用户,您好!欢迎您进入大气污染防治技术进展平台!登录 | 注册  NSTL重点领域信息门户  帮助中心  RSS
  您当前的位置: 首页 > 网页快照
Partially overlapping spatial environments trigger reinstatement in hippocampus and schema representations in prefrontal cortex | Nature Communications
Methods .Participants . A total of 32 right-handed participants were recruited from the Tucson community and were compensated for their time. Four participants were excluded from the analysis due to excessive movement (>1 voxel), and one participant was excluded due to an incidental finding. Therefore, the final sample size was comprised of 27 participants (17 females, mean age: 22.52 years, range: 18–35 years). All participants had normal or normal-to-corrected vision and normal color perception. Based on self-report, all participants were screened to ensure they had no neurological conditions. The study was approved by the Institutional Review Board at the University of Arizona and written Informed consent was obtained from each participant prior to the experiment.Materials . The experiment consisted of two sessions: an encoding session outside the scanner and a retrieval session inside the scanner. Three different virtual environments were created using Unity3D ( https://unity3d.com ). The three different cities contained stores arranged in a circle, and each consisted of six different stores located on the edge of the circle. One store (“Camera Store”, which was consistent across three environments) was in the center of the circle (Fig.? 1a ), allowing us to manipulate temporal duration while holding spatial distance constant. All three cities had the same basic layout (Fig.? 1c ), including the same ground and wall textures; thus, cities only varied in terms of what stores were shared or distinct across cities (Fig.? 1a ).The degree of similarity between cities was proportionate to the number of overlapping edge stores between cities (we did not take the center “Camera Store” into consideration, because the center store was the same across the three cities, Fig.? 1b ). Specifically, Cities 1 and 2 were the same except the “Store 4” in City 1 was changed to “Store 7” in City 2. Cities 2 and 3 were the same except the “Store 2” and “Store 6” in City 2 were changed to “Store 8” and “Store 9” in City 3, respectively. Therefore, Cities 1 and 2 had 5 overlapping stores (i.e., “Store 1”, “Store 2”, “Store 3”, “Store 5” and “Store 6”); Cities 2 and 3 had four overlapping stores (i.e., “Store 1”, “Store 3”, “Store 5” and “Store 7”); Cities 1 and 3 had three overlapping stores (i.e., “Store 1”, “Store 3” and “Store 5”). All the three cities also had three overlapping stores (i.e., “Store 1”, “Store 3” and “Store 5”, Fig.? 1b ).For the temporal interval task, participants encoded two different durations: 8?s (i.e., “Store 3”, “Store 5”, “Store 6” and “Store 9”) and 16?s (i.e., “Store 1”, “Store 2”,” Store 8”, “Store 4” and “Store 7”, Fig.? 1a ).Experimental procedures . Prescan encoding . During the encoding session, the participants were trained to perform two tasks involving navigating each city: a spatial distance task and temporal interval task (Fig.? 1d ). Participants were instructed that there would be three cities, with some stores shared across cities (see “Methods”), and that there would be a spatial distance/time interval retrieval task to test whether they could successfully learn this information in each city. At the beginning of each round of the spatial distance task, participants were placed at the center of the city and viewed videos of travel from the center store to peripheral stores in a randomized order. All the participants were asked to learn as much as they could about store locations during traversals. Similarly, in the time interval task, participants were asked to learn the time interval between the center and each store. We consider differences between the distance and interval tasks in a separate paper, with our focus here on retrieval of the different spatial environments, which would be common to both tasks. To ensure that participants did not merely use a counting strategy to encode time interval information, we added a distractor task (“math problem”) during the route from the center of the city to each store. A math problem would pop up in a pseudorandom position during the route to each store. Participants were asked to focus on processing the time interval while at the same time correctly answering the arithmetic question by the time they arrived at the store. This ensured that participants were not using a counting task to encode temporal intervals. Participants needed to determine their answer and submit it when they reached the store. Furthermore, the travel speed from the center to each store was not constant but variable to avoid the possibility that participants could merely take advantage of the speed differences to discriminate time intervals.The encoding process repeated as many times as the participants needed in order to learn spatial distances and temporal intervals before they moved on to the next city. The learning order of type of task (space/time) was randomized across all participants. After participants learned all three cities for one type of task (e.g., spatial distance), they then learned three cities for the other task (e.g., temporal interval task). Before starting the main encoding task, participants also performed a practice session in which they visited three additional stores in a virtual city to familiarize themselves with the main task.After the spatial and time encoding task, we tested participants’ memory for each of the three cities for both the spatial distance and temporal interval by performing a shorter version of the memory retrieval task they would experience in the scanner. The short version retrieval task was the same as the main fMRI memory retrieval task (see fMRI memory retrieval task, Fig.? 1e ) but only included 12 trials of spatial distance questions (not used in fMRI task) and 5 trials of temporal interval questions (used in fMRI task) for each city. If a subject failed to reach the memory accuracy criterion (i.e., 80%), they re-learned and were re-tested on their memory for all three cities.fMRI memory retrieval task . The fMRI retrieval session consisted of six consecutive spatial runs and six consecutive temporal runs (two per city), each including 15 trials and lasting 4?min and 40?s pertaining to a single city and a single task. The order of retrieval runs (spatial or temporal) across participants was fully counterbalanced and was pseudo randomized with rules dictating that no city could be tested twice in a row and that each city must be tested once before a city could be repeated. The spatial and temporal retrieval probes were rendered identically during retrieval. Before starting each retrieval run, text and verbal instructions reminded participants of which city and which type of task they would be retrieving next, followed by a 7.77?s (3 TRs) refresher picture which included all the stores of that city. There were not shown the actual layout of the city just pictures of stores to cue the correct city.A slow event-related design (18.13?s for each trial) was used in this study to better characterize the activation pattern for each trial (Fig.? 1e ). During spatial distance blocks, participants were instructed to retrieve the spatial distance by making judgments of the relative distances of stores in that city. For each trial, participants saw three stores on the screen for 9?s, with one store on the top and two below (Fig.? 1e ). Participants were asked to compare which of the two bottom stores was closer to the upper reference store and indicate their choice by pressing the corresponding key on an MR-compatible button box. A “one” response indicated that the lower-left store was closer to the top store, a “two” indicated the lower right store, and a “three” indicated that the distance from the two bottom stores to the reference store was equal. For temporal trials, the store in the center of the city (“Camera Store”) always appeared on the top of the screen, and two peripheral stores appeared on the bottom. Participants were instructed to judge which of two intervals between the center (top) and bottom stores was shorter. Once participants pressed the button within 9?s, a black outline would appear to indicate that they have completed the current question, while these three stores would stay on the screen until 9?s finished. Next, participants performed an active baseline task for 7.77?s, in which they pressed “one” for the appearance of an “X”, and “two” for the appearance of an “O”95. A self-paced procedure was used to make this task engaging; each letter appeared 0.2?s after the response.One hundred and eighty trials were presented in 12 runs, with half as spatial runs and half as temporal runs. One hundred and eight of these trials (60% of total trials) presented “unequal” comparisons in which the two bottom stores were an unequal spatial or temporal distance from the reference store. Seventy-two of these trials (40% of total trials) presented “equal” comparisons in which the two bottom stores shared an equal spatial or temporal distance from the reference store.fMRI localizer task . After the retrieval task, participants were asked to complete a localizer task involving a vowel counting task, which included two runs, each containing 18 trials (~6?min). This task served as a localizer task to allow the creation of multivariate pattern templates for each of six stores (i.e., “Store 2”, “Store 4”, “Store 6”, “Store 7”, “Store 8”, “Store 9”, see more details in tSNR based fMRI MPS). The structure of each trial in the vowel counting task was the same as in the retrieval task (Fig.? 1e ). Here, participants were asked to count the number of vowels (i.e., “A”, “E”, “I”, “O”, “U”; “Y” did not count) in each of the three store names and then select which of the two bottom stores had the closest number of vowels compared to the store on the top. Participants were asked to perform both vowel counting and X/O judgment task as accurately and quickly as possible.To allow us to build store “templates”, each triad of stores included one old store which was presented in the retrieval task with two new stores for vowel counting. These new stores were randomly selected from 24 unstudied stores. The purpose of the new stores was to allow us to identify unique activation patterns associated with a specific old store while at the same time allowing us to keep the trial structure the same as during the spatial retrieval questions. The positions of the old stores in each triad were counterbalanced such that all position arrangements for each old store were presented (i.e., old store-new store A-new store B, new store C-old store-new store D, new store E-new store F-old store). Thus, each old store was repeated three times in a different position of the triad within a run with an inter-repetition interval ranging from 2 to 12 trials.MRI image data acquisition . All participants were tested immediately following encoding in the Siemens 32-Channel 3?T “Skyra” scanner, located in the University of Arizona. Visual stimuli were projected onto a screen behind the scanner, which was made visible to the participant through a mirror attached to the head coil. Stimuli and responses were presented and recorded by PsychoPy ( https://www.psychopy.org ) on a Windows laptop. High-resolution functional images were acquired using a simultaneous multi-slice whole-brain echo planar imaging (EPI) sequence (interleaved acquisition, TR?=?2590?ms, TE?=?30?ms, flip angle?=?82 degree, field of view (FOV)?=?234?mm, matrix?=?128?×?128, slice thickness?=?1.8?mm, slices?=?84, slice acceleration factor?=?3, phase encoding direction?=?right to left, bandwidth?=?1562?Hz/pixel), adapted from a previous study96. High-resolution structural images were obtained using a 3D, T1-weighted MPRAGE (1?mm 3 isotropic) sequence acquired for the whole brain (FOV?=?256?mm, matrix?=?256?×? 256, slice thickness?=?1?mm, TR?=?2100?ms, TE?=?2.33?ms, flip angle?=? 12 degree, bandwidth?=?190?Hz/pixel). High-resolution anatomical images of the hippocampus and surrounding cortex were acquired with a T2-weighted turbo-spin echo (TSE) anatomical sequence (FOV?=?200?mm?×? 200?mm, matrix?=?448?×;448, TR?=?4200.0?ms, TE?=?93.0?ms, flip angle?=?139 degree, slice thickness?=?1.8?mm, 28 slices, bandwidth?=?199?Hz/pixel). Sequences were acquired perpendicular to the long axis of the hippocampus. An additional coplanar matched-bandwidth high-resolution gradient-echo EPI sequence (TR?=?6120?ms, TE?=?39?ms, slices?=?84, FOV?=?245?mm, matrix?=?128?×?128, flip angle?=?90 degree, bandwidth?=?1446?Hz/pixel) was acquired to aid in registration of the EPI sequence to the high-resolution structural images. B0-field maps were acquired immediately with a gradient recalled echo sequence (TR?=?888.0?ms, TE1??=?4.92?ms,TE2?=?7.38?ms, flip angle?=?90 degree, FOV?=?256?mm, slice thickness?=?3?mm, slices?=?84) following the coplanar matched-bandwidth sequence to correct for inhomogeneities of the magnetic field97. This sequence covered the whole brain, allowing us to correct field distortions for the entire EPI sequence.fMRI data preprocessing . Image preprocessing was performed by using FEAT (FMRI Expert Analysis Tool), version 6.00, implemented in FSL (part of the FSL package; http://www.fmrib.ox.ac.uk/fsl ). The first seven images were automatically discarded from each run by the scanner to allow for scanner to equilibrate. We additional discarded 3 volumes in which the refresher picture was present before the retrieval task started. The EPI images were first corrected for geometric distortion using participants’ field maps97 , 98and underwent motion-correction, temporal filtering (nonlinear high-pass filter with a 100?s cutoff), and slice-timing correction. Six motion parameters were added as confound variables to the model. Residual outlier timepoints were identified using FSL’s motion outlier detection program and integrated as additional confound variables in the first-level general linear model (GLM) analysis. No spatial smoothing was applied for single-trial estimation (see below). All functional images were linearly registered to individual-subject T1 MPRAGE structural volumes in a two-step process via a coplanar matched-bandwidth sequence described above using FLIRT. Registration from structural images to the standard MNI-152 template was further refined using FNIRT nonlinear registration for higher-level group analysis when needed (see below).Single-trial response estimates . The GLMs were performed separately to estimate the activation pattern for each of 180 retrieval trials and 36 localizer trials. In this single-trial model, a Least Square–Separate (LS-S) approach was used, in which the trial of interest was modeled as one regressor, with all other trials modeled as a separate regressor99. Specifically, each single-trial GLM included five regressors: (1) the trial of interest; (2) all other trials; (3) black outline stage; (4) fixation; (5) all incorrect trials within the active baseline task. Each event was modeled at the time of stimulus onset and convolved with a canonical hemodynamic response function (double gamma), whereas the correct baseline trials (X/O judgment task) were not coded and thus were treated as an implicit baseline. To control for the effects of head motion, six motion parameters were included in the GLM model as a covariate. The t-map for each trial was used for multivariate pattern similarity?analysis (MPS) and SVR classification analysis to increase the reliability by normalizing for noise100.Run-based response estimates for SVM classification analysis . The GLMs were performed separately to estimate the activation pattern for each retrieval run. Each single-run GLM included 6 regressors: (1) the remembered trials; (2) forgotten trials; (3) missed trials; (4) black outline stage; (5) fixation; (6) all incorrect trials within the active baseline task. Each event was modeled at the time of stimulus onset and convolved with a canonical hemodynamic response function (double gamma), whereas the correct baseline trials (X/O judgment task) were not coded and thus were treated as an implicit baseline. To control for the effects of head motion, six motion parameters were included in the GLM model as a covariate. This resulted in 4 run-based data points per city per subject. The run-based t-map has greater reliability101and could be used for SVM classification analysis to increase accuracy102and power103.Subfield demarcation and ROIs . Automatic hippocampal subfield segmentation software (ASHS)104 , 105was used to segment the subregions of the MTL based on each participant’s high-resolution T2-weighted MRI image. The MTL was segmented into CA1, CA2/3, DG, and subiculum (SUB), perirhinal cortex (PRC) and entorhinal cortex (ERC) and parahippocampus cortex (PHC). We combined the CA2/3 and DG subfields as finer distinctions cannot be made at the acquired resolution. Single-trial t-map were then obtained within those 6 ROIs (CA1, CA2/3/DG, SUB, ERC, PRC, PHC) for each subject for further MPS and classification analysis. Following a previous study43, the medial PFC mask was defined as a set of three regions within the Brodmann areas (BA) 10, 11, and 32.Temporal signal-to-noise ratio (tSNR) . We adopted voxel-wise tSNR to define the fMRI time series stability106. Specifically, for each MTL ROI, we obtained the voxel-wise tSNR of localizer task by calculating the mean of each voxel’s time series divided by its standard deviation. Then the voxels in each ROI could be ranked from high to low by the intensity tSNR and could be further divided into eight portions by different levels of percentile (i.e., 10th, 20th, 30th, 40th, 50th, 60th, 70th, 80th). For example, “10th percentile of tSNR” means that voxels with tSNR less than the bottom 10% of tSNRs in the ROI were removed from the analysis. Therefore, by applying different percentile tSNR as the threshold of t-stat, we could exclude different levels of influence of spurious voxels in the MPS.tSNR based fMRI multivariate pattern similarity analysis (MPS) .Multivariate patterns of stores . In the localizer task, six studied stores (i.e., “Store 2”, “Store 4”, “Store 6”, “Store 7”, “Store 8”, “Store 9”) were repeated 6 times (3 times per run, see Procedures). Based on the specificity, these six stores could be classified into two categories: unique city stores and stores shared across cities. For example, “Store 4” (only belongs to City 1),”Store 8” (only belongs to City 3), and”Store 9” (only belongs to city 3) are unique city stores because they were only presented in one city, while “Store 2” (belongs to City 1 and City 2), “Store 6” (belongs to City 1 and City 2) and “Store 7” (belongs to City 2 and 3) are stores shared across cities, because they were presented in two cities. Then, we constructed a multivariate pattern template for each of the six studied stores that were presented in the localizer task by averaging the activation patterns (i.e., single-trial t-maps) across six repetitions of a given store. The template of each store could provide a neural measure for a memory trace of each store during memory retrieval. Because the vowel counting task, which occurred at the end of the fMRI session, did not involve spatial retrieval and occurred after participants had retrieved information from all three environments, it is unlikely that the templates contained any environment-specific information and therefore could provide indices to store identity.We then applied MPS by measuring the similarity of activation patterns between each of the six store templates and each remembered trial in both retrieval tasks (spatial and temporal) based on the different thresholds of tSNR in each hippocampal subfield. We followed the approach of Power et al.107and censored TRs with a framewise displacement >0.5?mm. Specifically, to quantify unique store within-city pattern similarity (PS), pairwise Pearson correlation coefficients were calculated by correlating each unique city store template with the activity patterns evoked by correctly retrieved trials that did not include the given store within that specific city. For example, to determine whether participants retrieved a store not contained in a retrieval triad, we correlated the template of “Store 4” (belongs to City 1) from the control task with any correctly retrieved trial which did not include “Store 4” within City 1). Similarly, the unique store between-city PS was the correlation between each unique city store template and the activity pattern elicited by the remembered trials that did not belong to the current city. For example, to determine whether participants retrieved a store not contained in a retrieval triad, we correlated the template of “Store 4” (belongs to City 1 only) from the localizer task with any remembered trial of City 2 and City 3. Finally, the shared store within-city PS was the correlation between each store shared across city template and the activity pattern of each remembered trial that did not include the current store within all the shared cities. For example, to determine whether participants retrieved a store not contained in a retrieval triad, we correlated the template of “Store 2” (belongs to City 1 and City 2) from the localizer task with the activity pattern of remembered of trials without “Store 2” within City 1 and City 2. Since correlations are inherently a pairwise comparison, many correlations were performed and then averaged together for a metric of within-condition similarity. The resulting correlation coefficients were transformed into Fisher’s z-scores and then input into further group analyses.To examine the repulsion hypothesis, we also applied the tSNR based MPS. All the remembered trials of the retrieval task could be divided into three conditions: unique city trials (i.e., the trial that could only be attributed to one city, for example, the triads “Store 1-Store 4-Store 5” was only attributable to City 1 and the triads “Store 8-Store 9-Store 3” was only attributable to City 3, Fig.? 4c , top panel), two shared city trials (i.e., the trial could be attributed to two possible cities, for example, the triads “Store 2-Store 1-Store 6” and “Store 2-Store 6-Store 1” could only be attributed to City 1 and City 2 but not be attributed to city 3, Fig.? 4c , middle panel) and three shared city trials (i.e., the trial could be attributed to three possible cities, for example, the triads “Store 1-Store 3-Store 5” and “Store 1-Store 5-Store 3” could be attributed to City 1, City 2 and City3, Fig.? 4c , bottom panel). We calculated the between-city pattern similarity of the independent trials that corresponded to a specific condition (unique city trials, two shared city trials and three shared city trials) separately for spatial and temporal retrieval tasks. Note that all MPS analyses involved correlating between different runs of retrieval, thus avoiding temporal autocorrelations artificially inflating or biasing results.In addition, we also calculated between-city pattern similarity of trials for a specific condition across spatial and temporal tasks to utilize as many as possible correlations to obtain stable metric within a condition. Note, for those pairs which were included in the between-city PS of unique-city trials calculation, we excluded the unique-city pairs that have overlapping stores, for example, the triad “Store 1-Store 4-Store 6 and triad “Store 1-Store 8-Store 7”, to make the representation of each triad as distinct from each other as possible. Because three shared city trials that involved in the between-city PS calculation are perceptual identical, for example, the correlation between triad “Store 1-Store 3-Store 5” and triad “Store 1-Store 5-Store 3”. We also matched the two shared city trial pairs by selecting the pairs that had the same stores, for example, we only calculated the correlation between triad “Store 1-Store 2-Store 6” and “Store 1-Store 6-Store 2”, we did not calculate the correlation between triad “Store 1- Store 2-Store 6” and triad “Store 1-Store 3-Store 6”. The resulting correlation coefficients were transformed into Fisher’s z-scores and then input into further group analysis.Searchlight-based MPS . To examine the shared spatial layout information across three cities, we applied the MPS throughout the whole brain using searchlight approach60. For each voxel, signals (i.e., single-trial t-maps) were extracted from a cubic ROI containing 343 surrounding voxels throughout each subject’s whole brain. Specifically, to quantify shared-layout information, the same location between-city PS was calculated by correlating the remembered trials (in both spatial and temporal task) that share the same location from different cities using Pearson correlations. For example, the triads “Store 1-Store 2-Store 6” in City 1 were correlated with the triads “Store 1-Store 8-Store 9” in City 3 (Fig.? 1a ). In contrast, the different locations between-city PS was calculated by correlating the remembered trials (in both spatial and temporal task) that come from different locations in different cities. For example, the triads “Store 1-Store 2-Store 6” in City 1 were correlated with the triads “Store 1-Store 8-Store 5” in City 3 (Fig.? 1a ). Note, given that different location pairs usually contained perceptual differences (different stores), which in turn could contribute to lower PS for different locations than same locations, we matched the number of different stores between pairs when calculating PS. For example, there are two different stores between the triads “Store 3-Store 7-Store 6” and “Store 3-Store 7-Store 9” when calculating the same location between-city PS. Accordingly, when calculating the different locations between-city PS, we only consider the pairs that also involve two different stores, for example, the triads between “Store 1-Store 7-Store 6” and “Store 1-Store 7-Store 8”. We transformed these similarity scores into Fisher’s z-scores and compared the differences between the same location and different location pairs. Notably, we only included correctly retrieved trials into consideration and excluded any trials with any censored frames during the duration of the modeled GLM response using a framewise displacement threshold of 0.5?mm. A random-effects model was used for group analyses within the mPFC mask using a cluster-forming threshold of Z ?>?2.6, with p ?error rate, using random field theory).Correlating frontal activity with hippocampal PS . We also examined the role of prefrontal activity in modulating hippocampal PS during retrieval. Because the CA2/3/DG and CA1 were the regions that showed significant city-specific PS and conformed to our holistic hypothesis (see Results), we focused on these two regions and tested whether the ROIs could be modulated by PFC activity. We correlated the activation of each condition (unique city store within-city PS/unique city store between-city PS) in each voxel of the whole brain during retrieval with the corresponding PS in CA2/3/DG and CA1, separately. Because frontal activity was associated with the activity level in other brain regions, which was in turn associated with PS, we conducted a partial correlation analysis by correlating the activation level in each voxel across the whole brain and the corresponding PS of CA2/3/DG and CA1 while controlling for the activation levels of CA2/3/DG and CA1. The resulting Spearman’s rank correlation coefficients were transformed into Fisher’s z-scores and then directly compared between the unique store within-city vs. unique store between-city trials, which was put into further group analyses using a cluster-forming threshold of Z ?>?3.1, with p ?error rate using random field theory, Fig.? 2c ).Classification analysis . Searchlight-based leave-one-city-out SVR classification . To test whether the medial PFC represented the shared spatial layout schema, we performed a linear Support Vector Regression (SVR)108using LIBSVM 3.12 ( https://www.csie.ntu.edu.tw/~cjlin/libsvm/ ) as implemented in MATLAB (The MathWorks) to classify spatial distance using a searchlight approach60. Briefly, for each voxel, t-maps were extracted from a cubic ROI containing 343 surrounding voxels throughout each subject’s whole brain. The idea of the leave-one-city-out classification was that if the mPFC could support shared layout schema, the spatial distances learned from two cities should be able to generalize to the third new city. This is because the new city had the same layout as the two learned cities, even though the new city had new stores which they were not presented in the two learned cities (Fig.? 1c ). First, for each triad, we measured the physical distances by calculating the Euclidean distance between the top store displayed and each of the bottom stores (Fig.? 1e ). Then, we calculated the sum of the two Euclidean distances as the behavioral index of spatial distance of each triad. Notably, we only took correctly retrieved trials into consideration and excluded any trials with any censored frames during the duration of the modeled GLM response using a framewise displacement threshold of 0.5?mm. There were 18 possible Euclidean distances in all, and the Kolmogorov-Smirnov Test (KS-test) for group-level uniform distributions indicated that the frequency of the distance was not uniform ( t ?=?0.168, p ?=?0.016, Supplementary Figure? 5a ) because there were too many trials in the shortest distance category (i.e., distance?=?66) compared to the other distances. However, when the number of the shortest distance trials decreased from 14 to 11, the distribution became uniform ( t ?=?0.137, p ?=?0.08, Supplementary Fig.? 5b ). Then, this step was performed for each individual participant to identify the proper number of the shortest distance trials to ensure a uniform distribution of distances ( P s?>?0.05 in KS-Test). In each iteration of the leave-one-city-out cross-validation, a SVR model was trained on runs from two cities, which generated a prediction value of the runs of the third city based on each cubic ROI’s activation patterns. The accuracy of the SVR prediction was then calculated as Spearman’s rank correlation coefficient between actual and predicted values of the spatial distance index. The resulting correlation coefficients were transformed into Fisher’s z-scores and then input into further group analyses within the mPFC mask using a cluster-forming threshold of Z ?>?3.1, with p ?error rate, using random field theory). Similarly, the same SVR classification analysis was performed on hippocampal ROIs to test whether hippocampus could support shared spatial layout schema during spatial distance retrieval.For temporal interval task, we also performed the searchlight-based leave-one-city-out SVM classification analysis across the whole brain. Similar to spatial distance task, for each triad, the sum of time durations between the top store (the center of the city) and each of the two bottom stores (i.e., 16?+?16?=?32?s; 8?+?8?=?16?s; 8?+?16?=?24?s) was taken as the temporal interval index of the given triad (Fig.? 1e ). Since temporal interval could be divided into only three categories, here, a multi-class SVM classification was more appropriate to be adopted to decode the temporal interval information based on the activation patterns of each cubic ROI. In each iteration, we classify three categories (32?s/16?s/24?s) on two cities and tested on the left-out city. Classification accuracy thus represented the percentage of trials that were correctly categorized by the classifier. We balanced the number of trials in each condition of our classification analysis by randomly selecting the same number of trials for each condition (this procedure was performed both for training and testing set). The resulting classification accuracy map for all participants were input into further group analysis using a cluster-forming threshold of Z ?>?3.1, with p ?error rate, using random field theory). Similarly, the same SVM classification analysis was performed on hippocampal ROIs to test whether hippocampus could support shared layout schema during time duration retrieval.ROI-based classification analysis for city . To examine whether hippocampal subfields contained city-specific information, we performed an ROI-based multivoxel pattern classification analysis to classify three cities using a linear support vector machine (SVM)109using LIBSVM 3.12 ( https://www.csie.ntu.edu.tw/~cjlin/libsvm/ ) implemented in MATLAB (The MathWorks). The classification analysis was conducted on 12 run-based t-maps and with a penalty parameter of 1. Since there were 4 runs per city, in each iteration of the leave-three-run-out cross-validation, we trained the classifier on 9 of retrieval runs (3 runs per city) and used the left out three runs (1 run per city) to test classification accuracy based on each hippocampal ROI’s activation patterns. Specifically, for each iteration in the testing run, the SVM classifier generated a scalar probability estimate of the trial corresponding to 3 categories (City 1, City 2, and City 3). The category with the higher probability was then set as the classifier’s prediction. Classification accuracy thus represented the percentage of runs that were correctly categorized by the classifier. We performed a group analysis on each hippocampus subfield using two-tailed, one-sample t-tests to determine whether the accuracy was above chance levels (i.e., 33.33%).Reporting summary . Further information on research design is available in the? Nature Research Reporting Summary linked to this article. .
From:
系统抽取主题     
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)