neurosci3211112322219274|
|28.||Maris E,van Vugt M,Kahana M. Spatially distributed patterns of oscillatory coupling between high-frequency amplitudes and low-frequency phases in human iEEGNeuroimage5483685020851192|
|29.||Canolty RT,et al. Oscillatory phase coupling coordinates anatomically dispersed functional cell assembliesProc Natl Acad Sci U S A107173561736120855620|
|30.||Hipp JF,Hawellek DJ,Corbetta M,Siegel M,Engel AK. Large-scale cortical correlation structure of spontaneous oscillatory activityNat Neurosci|
|31.||Lega BC,Jacobs J,Kahana M. Human hippocampal theta oscillations and the formation of episodic memoriesHippocampus2274876121538660|
|32.||Anderson KL,Rajagovindan R,Ghacibeh GA,Meador KJ,Ding M. Theta oscillations mediate interaction between prefrontal cortex and medial temporal lobe in human memoryCereb Cortex201604161219861635|
|33.||Lavenex P,Amaral DG. Hippocampal-neocortical interaction: a hierarchy of associativityHippocampus10420430Year: 200010985281|
|34.||Bullmore E,Sporns O. Complex brain networks: graph theoretical analysis of structural and functional systemsNat Rev Neurosci10186198Year: 200919190637|
|35.||Watrous AJ,Fried I,Ekstrom AD. Behavioral correlates of human hippocampal delta and theta oscillations during navigationJ Neurophysiol1051747175521289136|
|36.||Mormann F,et al. Independent delta/theta rhythms in the human hippocampus and entorhinal cortexFrontHum Neurosci23Year: 2008|
|37.||Libby LA,Ekstrom AD,Ragland JD,Ranganath C. Differential connectivity of perirhinal and parahippocampal cortices within human hippocampal subregions revealed by high-resolution functional imagingJ Neurosci326550656022573677|
|38.||Rutishauser U,Ross IB,Mamelak AN,Schuman EM. Human memory strength is predicted by theta-frequency phase-locking of single neuronsNature464903907Year: 201020336071|
|39.||Rizzuto DS,et al. Reset of human neocortical oscillations during a working memory taskProc Natl Acad Sci U S A10079317936Year: 200312792019|
|40.||Rizzuto DS,Madsen JR,Bromfield EB,Schulze-Bonhage A,Kahana MJ. Human neocortical oscillations exhibit theta phase differences between encoding and retrievalNeuroimage3113521358Year: 200616542856|
|41.||Womelsdorf T,et al. Modulation of neuronal interactions through neuronal synchronizationScience31616091612Year: 200717569862|
|42.||Yonelinas AP,et al. Effects of extensive temporal lobe damage or mild hypoxia on recollection and familiarityNat Neurosci512361241Year: 200212379865|
|43.||Zhang H,Ekstrom AD. Human Neural Systems Underlying Rigid and Flexible Forms of Allocentric Spatial RepresentationHuman Brain Mapping (in press).|
|44.||Marshuetz C,Reuter-Lorenz PA,Smith EE,Jonides J,Noll DC. Working memory for order and the parietal cortex: an event-related functional magnetic resonance imaging studyNeuroscience139311316Year: 200616417974|
|45.||Tolman EC. Cognitive Maps in Rats and MenPsychol Rev55189208Year: 194818870876|
|46.||Howard MW,Kahana MJ. A distributed representation of temporal contextJournal of Mathematical Psychology46269299Year: 2002|
|47.||Tulving E. Episodic memory: from mind to brainAnnu Rev Psychol53125Year: 200211752477|
|48.||Conner CR,Ellmore TM,Pieters TA,DiSano MA,Tandon N. Variability of the relationship between electrophysiology and BOLD-fMRI across cortical regions in humansJ Neurosci31128551286521900564|
|49.||Tandon N. Cortical mapping by electrical stimulation of subdural electrodes: language areasTextbook of epilepsy surgery10011015Informa HealthcareNew York, NYYear: 2008|
|50.||Nunez PL,Srinivasan R. Electric fields of the brain : the neurophysics of EEGOxford University Press, OxfordNew YorkYear: 2006|
|51.||Delorme A,Makeig S. EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics including independent component analysisJ Neurosci Methods134921Year: 200415102499|
|52.||Vinck M,van Wingerden M,Womelsdorf T,Fries P,Pennartz CM. The pairwise phase consistency: a bias-free measure of rhythmic neuronal synchronizationNeuroimage5111212220114076|
|53.||Cox RW. AFNI: software for analysis and visualization of functional magnetic resonance neuroimagesComput Biomed Res29162173Year: 19968812068|
Keywords: human memory, ECoG, network analysis, oscillations, phase synchronization, delta, theta.
Previous Document: Prefrontal atrophy, disrupted NREM slow waves and impaired hippocampal-dependent memory in aging.
Next Document: Mephedrone (Methylmethcathinone) in Toxicology Casework: A Northern Ireland Perspective.