Excitability at encoding determines sparse coding of remembered episodic memories in the human hippocampus

Excitability at encoding determines sparse coding of remembered episodic memories in the human hippocampus | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Excitability at encoding determines sparse coding of remembered episodic memories in the human hippocampus Catherine W. Tallman, Peter N. Steinmetz, John T. Wixted This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5731906/v2 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 18 Nov, 2025 Read the published version in Scientific Reports → Version 2 posted 10 You are reading this latest preprint version Show more versions Abstract Neurocomputational models hold that individual episodic memories are represented by a sparse, pattern-separated coding scheme in the hippocampus. In addition, recent theories of neuronal allocation suggest that the assignment of individual neurons to a sparse code is non-random and is associated with intrinsic neural excitability. Here, utilizing an independent dataset of single-unit recordings from epilepsy patients, we demonstrate that a relatively small proportion of high-firing hippocampal neurons represent a single item within a recognition memory test. Critically, only items that were both remembered and showed heightened excitability during encoding were preferentially allocated to a sparse, pattern-separated code, one that was selectively present in the hippocampus. Our findings suggest that individual episodic memories are represented by a sparse distributed coding scheme and that neuronal excitability guides the preferential allocation of hippocampal neurons into sparse codes, which in turn supports subsequent retrieval. Biological sciences/Neuroscience/Cognitive neuroscience Biological sciences/Psychology/Human behaviour sparse coding episodic memory human hippocampus single-unit recording neuronal allocation subsequent memory Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Introduction The structures of the medial temporal lobe (MTL), particularly the hippocampus, support the encoding, consolidation, and retrieval of declarative memory (Squire, 2004). One form of declarative memory—episodic memory—refers to the ability to recollect a single event within its original spatial and temporal context (Tulving, 1972). By definition, episodic memories are formed rapidly, and they are often studied using experimental paradigms that involve single-trial learning. Here, we focus on how single neurons in the human hippocampus are allocated to encode and subsequently retrieve individual episodic memories. Neurocomputational models hold that episodic memories are represented in the hippocampus according to a sparse, distributed, and pattern-separated coding scheme (Marr, 1971; McClelland et al., 1995; Norman & O’Reilly, 2003; Treves & Rolls, 1994). In a sparse coding scheme, a single episodic memory is associated with relatively few hippocampal neurons (lifetime sparseness), and each neuron is associated with relatively few individual episodic memories (population sparseness) (Willmore & Tolhurst, 2001) . Moreover, individual episodic memories are theoretically encoded in separate and distinctive neural assemblies, not in assemblies that share neurons. Such non-overlapping sparse codes are predicted to reduce catastrophic interference, protecting against massive forgetting that would otherwise occur (French, 1999; McClelland et al., 1995). Another theoretically important issue concerns the assignment of hippocampal neurons to a sparse code during the encoding of an episodic memory. This allocation process may not be random. For example, work with rodents has found that “intrinsically excitable” neurons are preferentially recruited to code fear-conditioning memories in the amygdala (Kim et al., 2013; Silva et al., 1998, 2009) and CA1 (Cai et al., 2016). It seems reasonable to hypothesize that a similar principle may apply to the recruitment of neurons to form non-overlapping episodic memory representations in the human hippocampus. Single-unit recordings of neurons in the human MTL provide a unique opportunity to test the predictions of neurocomputational models of sparse coding and the neuronal allocation mechanisms that theoretically underlie episodic memory. In the laboratory, episodic memory is often operationalized in terms of recognizing items from a recently presented list, which is the methodology used in the current report . The participant’s task on a recognition memory test is to distinguish between previously encountered old items from the list (targets) and newly presented items (foils). Therefore, the neural episodic memory signal should theoretically not only reflect activity that is preferentially associated with old items compared to new items, but should also be (1) specific to one or, perhaps, a few old items from the list (i.e., lifetime sparseness), (2) specific to a small assembly of neurons (i.e., population sparseness), (3) preferentially associated with neurons that were excitable during encoding, and (4) specific to the hippocampus. A sparse episodic memory signal contrasts with the neural signature of the more commonly investigated generic memory signal, whereby neurons respond differentially to the entire category of old items vs. new items. This generic episodic memory signal is not item specific, and it has been observed not only in the hippocampus but also in the amygdala, anterior cingulate, prefrontal cortex, and parietal cortex (Rutishauser et al., 2008, 2010, 2015; Urgolites et al., 2022). The fact that this signal is both generic and ubiquitous suggests that it may not be the fundamental episodic memory signal that neurocomputational models have long predicted. Several previous studies have reported evidence of sparsely coded episodic memories based on single-unit recordings in the human hippocampus (Urgolites et al., 2022; Wixted, 2014, 2018). However, those studies were based on recordings produced from one research group that used the same continuous recognition memory task of word learning. Here, we take advantage of a different dataset of single-unit recordings from the MTL that was made publicly available by another research group (Chandravadia et al., 2020; Faraut et al., 2018). This independent dataset differed from the first dataset in ways that should not influence the predictions of neurocomputational models. Specifically, the dataset differed in the episodic memory task design (old/new recognition vs. continuous recognition), stimulus material (images vs. words), and spike sorting algorithm. Thus, it is an opportunity to replicate prior work in an independent dataset with several experimental differences that theoretically should not influence the sparsely coded item-specific episodic memory signal. More importantly, this dataset also afforded the opportunity to investigate the relationships among sparse coding, neuronal allocation, and subsequent memory. Methods We analyzed an independent dataset of single-unit neural recordings from the medial temporal lobe of epilepsy patients during an old/new recognition memory task. Details of the experimental protocol are described fully in previous work (Rutishauser et al., 2010, 2015). The open-source dataset was originally made available by (Faraut et al., 2018) and was updated by (Chandravadia et al., 2020). The experimental protocols associated with this dataset were approved by the Institutional Review Boards of the California Institute of Technology, the Huntington Memorial Hospital, and Cedars-Sinai Medical Center. Participants Patients with intractable epilepsy (N=59) underwent depth electrode monitoring in preparation for potential surgical resection of the seizure loci. All participants were volunteers and provided written informed consent. From the 87 sessions available in the OSF data repository (https://osf.io/hv7ja/), sessions with recordings not labeled as the medial temporal lobe (n=4), and sessions with negative d’ values were excluded (n=4), resulting in 79 sessions and 55 patients (F=23, M=32, µ age = 37.2 years) included in the analyses reported here. Microwire Implantation All recordings were performed with macro-micro depth electrodes, each with eight 40 μm diameter microwires. Broadband recordings (0.1–9,000 Hz filter) were sampled at 32 kHz using a Neuralynx Inc. system (either Atlas or Cheetah). Electrodes were included in the analysis if they were localized to the hippocampus or amygdala and were excluded if inter-ictal epileptic activity was observed. Spike detection and sorting Continuous recordings of the raw signal were bandpass filtered 300–3,000 Hz for each channel. Spikes were detected by convolving the filtered raw trace, defined by threshold crossings of an energy signal, with a kernel of approximately the width of an action potential. Detected spikes were then sorted using the semiautomatic template-matching algorithm OSort, which was used to identify clusters that represented single neurons. Electrodes were localized using MNI coordinates defined by post-operative MRI scans which were registered to the MNI152-aligned CIT168 template brain35 atlas. Experimental Design Each session consisted of a recognition memory task with a distinct encoding and retrieval phase. During the encoding phase (Figure 1a), participants studied 100 images from five different visual categories: animals, people, cars/vehicles, outdoor scenes/houses, and flowers/food items. They were instructed to carefully study the images as their memory would be tested later. Each learning trial consisted of a 1-second delay (blank screen), a 1- or 2-second presentation of an image, another .5-second delay (blank screen), and were then prompted with a yes/no encoding question (i.e., “Is this an animal?), with unlimited time to respond. Participants studied each image once except during 4 sessions in which patients studied only 50 images due to poor performance in categorizing images during the encoding question. After the encoding phase, participants completed a distractor task (approximately 15 minutes) to inhibit active rehearsal. They then completed an old/new recognition memory test with confidence ratings. During the retrieval phase (Figure 1b), 100 images were shown; 50 were previously shown during the learning phase (“old”; targets) and 50 were not presented before during learning (“new”; foils). Patients used a button box to respond on a 1-6 confidence scale (1=new, very sure; 2=new, sure; 3=new, guess; 4=old, guess; 5=old, sure; 6=old, very sure). Each recognition trial consisted of a 1-second delay (blank screen), a 1- or 2- second presentation of a target or foil item, another .5-second delay (blank screen), and were then prompted to submit a confidence rating (i.e., 1-6) with unlimited time to respond. Statistical Analysis For each single unit, and for each image presented during both the learning phase and the recognition phase, we analyzed spike counts that occurred during the pre-stimulus period (1000ms-200ms before the onset of the image) and the post-stimulus period (200ms-1000ms after the onset of each image). In addition to examining raw spike counts, for some analyses, the recordings were normalized as follows: for each image ( i ) and each neuron ( j ), spike counts during the pre-stimulus period were computed across all learning and recognition trials in that session, yielding the pre-stimulus mean (µ prestim j ) and standard deviation (σ prestim j ) for each neuron. Next, normalized post-stimulus spike counts were computed for every neuron’s response to every test image (N norm ij ) using the following equation: N norm ij = (N poststim ij - µ prestim j )/σ prestim j , where N poststim ij represents the number of post-stimulus spike counts recorded for neuron j in response to image i . The generic recognition memory signal is, by definition, not item specific and instead consists of a statistically significant difference in the average firing rate to old items compared to new items on the recognition test (see relevant analyses in supplemental materials). Our focus here is on detecting a sparsely coded episodic memory signal, if it exists. A sparse episodic memory signal consists of a small fraction of neurons selectively responding to a small fraction of “old” items. Thus, a different approach is needed to detect this signal. Critically, testing for a neuron’s signal associated with a single old item can be performed only once. Testing the same neuron’s response to the same item multiple times would create new episodic memories of the test itself, which might influence each subsequent test. Multiple testing of neuron-item pairs might also simply identify a concept neuron (i.e., a semantic memory signal). A single neuron exhibiting an item-specific response on a single test would not necessarily reflect a sparse episodic memory signal for that old item. For example, it might instead reflect a semantic memory signal, one that would have occurred even if the word were new, or it might simply reflect an artifactual noise signal that happened to coincide with the presentation of the test item. However, both semantic memory signals and artifactual noise signals would be expected to occur with equal probability to old items (targets) and new items (foils). Therefore, the prediction of interest applies to a comparison of the full distributions of item-by-neuron spike counts for targets vs. foils. To identify the presence of the sparse coding signal, the normalized target-by-neuron and foil-by-neuron response spike count distributions were compared both visually and statistically, and for each medial temporal lobe region separately (hippocampus: 73,539 item-by-neuron recordings, amygdala: 99,944 item-by-neuron recordings). To compare them visually, empirical quantile-quantile (QQ) plots were generated. The goal of this method is to visually assess whether two datasets come from the same or similar distributions. A linear pattern indicates that the two distributions have the same shape even if they have different means and standard deviations (i.e., even if they differ in the first two moments). Non-linear patterns reveal differences in skewness (the third moment) or tail behavior (i.e., kurtosis, the fourth moment) between the two distributions. Statistically, the moments associated with the two empirical distributions were compared using bootstrap analyses. The empirical data consisted of N t target-by-neuron normalized spike counts and N f foil-by-neuron normalized spike counts. The sparse coding signal would be evident as an interaction effect between the two MTL regions of interest (hippocampus and amygdala) and the skewness of the two distributions (target and foils) in which the difference in skewness between the target and foil distributions is greater in the hippocampus relative to the corresponding difference in the amygdala. First, for each statistical moment (i.e., mean, standard deviation, skewness, and kurtosis), the observed interaction value was calculated (e.g., the difference in skewness values between the target and foil distributions for the amygdala was subtracted from the corresponding difference for the hippocampus). Next, we combined the N t and N f empirical recordings to conduct the bootstrapping procedure. For each of B = 10,000 bootstrap trials, a bootstrapped “target” distribution of N t observations was created by randomly sampling with replacement from the combined empirical data, and a bootstrapped “foil” distribution of N f observations was created by randomly sampling with replacement from the combined empirical data. We then calculated the difference in the relevant statistical measure (e.g., skewness) between these two randomly sampled distributions. Repeating this process for B = 10,000 bootstrap trials yielded 10,000 bootstrap difference scores for the relevant statistic. The proportion of bootstrapped trials for which the absolute value of the interaction difference score was greater than or equal to the absolute value of the empirically observed interaction difference score determined the P value. Significance was defined as P < .05 (two-tailed), unless otherwise specified. The bootstrap procedure was repeated for the interaction difference score for each statistical moment separately. To determine the direction of the interaction effects, the same bootstrap procedure was used to examine if the difference in each statistical moment between the target and foil distributions was statistically reliable within the hippocampus and the amygdala separately. After performing this omnibus analysis, we investigated whether the sparse coding signature during retrieval (i.e., a significant interaction in skewness values between the target and foil distributions and between the hippocampus and amygdala) was selectively associated with two specific circumstances that should theoretically give rise to it (i.e., excitable neurons at encoding and subsequent memory). First, we investigated whether the sparse coding signature was selectively present at retrieval for targets associated with excitable neurons during encoding. Within each session, the mean pre-stimulus (1000ms-200ms) and mean post-stimulus (200ms-1000ms) raw spike counts during the encoding phase (i.e., during list presentation) were calculated for each item. Items were then rank ordered by the raw mean spike count value and placed into “Low” or “High” groups using a median split, first for pre-stimulus values and again for post-stimulus values. Items that were at the median value pre- or post-stimulus were excluded from the primary analysis. This resulted in target items being categorized into one of four categories based on the level of firing pre- and post-stimulus presentation: 1) increase in firing: Low-High, 2) consistently high firing: High-High, 3) decrease in firing: High-Low, and 4) consistently low firing: Low-Low. Items most strongly associated with heightened excitability were defined to be those in the Low-High category, and the question of interest was whether a difference in skewness between the target-by-neuron response distribution and foil-by-neuron response distribution would be limited to those items. Second, we investigated whether the sparse coding signature was selectively associated with items that were successfully remembered (i.e., correct “old” decisions on the recognition test). To address this issue, after classifying targets into one of four categories based on their responsivity during encoding (e.g., Low-Low), the targets were further classified within each category as remembered or forgotten. The question of interest was whether a difference in skewness between the target-by-neuron response distribution and foil-by-neuron response distribution would be limited not only to targets most associated with excitable neurons (Low-High) but also to a further subset of those targets that were successfully retrieved during the recognition test. Results Behavioral Performance On average, participants performed above chance on the recognition memory test for old items, 67.7% correct, SEM = 0.2% (i.e., hit rate = .677) and for foil items, 73.9% correct, SEM = 0.2% (i.e., false alarm rate = 1 - .739 = .261). The average dʹ across sessions was 1.24 (min = .05, max = 2.47). Four sessions with negative dʹ values were excluded. The item-specific episodic memory signal was selectively detected in the hippocampus A sparse episodic memory signal consists of a small fraction of neurons selectively responding to a small fraction of “old” items (items studied before the recognition test). In the simplest case, a single neuron would exhibit a strong response to a single old item but not to any other old item nor to any new item. In other words, the neuron’s response would be (1) item specific, and (2) specific to that item only if it is old. We investigated this issue by comparing the full distributions of item-by-neuron spike counts for targets vs. foils. Consider a simplified concrete example. If recordings were made from 5 single neurons during the presentation of 20 items on the recognition test (10 old items and 10 new items), the two distributions of interest would consist of the 50 recordings to old items (5 neurons X 10 old items) and the 50 recordings to new items (5 neurons X 10 new items). The two distributions—which we refer to as the target-by-neuron response distribution and the foil-by-neuron response distribution—should be a similar shape when compared to one another. For example, a singular neuron would respond strongly to only one target item, but not to any of the other target items, nor to any of the foil items. Thus, the distributions of the target-by-neuron recordings and the foil-by-neuron recordings may be a similar shape, with the only difference being a very small percentage of the target-by-neuron recordings showing a differentially strong response. Statistically, the prediction is that the target-by-neuron response distribution should be significantly more skewed to the right compared to the foil-by-neuron response distribution. Moreover, this difference in skewness should reflect the influence of a very small percentage of target-by-neuron responses. Therefore, the difference in skewness should disappear after removing a small percentage of the highest values from both distributions. To visualize the shapes of these distributions, we used empirical QQ plots (Chambers et al., 1983, see Methods). In this type of plot, the quantiles of one dataset are plotted on one axis, and the corresponding quantiles of the second dataset are plotted on the other axis. Here, we compared the quantiles of the foil-by-neuron normalized spike count distribution (x-axis) vs. the quantiles of the target-by-neuron normalized spike count distribution (y-axis), separately for the hippocampus and amygdala (Figure 2a & 2b). Deviations from linearity arise when the distributions differ in moments beyond the first (mean) and second (standard deviation). Of most interest here, an upward deflection away from the diagonal at the upper right of the QQ plot indicates a difference in the third moment (skewness). More specifically, it indicates that the target item distribution is more skewed to the right than the foil item distribution (a feature that might not be visually apparent in a plot of the frequency distributions). Note that points on a QQ plot are densely grouped and represent thousands of recordings that have similar values for both distributions, so there are many more points on the plot than it appears. For the hippocampus, an upward deflection is evident at higher x and y values, indicating that the target (i.e., old item) distribution is more positively skewed compared to the foil (i.e., new item) distribution. To investigate whether the apparent difference in skewness is due to a small percentage of observations in the target-by-neuron response distribution, we removed the top 0.25% of the highest-ranking recordings from both distributions. In a plot of the remaining 99.75% of recordings, the deflection was no longer apparent (Figure 2c & 2d), thus providing visual evidence of sparse coding in hippocampal neurons. For the amygdala, a less pronounced downward deflection was observed, indicating differentially higher spiking for a small proportion of neurons in the foil distribution. Similarly, after we removed the top 0.25% of the highest-ranking recordings from both distributions, the deflection towards the x-axis was no longer apparent. We next examined whether the visual deflections were statistically significant, focusing mainly on the skewness measures of the two distributions (Figure 3; Table 1). First, and most importantly, we tested for an interaction between the two MTL regions of interest (hippocampus and amygdala) and the skewness of the two distributions (target and foils) using bootstrap tests (see Methods). Theoretically, the sparse coding signal would be evident as an interaction effect such that the magnitude of the target–minus-foil difference in skewness would be larger in the hippocampus than in the amygdala. A one-tailed test was used based on the specific directional hypotheses predicted by Urgolites et al. (2022). The interaction was significant ( P = .049), replicating the significant interaction reported by Urgolites et al. (2022). Table 1. Statistical Results for the Statistical Moments of the Target vs. Foil Distributions at Retrieval Note. A significant interaction was defined as a greater difference in a given statistical moment (e.g., skewness) between the target-vs.-foil normalized spike count distributions in the hippocampus compared to the corresponding difference in the amygdala according to bootstrap tests (B = 10,000, P < .05, one-tailed). The interactions for mean ( P = .047) and skewness ( P = .049) were significant. Differences in each statistical moment between the two distributions (target-vs.-foil) were also considered separately within each structure (hippocampus and amygdala) using bootstrap tests (B = 10,000, P < .05, one-tailed). In the hippocampus, differences in the upper moments of the target and foil spike count distributions were not significant but trended in that direction (standard deviation: P = .054, skewness: P = .070, kurtosis: P = .086). In the amygdala, the differences were not significant (standard deviation: P = .344, skewness: P = .254, kurtosis: P = .178). Finally, mean firing in the amygdala distinguished between target and foil items ( P = .019) but not in the hippocampus ( P = .334). ~ = P < 0.10, * = P < .05. Next, one-tailed bootstrap tests were performed to assess the directionality of the effects (Table 1). In the hippocampus, the target-by-neuron response distribution was more positively skewed (4.17) than the foil-by-neuron response distribution (3.05), as also reported by Urgolites (2022), but the difference here was not quite significant ( P = .070). Conversely, in the amygdala, the skewness of the target-by-neuron response distribution (3.0) was slightly less than the skewness of the foil-by-neuron response distribution (3.28), but the difference did not approach significance ( P = .254). In the amygdala, the only significant difference was in the means of the distributions, with foil items eliciting significantly more firing on average than targets (targets = .14, foils = .16; P = .019). The theoretically important skewness interaction reported by (Urgolites et al., 2022) was evident in this independent dataset as well, and it is consistent with an item-specific sparse coding signal being selectively observed in the hippocampus. Moreover, as described in more detail next, the interaction became more visually apparent and more statistically reliable when examining subsets of targets that, theoretically, should exhibit the same effects. The item-specific episodic memory signal was related to excitability at encoding Next, we tested a theory of neuronal allocation according to which episodic memories are assigned to neurons with heightened excitability at learning. We operationalized a test of this theory by repeating the same sparse coding analysis just described (visual tests using QQ plots, statistical tests based on skewness), but this time by partitioning retrieval data for targets by the pattern of spiking observed before and during item presentation at encoding. Within session, each target was assigned to one of four categories determined by their relative firing pre- and post- stimulus presentation at encoding (see Methods, Low-Low, Low-High, High-High, High-Low). Based on the theoretical predictions of sparse coding and neuronal allocation, we expected the item-specific episodic memory signal would remain selective to the hippocampus and would further be selective to the subset of targets which were associated with a relative increase in excitability at encoding (low firing pre-stimulus and high firing post-stimulus item presentation at learning; Low-High). Additionally, we expected that the episodic memory signal would remain undetected in the amygdala (see Figures 2 & 3 and Table 1), regardless of spiking observed before and during item presentation at encoding. We repeated the analyses used to identify the sparse coding signal (see Methods and Section 1) based on the skewness of the target-by-neuron distributions and the foil-by-neuron distributions, separately for each category of relative spiking at encoding, beginning with a visual inspection of the relevant QQ plots (Figure 4). A sharp upward deflection of the QQ plot (indicative of greater skewness for the target distribution relative to the foil distribution) was present only for the targets that were categorized as Low-High at encoding (i.e., the excitable neurons) and was further selective to hippocampal neurons (Figure 4a). For the amygdala, by contrast, almost all points were densely grouped along the red diagonal line (Figure 4b), providing no evidence of a skewness difference for targets with heightened excitability at encoding. The upward deflection was not visually apparent on any of the remaining QQ plots, either in the hippocampus or the amygdala, for the subsets of targets in which spiking remained high (High-High: Figure 4c and 4d), spiking decreased (High-Low: Figure 4e and 4f), or remained low (Low-Low: Figure 4g and 4h) (Supplemental Figure 1). We next tested the statistical significance of the visual trends in the QQ plots using bootstrapping tests (see Methods; B = 10,000, Bonferroni corrected: P < 0.0125, two-tailed). With regard to the measure of most interest, skewness, a significant interaction was observed between the MTL regions (hippocampus vs amygdala) and trial type (Table 2: targets labeled Low-High at encoding vs foil items; P = 0.008). For the hippocampus, the Low-High target-by-neuron response distribution was significantly more skewed than the foil-by-neuron response distribution (6.79 vs. 3.06, P = .004). In contrast, in the amygdala, the skewness of the corresponding distributions did not differ significantly (3.27 vs. 3.28, P = .989). No significant interactions for the difference in skewness values between targets and foils and the MTL regions were present for the other subsets of targets in which spiking decreased (High-Low), remained high (High-High), or remained low (Low-Low). Additionally, interaction tests performed on the other statistical moments (mean, standard deviation, and kurtosis) were not significant for any of the subsets of targets, including the targets with heightened excitability at encoding (Low-High). (Supplemental Table 1). Table 2. Statistical Results for Skewness of the Target vs. Foil Distributions at Retrieval Partitioned by the Pattern of Target Firing at Encoding Note. A significant interaction was defined as a greater difference in a given statistical moment between the target-vs.-foil normalized spike count distributions in one region (e.g., the hippocampus) compared to the corresponding difference the other region (e.g., the amygdala) according to bootstrap tests (B = 10,000, Bonferroni corrected , P < .0125). The interaction was significant for the test involving targets associated with excitable neurons at encoding (Low-High: P =.008) but not for any other category of targets at encoding (High-High: P = .803, High-Low: P = .891, or Low-Low: P = .473). Within the hippocampus, a significant difference between target vs. foil skewness (theoretically indicative of the item-specific memory signal) was observed only for targets that were associated with excitability at encoding (Low-High: P = .004). Within the amygdala, no significant difference in skewness between the target and foil distributions was observed. Corresponding analyses for the other statistical moments are reported in Supplemental Table 1. ** = P < 0.01 Subsequent memory was associated with the item-specific episodic memory signal in the hippocampus Next, QQ plots were generated separately to compare the distributions of the normalized spike counts of remembered targets to foil items (hippocampus: Figure 5a & amygdala: Figure 5b, hippocampus) and forgotten targets to foil items (hippocampus: Figure 5c & amygdala: Figure 5d) In the hippocampus, greater skewness of the target-by-neuron distribution is apparent in the QQ plots when the targets were subsequently remembered (Figure 5a) but not when they were forgotten (Figure 5c). Indeed, when the targets were forgotten, the QQ plot trend is in the opposite direction. No apparent nonlinear trends were evident in the QQ plots for the amygdala whether the targets were remembered or forgotten (Figure 5b and 5d). The non-linear deflection in the hippocampal QQ plot (Figure 5a) was no longer apparent after removing a small fraction of the target-by-neuron and foil-by-neuron recordings (Supplemental Figure 2), indicating relatively few high firing neurons contributed to the greater skewness of the remembered target-by-neuron distribution when compared to foil items. Statistically, bootstrapping tests determined the interaction between the skewness of the two distributions (remembered targets vs. foils) and brain regions (hippocampus vs. amygdala) was significant (Table 3, B = 10,000, P = .027, two-tailed). Considering the hippocampus only, skewness was significantly greater for the remembered target distribution compared to the foil item distribution (4.82 vs. 3.05, P = .036). The interaction test for a difference in kurtosis of the normalized spike count distributions between MTL regions (remembered targets vs. foil items) was significant ( P = .037) and marginally significant for standard deviation ( P = .083). Within each region considered separately, a significant difference in the remembered vs. foil item distributions was also selectively observed in the hippocampus for standard deviation (1.26 vs. 1.18, P = .015) and a marginally significant difference was present for kurtosis (72.3 vs. 29.9, P = .064). Similar effects in the hippocampus were not observed for forgotten targets, and no such effects were observed in the amygdala for either remembered (Table 3) or forgotten targets (Table 4). Table 3. Analyses of the Statistical Moments of Remembered Target vs. Foil Distributions at Retrieval Note. A significant interaction was defined as a greater difference in a given statistical moment between the target-vs.-foil normalized spike count distributions in one region (e.g., the hippocampus) compared to the corresponding difference the other region (e.g., the amygdala) according to bootstrap tests (B = 10,000 , P < .05). The interaction between the two distributions (remembered targets vs. foils) and brain region (hippocampus vs. amygdala) was significant for skewness ( P = .027) and kurtosis ( P = .037). Within each region considered separately, significant differences were observed for standard deviation ( P = .015) and skewness ( P = .036) and was marginally significant for kurtosis ( P = .064). These effects were not present in the amygdala. ~ = P < .10, * = P < .05. Table 4. Analyses of the Statistical Moments of Forgotten Target vs. Foil Distributions at Retrieval Note. A significant interaction was defined as a greater difference in a given statistical moment between the target-vs.-foil normalized spike count distributions in one region (e.g., the hippocampus) compared to the corresponding difference the other region (e.g., the amygdala) according to bootstrap tests (B = 10,000 , P < .05). The interaction between a given statistical moment of the two distributions (forgotten targets vs. foils) and brain region (hippocampus vs. amygdala) was significant for the mean only ( P < .001). Within each region considered separately, the difference between the mean of the forgotten vs. foils item distribution was significant in the amygdala ( P < .001) but not the hippocampus ( P = .470). ~ = P < 0.10, *** = P < .001. We next performed a direct comparison between the remembered and forgotten targets. The analyses just described compared the target-by-neuron normalized spike count distribution (y-axis of the QQ plot) and the foil-by-neuron normalized spike count distribution (x-axis of the QQ plot), separately for remembered targets and forgotten targets. In the next analysis, we examined the QQ plot with the target-by-neuron distribution for remembered targets plotted on the y-axis and the target-by-neuron distribution for forgotten targets on the x-axis (Figure 6). In other words, unlike the analyses described thus far, recordings to foil items were not included in this analysis. In the hippocampus, a difference in skewness was visibly apparent (Figure 6a). Conversely for amygdala recordings, almost all points fell on the diagonal line, and no deflection was observed in the QQ plot (Figure 6b). The non-linear deflection in the hippocampal QQ plot (Figure 6a) was no longer apparent after removing a small fraction of the target-by-neuron and foil-by-neuron recordings (Supplemental Figure 3), indicating relatively few high firing neurons contributed to the greater skewness of the remembered target-by-neuron distribution when compared to forgotten items. A significant interaction was observed for skewness values between brain region (hippocampus vs. amygdala) compared to the status of the test item (Table 5; remembered targets compared to forgotten targets, P = .034). Within the hippocampus, skewness was significantly greater for the remembered target distribution compared to the forgotten item distribution (4.82 vs. 2.30, P = .022), but the corresponding comparison in the amygdala was not significant (2.94 vs. 3.15, P = .732). Although not relevant to the theoretical issues of interest here, it is worth noting that a significant interaction was observed between the MTL regions for the difference in the means for remembered compared to forgotten targets (Table 5, P = .0015). Within the amygdala, mean firing was significantly greater for the remembered item distribution (.16) compared to the forgotten item distribution (.09), but the corresponding difference in the hippocampus was not significant. Note that this effect in the amygdala is the generic memory signal that distinguished between general categories of stimuli (remembered versus forgotten), rather than representing a single episodic memory. Table 5. Analyses of the Statistical Moments of Remembered Target vs. Forgotten Target Distributions at Retrieval Note. A significant interaction reflects a greater difference in a statistical moment (e.g., skewness) of the spike count distributions for the remembered vs. forgotten targets in one region compared to the other region (hippocampus vs. amygdala) according to bootstrap tests (B = 10,000, P < 0.05). The interactions were significant for the mean ( P = .0015) and skewness ( P = .034), and marginally significant for kurtosis ( P = .059). Within each region considered separately, differences in the upper moments of the distributions were detected in the hippocampus (skewness: P = .022, kurtosis: P = .0501) but not the amygdala (skewness: P = .732 and kurtosis: P = .789). Conversely, mean firing was greater for remembered compared to forgotten items in the amygdala ( P < .001) but not in the hippocampus ( P = .156). Although the interaction was not significant for standard deviation, within each region, both the hippocampus exhibited a significant difference ( P = .020) and the amygdala exhibited a marginally significant difference ( P = .055), with greater SD values for remembered compared to forgotten items. * = P < 0.05, ** = P < 0.01, *** = P < 0.001. The item-specific episodic memory signal was selectively detected for remembered but not forgotten targets associated with excitability at encoding Thus far, we have identified two theoretically motivated and distinct subsets of targets for which the target-by-neuron response distribution exhibits greater skewness than the foil-by-neuron response distribution: 1) targets with heightened excitability at encoding and 2) targets that were subsequently remembered. Thus, we posited that the item-specific episodic memory signal in the hippocampus (i.e., elevated skewness resulting from a small percentage of recordings) would be selectively identified for targets that were both excitable at encoding and subsequently remembered. Indeed, as might be expected, the skewness difference (theoretically reflective of sparse coding) was visually identified only in the QQ plots that were selective to the hippocampus and involving targets that were (1) associated with increased excitability during encoding and (2) were also subsequently remembered (Figure 7a). If the targets that were associated with increased excitability during encoding were subsequently forgotten, no such effect was evident (Figure 7c). This pattern also did not emerge in the amygdala for targets that exhibited increased excitability during encoding, whether those targets were later remembered (Figure 7b) or forgotten (Figure 7d). Additionally, no other encoding level-by-subsequent memory combinations within the hippocampus or amygdala (Supplemental Figure 5) exhibited the skewness difference that we hypothesize arises from sparse coding. The remarkable specificity of these findings is best appreciated by examining a bar graph (Figure 8) showing the four statistical moments of the target-by-neuron response distribution (mean, standard deviation, skewness, and kurtosis) broken down by brain region (amygdala and hippocampus) and excitability at encoding. The four statistical moments are shown separately in the four plots arrayed from left to right, with remembered targets shown in the top row of four graphs (Figure 8a-d) and forgotten targets shown in the bottom row of four graphs (Figure 8e-h). Excitability at encoding is color-coded within each of the eight graphs, with High-High in red, High-Low in orange, Low-Low in green, and Low-High in blue. The moments for the foil-by-neuron response distribution are shown in gray. For the foil items, the value for a given moment (e.g., the mean) is the same for the remembered-target graphs on top and the forgotten-target graphs on the bottom (i.e., the remembered vs. forgotten distinction does not apply to foil items). Note in particular, the measures shown in column C (skewness), where the blue bar in the top right graph of that column (neural recordings from the hippocampus, involving excitable targets that were also subsequently remembered) conspicuously stands out among all other bars. Essentially the same pattern is also evident for kurtosis in column D and is also somewhat apparent for the standard deviation in column B. In summary, the most specific and theoretically relevant subset of targets, items that were later remembered and were also associated with heightened excitability at encoding, were the only targets with evidence of sparse coding. Moreover, this effect was evident only in the hippocampus. Discussion Neurocomputational models hold that relatively few neurons represent distinct episodic memories, even similar ones, in distinct, pattern-separated sparse neural assemblies (Marr, 1971; McClelland et al., 1995; Norman & O’Reilly, 2003; Treves & Rolls, 1994). The formation of such representations should avoid catastrophic forgetting and facilitate the later retrieval of events that occurred in a particular spatial and temporal context (i.e., it should facilitate the retrieval of episodic memories). In addition, more recent theories of neural allocation hold that, during encoding, excitable neurons are differentially allocated to sparsely represent episodic memories (Silva et al., 2009), an idea that has been supported by work with rodents (Kim et al., 2013; Silva et al., 1998). To test these mechanisms of sparse coding and neuronal allocation in humans, we analyzed an open-source dataset consisting of single-unit recordings from the MTL that were made while epilepsy patients were engaged in an old/new recognition task for a list of images (Chandravadia et al., 2020; Faraut et al., 2018). Our findings replicated previously reported evidence of an item-specific (pattern-separated) sparse code in humans that was also specific to the hippocampus (Urgolites et al., 2022; Wixted, 2014, 2018). Additionally, here, we found that the most specific and theoretically relevant subset of targets—those that were associated with heightened excitability during encoding and were also subsequently remembered—yielded the clearest evidence of a sparse, pattern-separated code. Critically, the presence of this sparse code was only evident in the hippocampus. Together, the results provide clear evidence that excitable neurons are preferentially allocated to create a sparse, pattern-separated episodic memory code in the hippocampus that, in turn, supports subsequent retrieval. An alternative view accepts the possibility that episodic memories may be represented as a pattern-separated code in rats and monkeys but rejects the idea that the same is true of humans (Quiroga, 2020). According to this perspective, single unit recording studies from the human MTL have thus far yielded no evidence in support of predictions made by longstanding neurocomputational models. The findings reported here serve as a rebuttal to this claim. Moreover, some of the work that gave rise to the idea that newly formed episodic memories are encoded via overlapping neural assemblies (Ison et al., 2015) do not actually support that perspective (see supplemental material). Related work has identified “episode-specific” neurons in the human hippocampus, which are neurons that encode the combination of elements within an individual episode (Kolibius et al., 2023). The episodic memory signal of interest here was item-specific, not generic. As noted earlier, a generic memory signal—whereby a neuron fires differentially to the category of old items vs. new items—has often been detected in the hippocampus , the amygdala, and multiple regions of the cortex (Rutishauser et al., 2008, 2010, 2015; Urgolites et al., 2022). This research has provided valuable insights, but our emphasis here is on the item-specific memory signal in the hippocampus that neurocomputational models have long proposed as the foundation of episodic memory. In accordance with those models and with later theoretical developments, our findings provide the first clear evidence that the human hippocampus selectively represents single episodic memories and facilitates remembering using a sparse pattern-separated code, which is preferentially allocated to excitable neurons during encoding. Declarations Author contributions: C.W.T contributed data analysis, data interpretation, and manuscript writing. P.N.S. contributed data curation and statistical guidance at all stages of the analysis and manuscript writing. J.T.W. supervised the project and contributed to all aspects of the manuscript. All of the authors discussed the results at all stages of the project. Competing interests: C.W.T., P.N.S., and J.T.W. have no competing interests. Data Availability Statement The dataset analyzed during the current study is available in the following Open Science Framework repository: https://osf.io/hv7ja/ (Chandravadia et al., 2020) References Cai, D. J., Aharoni, D., Shuman, T., Shobe, J., Biane, J., Song, W., Wei, B., Veshkini, M., La-Vu, M., Lou, J., Flores, S. E., Kim, I., Sano, Y., Zhou, M., Baumgaertel, K., Lavi, A., Kamata, M., Tuszynski, M., Mayford, M., … Silva, A. J. (2016). A shared neural ensemble links distinct contextual memories encoded close in time. Nature , 534 (7605), 115–118. https://doi.org/10.1038/nature17955 Chambers, J. M., Cleveland, W. S., Kleiner, B., & Tukey, P. (1983). Graphical Methods of Data Analysis . Duxbury Press. Chandravadia, N., Liang, D., Schjetnan, A. G. P., Carlson, A., Faraut, M., Chung, J. M., Reed, C. M., Dichter, B., Maoz, U., Kalia, S. K., Valiante, T. A., Mamelak, A. N., & Rutishauser, U. (2020). A NWB-based dataset and processing pipeline of human single-neuron activity during a declarative memory task. Scientific Data , 7 (1), 78. https://doi.org/10.1038/s41597-020-0415-9 Faraut, M. C. M., Carlson, A. A., Sullivan, S., Tudusciuc, O., Ross, I., Reed, C. M., Chung, J. M., Mamelak, A. N., & Rutishauser, U. (2018). Dataset of human medial temporal lobe single neuron activity during declarative memory encoding and recognition. Scientific Data , 5 (1), 180010. https://doi.org/10.1038/sdata.2018.10 French, R. M. (1999). Catastrophic forgetting in connectionist networks. Trends in Cognitive Sciences , 3 (4), 128–135. https://doi.org/10.1016/S1364-6613(99)01294-2 Ison, M. J., Quiroga, R. Q., & Fried, I. (2015). Rapid encoding of new memories by individual neurons in the human brain. Neuron , 87 , 220–230. Kim, J., Kwon, J.-T., Kim, H.-S., & Han, J.-H. (2013). CREB and neuronal selection for memory trace. Frontiers in Neural Circuits , 7 . https://doi.org/10.3389/fncir.2013.00044 Kolibius, L. D., Roux, F., Parish, G., Ter Wal, M., Van Der Plas, M., Chelvarajah, R., Sawlani, V., Rollings, D. T., Lang, J. D., Gollwitzer, S., Walther, K., Hopfengärtner, R., Kreiselmeyer, G., Hamer, H., Staresina, B. P., Wimber, M., Bowman, H., & Hanslmayr, S. (2023). Hippocampal neurons code individual episodic memories in humans. Nature Human Behaviour , 7 (11), 1968–1979. https://doi.org/10.1038/s41562-023-01706-6 Marr, D. (1971). Simple memory: A theory for archicortex. Philos (R. S. Lond, Trans.). B Biol. Sci , 262 , 23–81. McClelland, J., McNaughton, B., & O’Reilly, R. (1995). Why there are complementary learning systems in the hippocampus and neocortex: Insights from the successes and failures of connectionist models of learning and memory. PSYCHOLOGICAL REVIEW , 102 (3), 419–457. https://doi.org/10.1037/0033-295X.102.3.419 Norman, K. A., & O’Reilly, R. C. (2003). Modeling hippocampal and neocortical contributions to recognition memory: A complementary-learning-systems approach. Psychological Review , 110 (4), 611–646. https://doi.org/10.1037/0033-295X.110.4.611 Quiroga, R. Q. (2020). No pattern separation in the human hippocampus. Trends Cogn. Sci , 24 , 994–1007. Rutishauser, U., Ross, I. B., Mamelak, A. N., & Schuman, E. M. (2010). Human memory strength is predicted by theta-frequency phase-locking of single neurons. Nature , 464 (7290), 903–907. https://doi.org/10.1038/nature08860 Rutishauser, U., Schuman, E. M., & Mamelak, A. N. (2008). Activity of human hippocampal and amygdala neurons during retrieval of declarative memories. Proceedings of the National Academy of Sciences , 105 (1), 329–334. https://doi.org/10.1073/pnas.0706015105 Rutishauser, U., Ye, S., Koroma, M., Tudusciuc, O., Ross, I. B., Chung, J. M., & Mamelak, A. N. (2015). Representation of retrieval confidence by single neurons in the human medial temporal lobe. Nature Neuroscience , 18 (7), 1041–1050. https://doi.org/10.1038/nn.4041 Silva, A. J., Kogan, J. H., Frankland, P. W., & Kida, S. (1998). CREB AND MEMORY. Annual Review of Neuroscience , 21 (Volume 21, 1998), 127–148. https://doi.org/10.1146/annurev.neuro.21.1.127 Silva, A. J., Zhou, Y., Rogerson, T., Shobe, J., & Balaji, J. (2009). Molecular and Cellular Approaches to Memory Allocation in Neural Circuits. Science , 326 (5951), 391–395. https://doi.org/10.1126/science.1174519 Squire, L. R. (2004). Memory systems of the brain: A brief history and current perspective. Neurobiology of Learning and Memory , 82 (3), 171–177. https://doi.org/10.1016/j.nlm.2004.06.005 Treves, A., & Rolls, E. T. (1994). Computational analysis of the role of the hippocampus in memory. Hippocampus , 4 , 374–391. Tulving, E. (1972). Episodic and semantic memory. In E. Tulving & W. Donaldson (Eds.), Organization of memory . Academic Press. Urgolites, Z. J., Wixted, J. T., Goldinger, S. D., Papesh, M. H., Treiman, D. M., Squire, L. R., & Steinmetz, P. N. (2022). Two kinds of memory signals in neurons of the human hippocampus. Proceedings of the National Academy of Sciences , 119 (19), e2115128119. https://doi.org/10.1073/pnas.2115128119 Willmore, B., & Tolhurst, D. J. (2001). Characterizing the sparseness of neural codes. Network: Computation in Neural Systems , 12 (3), 255. https://doi.org/10.1088/0954-898X/12/3/302 Wixted, J. T. (2014). Sparse and distributed coding of episodic memory in neurons of the human hippocampus. Proc. Natl. Acad. Sci. U.S.A , 111 , 9621–9626. Wixted, J. T. (2018). Coding of episodic memory in the human hippocampus. Proc. Natl. Acad. Sci. U.S.A , 115 , 1093–1098. Additional Declarations No competing interests reported. Supplementary Files SRSupplemental3425Tallman.docx Cite Share Download PDF Status: Published Journal Publication published 18 Nov, 2025 Read the published version in Scientific Reports → Version 2 posted Editorial decision: Revision requested 02 May, 2025 Reviews received at journal 28 Apr, 2025 Reviews received at journal 23 Apr, 2025 Reviewers agreed at journal 13 Mar, 2025 Reviewers agreed at journal 05 Mar, 2025 Reviewers invited by journal 05 Mar, 2025 Editor assigned by journal 05 Mar, 2025 Editor invited by journal 05 Mar, 2025 Submission checks completed at journal 04 Mar, 2025 First submitted to journal 18 Feb, 2025 You are reading this latest preprint version Show more versions Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-5731906","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[{"code":1,"date":"2025-01-24 05:59:50","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"articleType":"Article","associatedPublications":[],"authors":[{"id":446383625,"identity":"d7b4fe0c-4547-404d-b47e-4c1a07e8064e","order_by":0,"name":"Catherine W. Tallman","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAwklEQVRIiWNgGAWjYHACxgMPgCQ/iMkD5DUQ1MDGwHAgAUhLNpCsxeAAsVr45zc/OJCYY5dnfCP58Yc3DDayGw4Q0CJxjM3gQOK25GKzG2lmknMY0owJajFgYwBpYU7cdiOHjZmH4XAiEVrYPwC11CdunpHD/JmH4T8xWnhAtgANl8hhkOZhOEBYi8SxnAKgluOJM848A/rFINl4JiEt/M3HNz74uK06sb8dFGIVdrJ9hLSgu5M05aNgFIyCUTAKcAAAdspHo4dZ5TYAAAAASUVORK5CYII=","orcid":"","institution":"University of California San Diego, Department of Psychology, San Diego, CA, USA","correspondingAuthor":true,"prefix":"","firstName":"Catherine","middleName":"W.","lastName":"Tallman","suffix":""},{"id":446383626,"identity":"32b6cff1-ef8d-4954-ac0a-22d363f5e640","order_by":1,"name":"Peter N. Steinmetz","email":"","orcid":"","institution":"Neurtex Brain Research Institute, Dallas, TX, USA","correspondingAuthor":false,"prefix":"","firstName":"Peter","middleName":"N.","lastName":"Steinmetz","suffix":""},{"id":446383627,"identity":"18da361d-e292-425e-aba3-e0eab1bfabbd","order_by":2,"name":"John T. Wixted","email":"","orcid":"","institution":"University of California San Diego, Department of Psychology, San Diego, CA, USA","correspondingAuthor":false,"prefix":"","firstName":"John","middleName":"T.","lastName":"Wixted","suffix":""}],"badges":[],"createdAt":"2024-12-30 01:25:20","currentVersionCode":2,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-5731906/v2","doiUrl":"https://doi.org/10.21203/rs.3.rs-5731906/v2","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41598-025-21967-7","type":"published","date":"2025-11-18T15:58:53+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":81200551,"identity":"5fd93c42-edd7-4e1a-9c63-06cdd4f2319f","added_by":"auto","created_at":"2025-04-23 11:10:24","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":408818,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eRecognition Memory Procedure used by Faraut et al. (2018)\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eNote\u003c/em\u003e. Recognition memory task experimental design (reproduced and adapted from Faraut et al., 2018 under Creative Commons License 4.0). The published dataset consists of single-unit recordings from the hippocampus and amygdala from 59 epilepsy patients while completing a recognition memory test, for a total of 89 experimental sessions (Faraut et al., 2018, Chandravadia et al., 2020). Each session consisted of an encoding and retrieval phase. During the encoding phase (a), participants studied images (n=100) from five different visual categories: animals, people, cars/vehicles, outdoor scenes/houses, and flowers/food items. Each trial consisted of a delay (1s blank screen), the presentation of an image (1- or 2- seconds), a second delay (0.5s blank screen), followed by a yes/no question to promote encoding (i.e., “Is this an animal?), with unlimited time to respond. (b) Approximately 15 minutes after the encoding phase, participants completed an old/new recognition memory test during the retrieval phase. Each trial consisted of a delay (1s blank screen), the presentation of an image (1- or 2- seconds), a second delay (0.5s blank screen), followed by an old/new recognition judgement with confidence ratings (unlimited time to respond). A total of 100 images were shown; 50 were previously studied during the encoding phase (“old”; targets) and 50 were not previously presented during the encoding phase (“new”; foils).\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-5731906/v2/919930bd9c3cf420b2e10ade.png"},{"id":81201251,"identity":"dd7b287f-682c-4a60-8f3c-5bd35878a7ca","added_by":"auto","created_at":"2025-04-23 11:18:24","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":480997,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eEmpirical QQ Plots of\u003c/strong\u003e \u003cstrong\u003eTarget- vs. Foil-by-Neuron Normalized Spike-Count Distributions at Retrieval\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eNote. \u003c/em\u003eEmpirical quantile-quantile (QQ) plots of target-by-neuron normalized spike count distributions (y-axis) vs. foil-by-neuron normalized spike-count distributions (x-axis) in the hippocampus (a and b) and the amygdala (c and d). The top panels plot the distributions for 100% of the recordings from the hippocampus (panel a, 72,538 recordings) and the amygdala (panel b, 99,944 recordings). Points that are densely grouped (dark grey) represent thousands of recordings whereas less densely grouped points (light grey) represent relatively few recordings. If the points all fell on the red line of equivalence, it would indicate that the two distributions are identical. If the points deviate from the red line but still yield a linear relationship, it would indicate that the two distributions have the same form (e.g., both exponential or both lognormal) but differ in their location and scale parameters (e.g., \u0026nbsp;different mean and/or standard deviation). The observed deflection of points away from the red line of equivalence towards one axis in (a) indicates that the target-by-neuron distribution has a different form than the foil-by-neuron distribution. More specifically, the deflection indication that the target distribution is more skewed to the right relative to the foil distribution. A less pronounced effect in the opposite direction is evident in the amygdala (b). The deflection in (a) is predicted by a sparse coding account, which further predicts that the deflection reflects the strong responses of a small percentage of neurons in response to target items. After removing the top 0.25% of both the target-by-neuron recordings and foil-by-neuron recordings (c and d), the deflections in (a) and (b) were no longer apparent. The skewness values of the two distributions (targets vs. foils) were compared to determine if the theoretically predicted deflection in panel (a) was statistically reliable. The difference between the target vs. foil distributions was significantly different between the brain regions (hippocampus vs. amygdala) and was marginally significant only within the hippocampus (refer to Methods and Table 1 for detailed statistical reporting; * = \u003cem\u003eP\u003c/em\u003e \u0026lt; .05, ~ = \u003cem\u003eP\u003c/em\u003e \u0026lt; .10).\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-5731906/v2/11c9b90a28dbea82233c261c.png"},{"id":81202014,"identity":"d65b24f0-580b-4c44-9738-18232aebdbb0","added_by":"auto","created_at":"2025-04-23 11:26:24","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":258757,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eStatistical Moments of Target- and Foil-by-Neuron Normalized Spike-Count Distributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eNote. \u003c/em\u003eStatistical moments of the target-by-neuron and foil-by-neuron normalized spike count distributions that yielded the QQ plots shown in Figures 1a and 1b: (a) mean, (b) standard deviation, (c) skewness, and (d) kurtosis. Within each panel, data from the amygdala are shown on the left, and data from the hippocampus are shown on the right. Significant differences between the target (blue) and foil (grey) distributions within a region, and significant interactions between regions (hippocampus vs. amygdala), are indicated by horizontal lines above the relevant bars (~ = \u003cem\u003eP\u003c/em\u003e \u0026lt; .10, * = \u003cem\u003eP\u003c/em\u003e \u0026lt; .05). As detailed in Table 1, a significant interaction was observed for mean values (a), with greater average spiking in the amygdala in response to foil items compared to targets, with no significant difference observed for the hippocampus. A marginally significant interaction was present for standard deviation (b) and kurtosis (d), with greater values for the target distribution relative to the foil distribution in the hippocampus, but not the amygdala. Critically, a significant interaction between the hippocampus and amygdala was present for skewness (c), the statistical moment most relevant to the item-specific memory signal under investigation here.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-5731906/v2/cd814eb93f11812c47ff8fc1.png"},{"id":81199548,"identity":"016227a9-4919-4dfb-8d04-8d56916ded0f","added_by":"auto","created_at":"2025-04-23 11:02:24","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":428812,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eEmpirical QQ Plots of the Normalized Spike-Count Distributions at Retrieval Partitioned by Neuronal Excitability at Encoding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eNote\u003c/em\u003e. As in Figure 2, normalized spike count distributions were plotted to visualize differences in skewness between the target-by-neuron distributions (y-axis) and foil-by-neuron (x-axis) distributions. Here, separate plots were created for each of the four spiking patterns at encoding (Low-High, High-High, High-Low, and Low-Low; see Methods). Thus, four panels show retrieval data from the hippocampus partitioned by the encoding pattern observed in the hippocampus (panels a, c, e, and g), and four panels show retrieval data from the amygdala partitioned by the encoding pattern observed in the amygdala (panels b, d, f, and h). The same foil-by-neuron (x-axis) distribution based on recordings from the hippocampus is used in all four plots for the hippocampus, and the same foil-by-neuron (x-axis) distribution based on recordings from the amygdala is used in all four plots for the amygdala. Visually, a sharp deflection towards the y-axis (indicative of greater skewness for the target distribution relative to the foil distribution) is evident for targets associated with heightened excitability at encoding (Low-High) in the (a) hippocampus but not in the (b) amygdala. As detailed in Table 2, the interaction between the skewness of the two distributions (targets vs. foils) and brain region (hippocampus vs. amygdala) was significant, as denoted by the horizontal line above panels a \u0026amp; b with double asterisks. Within the hippocampus, the apparent visual difference in skewness was also significant (as denoted by ** in panel (a)), but no significant difference was detected for the amygdala. No visual or statistical evidence of a skewness difference between the target and foil distributions was observed for targets associated with the other spiking patterns at encoding, for either the hippocampus or amygdala Thus, the difference in skewness was statistically reliable only in the hippocampus, and only for targets that were associated with excitable hippocampal neurons at encoding. Supplemental Figure 1 reports graphs with the top 0.25% of both the target-by-neuron and foils-by-neuron distributions were removed. The deflection observed in panel (a) disappeared after removing the top 0.25% of data, indicating that relatively few neurons fired strongly in response to targets mostly associated with excitable neurons at encoding compared to foil items, and only within the hippocampus. ** = \u003cem\u003eP\u003c/em\u003e \u0026lt; 0.01\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-5731906/v2/842f374eaef5b2ca640ecb67.png"},{"id":81199552,"identity":"5c7d1fe6-0634-4913-8c79-31bce685d811","added_by":"auto","created_at":"2025-04-23 11:02:24","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":395308,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eEmpirical QQ Plots for Remembered Targets vs. Foils and Forgotten Targets vs. Foils\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNote. As in Figures 2 \u0026amp; 3, normalized spike count distributions were plotted to visualize differences in skewness between the target-by-neuron (y-axis) and foil-by-neuron (x-axis) distributions, separately for targets that were remembered (top row) or forgotten (bottom row), and for both the hippocampus (a \u0026amp; c) and amygdala (b \u0026amp; d). Visually, in the hippocampus, a sharp deflection towards the y-axis (indicative of greater skewness of the target distribution relative to the foil distribution) is visually evident for remembered targets (a). A smaller trend in the opposite direction was observed in the amygdala (b). As detailed in Table 3, the interaction between the skewness of the two distributions (remembered targets vs. foils) and brain regions (hippocampus vs. amygdala) was significant, as denoted by the line with an asterisk above panels (a) and (b). Within each region considered separately, the difference in skewness for the remembered target-by-neuron distribution was significantly greater than the foil-by-neuron distribution (denoted by * in the lower right of panel a) in the hippocampus, but the slight difference in the opposite direction in the amygdala was not significant. No visual or statistical evidence of a difference in skewness was evident for forgotten targets in either the hippocampus (c) or amygdala (d). Thus, the difference in skewness associated with the item-specific memory signal (the theoretical signature of a sparse episodic memory code) was statistically reliable only in the hippocampus, and only for targets that were remembered. Supplemental Figure 2 reports graphs with the top 0.25% of both the target-by-neuron and foils-by-neuron distributions were removed. The deflection observed in panel (a) disappeared after removing the top 0.25% of data, indicating that relatively few neurons fired strongly in response to remembered targets compared to foil items, and only within the hippocampus. * = \u003cem\u003eP\u003c/em\u003e\u0026lt; .05.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-5731906/v2/da29a49c81fda73aa632c7ea.png"},{"id":81199553,"identity":"31c3149f-e178-459b-893a-525ee91b5fec","added_by":"auto","created_at":"2025-04-23 11:02:24","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":228819,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eEmpirical QQ Plots for Remembered vs. Forgotten Targets\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNote. Normalized spike count distributions were again plotted to visualize differences in skewness between the remembered target-by-neuron (y-axis) and forgotten item-by-neuron (x-axis) distributions for the (a) hippocampus and (b) amygdala. Visually, a sharp deflection towards the y-axis was evident for the remembered vs. forgotten item spike count distributions in the (a) hippocampus but not in the (b) amygdala recordings. As detailed in Table 5, the interaction between the skewness of the two distributions (remembered targets vs. forgotten targets) and brain region (hippocampus vs. amygdala) was significant, as denoted by the * above the horizontal line connecting panels (a) and (b). Within each region considered separately, the difference in skewness for the remembered target distribution was significantly greater than the forgotten target distribution in the hippocampus (denoted by * in lower right of panel a) but not in the amygdala. The deflection observed in panel (a) disappeared after removing the top 0.25% of data (Supplemental Figure 3), indicating that relatively few neurons fired strongly in response to remembered targets compared to forgotten targets, and only within the hippocampus. * = P \u0026lt; 0.05.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-5731906/v2/52458c2f2578442530888d4e.png"},{"id":81199554,"identity":"51723e90-6a24-4b61-9dcf-089e04baf70a","added_by":"auto","created_at":"2025-04-23 11:02:24","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":375628,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eEmpirical QQ Plots for Excitable Neurons as a Function of Subsequent Memory\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eNote. \u003c/em\u003eNormalized empirical QQ plots of the excitable (i.e., Low-High pattern at encoding) target-by-neuron response distribution (y-axis) and the foil-by-neuron response distribution (x-axis). Panel (a) plots remembered targets associated with excitable hippocampal neurons at encoding, panel (b) plots remembered targets associated with excitable amygdala neurons at encoding, panel (c) plots forgotten targets associated with excitable hippocampal neurons at encoding, and panel (d) plots forgotten targets associated with excitable amygdala neurons at encoding. In panel (a), a sharp deflection towards the y-axis is evident in the hippocampus for remembered targets that were associated with heightened excitability at encoding. By contrast, in the amygdala (b), most points fell densely on the diagonal line. Additionally, no other subset of targets split by other levels of spiking at encoding (Low-Low, High-High, High-Low) for remembered or forgotten items showed visual evidence of a difference in skewness (Supplemental Figure 4). The deflection observed in panel a disappeared after removing the top 0.25% of data, indicating that relatively few neurons fired strongly in response to remembered targets compared to foil items, and only within the hippocampus (Supplemental Figure 5). Thus, the difference in skewness associated with the item-specific memory signal was statistically reliable and selective to only the hippocampus, only to targets that were remembered, and only to targets associated with heightened excitability at encoding (bootstrapped, B\u003cem\u003e \u003c/em\u003e= 10,000, \u003cem\u003eP\u003c/em\u003e \u0026lt; 0.05). ** = \u003cem\u003eP\u003c/em\u003e \u0026lt; 0.01.\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-5731906/v2/1f7e12f9f35bf2bf45486d8c.png"},{"id":81199555,"identity":"144ca042-40d1-4fad-89a2-274c5f304b69","added_by":"auto","created_at":"2025-04-23 11:02:24","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":370332,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eStatistical Moments of Item-by-Neuron Normalized Spike Distributions by Neuronal Excitability, Subsequent Memory, and Region\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eNote.\u003c/em\u003e Bar charts representing the statistical moments of the normalized spike counts at retrieval for targets, categorized by their 1) level of spiking at encoding (red: neurons that were consistently high spiking [High-High]; orange: neurons that decreased in spiking [High-Low]; green: neurons that were consistently low spiking [Low-Low]; blue: neurons with heightened excitability [Low-High]) and foils (grey) and 2) subsequent memory response (remembered items: top row, forgotten items: bottom row). The primary statistical moment suggestive of sparse coding, (c) skewness, was conspicuously elevated for targets with heightened excitability at encoding (changed from low pre-stimulus to high post-stimulus firing at encoding, [Low-High: blue]). Additionally, elevations in the (a) mean, (b) standard deviation, and (d) kurtosis values of the spike counts at retrieval were observed for targets that changed from low pre-stimulus to high post-stimulus firing at encoding (Low-High; blue) compared to foils (grey) for the hippocampus. This pattern was selective only to targets that were later remembered (top row) and not forgotten (bottom row).\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-5731906/v2/5c4fb3e1052f94ab094d10cc.png"},{"id":96650460,"identity":"2e8e6f6a-359c-4911-9f1b-b8d69a8aef35","added_by":"auto","created_at":"2025-11-24 16:12:08","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4449284,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5731906/v2/b75c223e-d9f9-48da-94a5-daadeae83df6.pdf"},{"id":81199557,"identity":"a61e6c4b-bd6e-4346-b876-769594762c41","added_by":"auto","created_at":"2025-04-23 11:02:26","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":27464776,"visible":true,"origin":"","legend":"","description":"","filename":"SRSupplemental3425Tallman.docx","url":"https://assets-eu.researchsquare.com/files/rs-5731906/v2/935991113658248c52300585.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"\u003cp\u003eExcitability at encoding determines sparse coding of remembered episodic memories in the human hippocampus\u003c/p\u003e","fulltext":[{"header":"Introduction","content":"\u003cp\u003eThe structures of the medial temporal lobe (MTL), particularly the hippocampus, support the encoding, consolidation, and retrieval of declarative memory (Squire, 2004). One form of declarative memory\u0026mdash;episodic memory\u0026mdash;refers to the ability to recollect a single event within its original spatial and temporal context (Tulving, 1972). By definition, episodic memories are formed rapidly, and they are often studied using experimental paradigms that involve single-trial learning. Here, we focus on how single neurons in the human hippocampus are allocated to encode and subsequently retrieve individual episodic memories.\u003c/p\u003e\n\u003cp\u003eNeurocomputational models hold that episodic memories are represented in the hippocampus according to a sparse, distributed, and pattern-separated coding scheme (Marr, 1971; McClelland et al., 1995; Norman \u0026amp; O\u0026rsquo;Reilly, 2003; Treves \u0026amp; Rolls, 1994). In a sparse coding scheme, a single episodic memory is associated with relatively few hippocampal neurons (lifetime sparseness), and each neuron is associated with relatively few individual episodic memories (population sparseness) (Willmore \u0026amp; Tolhurst, 2001)\u0026nbsp; . Moreover, individual episodic memories are theoretically encoded in separate and distinctive neural assemblies, not in assemblies that share neurons. Such non-overlapping sparse codes are predicted to reduce catastrophic interference, protecting against massive forgetting that would otherwise occur (French, 1999; McClelland et al., 1995).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAnother theoretically important issue concerns the assignment of hippocampal neurons to a sparse code during the encoding of an episodic memory. This allocation process may not be random. For example, work with rodents has found that \u0026ldquo;intrinsically excitable\u0026rdquo; neurons are preferentially recruited to code fear-conditioning memories in the amygdala (Kim et al., 2013; Silva et al., 1998, 2009) and CA1 (Cai et al., 2016). It seems reasonable to hypothesize that a similar principle may apply to the recruitment of neurons to form non-overlapping episodic memory representations in the human hippocampus.\u003c/p\u003e\n\u003cp\u003eSingle-unit recordings of neurons in the human MTL provide a unique opportunity to test the predictions of neurocomputational models of sparse coding and the neuronal allocation mechanisms that theoretically underlie episodic memory. In the laboratory, episodic memory is often operationalized in terms of recognizing items from a recently presented list, which is the methodology used in the current report . The participant\u0026rsquo;s task on a recognition memory test is to distinguish between previously encountered old items from the list (targets) and newly presented items (foils). Therefore, the neural episodic memory signal should theoretically not only reflect activity that is preferentially associated with old items compared to new items, but should also be (1) specific to one or, perhaps, a few old items from the list (i.e., lifetime sparseness), (2) specific to a small assembly of neurons (i.e., population sparseness), (3) preferentially associated with neurons that were excitable during encoding, and (4) specific to the hippocampus.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eA sparse episodic memory signal contrasts with the neural signature of the more commonly investigated \u003cem\u003egeneric\u003c/em\u003e memory signal, whereby neurons respond differentially to the entire category of old items vs. new items. This generic episodic memory signal is not item specific, and it has been observed not only in the hippocampus but also in the amygdala, anterior cingulate, prefrontal cortex, and parietal cortex (Rutishauser et al., 2008, 2010, 2015; Urgolites et al., 2022). The fact that this signal is both generic and ubiquitous suggests that it may not be the fundamental episodic memory signal that neurocomputational models have long predicted.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eSeveral previous studies have reported evidence of sparsely coded episodic memories based on single-unit recordings in the human hippocampus (Urgolites et al., 2022; Wixted, 2014, 2018). However, those studies were based on recordings produced from one research group that used the same continuous recognition memory task of word learning. Here, we take advantage of a different dataset of single-unit recordings from the MTL that was made publicly available by another research group (Chandravadia et al., 2020; Faraut et al., 2018). This independent dataset differed from the first dataset in ways that should not influence the predictions of neurocomputational models. Specifically, the dataset differed in the episodic memory task design (old/new recognition vs. continuous recognition), stimulus material (images vs. words), and spike sorting algorithm. Thus, it is an opportunity to replicate prior work in an independent dataset with several experimental differences that theoretically should not influence the sparsely coded item-specific episodic memory signal. More importantly, this dataset also afforded the opportunity to investigate the relationships among sparse coding, neuronal allocation, and subsequent memory.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003eWe analyzed an independent dataset of single-unit neural recordings from the medial temporal lobe of epilepsy patients during an old/new recognition memory task. Details of the experimental protocol are described fully in previous work (Rutishauser et al., 2010, 2015). The open-source dataset was originally made available by (Faraut et al., 2018) and was updated by (Chandravadia et al., 2020). The experimental protocols associated with this dataset were approved by the Institutional Review Boards of the California Institute of Technology, the Huntington Memorial Hospital, and Cedars-Sinai Medical Center.\u003c/p\u003e\n\u003ch2\u003e\u003cem\u003eParticipants\u003c/em\u003e\u003c/h2\u003e\n\u003cp\u003ePatients with intractable epilepsy (N=59) underwent depth electrode monitoring in preparation for potential surgical resection of the seizure loci. All participants were volunteers and provided written informed consent. From the 87 sessions available in the OSF data repository (https://osf.io/hv7ja/), sessions with recordings not labeled as the medial temporal lobe (n=4), and sessions with negative \u003cem\u003ed’\u003c/em\u003e values were excluded (n=4), resulting in 79 sessions and 55 patients (F=23, M=32,\u0026nbsp;µ\u003csub\u003eage\u003c/sub\u003e = 37.2 years) included in the analyses reported here.\u0026nbsp;\u003c/p\u003e\n\u003ch2\u003e\u003cem\u003eMicrowire Implantation\u003c/em\u003e\u003c/h2\u003e\n\u003cp\u003eAll recordings were performed with macro-micro depth electrodes, each with eight 40 μm diameter microwires. Broadband recordings (0.1–9,000 Hz filter) were sampled at 32 kHz using a Neuralynx Inc. system (either Atlas or Cheetah). Electrodes were included in the analysis if they were localized to the hippocampus or amygdala and were excluded if inter-ictal epileptic activity was observed.\u0026nbsp;\u003c/p\u003e\n\u003ch2\u003e\u003cem\u003eSpike detection and sorting\u003c/em\u003e\u003c/h2\u003e\n\u003cp\u003e\u0026nbsp;Continuous recordings of the raw signal were bandpass filtered 300–3,000 Hz for each channel. Spikes were detected by convolving the filtered raw trace, defined by threshold crossings of an energy signal, with a kernel of approximately the width of an action potential. Detected spikes were then sorted using the semiautomatic template-matching algorithm OSort, which was used to identify clusters that represented single neurons. Electrodes were localized using MNI coordinates defined by post-operative MRI scans which were registered to the MNI152-aligned CIT168 template brain35 atlas.\u0026nbsp;\u003c/p\u003e\n\u003ch2\u003e\u003cem\u003eExperimental Design\u003c/em\u003e\u003c/h2\u003e\n\u003cp\u003eEach session consisted of a recognition memory task with a distinct encoding and retrieval phase. During the encoding phase (Figure 1a), participants studied 100 images from five different visual categories: animals, people, cars/vehicles, outdoor scenes/houses, and flowers/food items. They were instructed to carefully study the images as their memory would be tested later. Each learning trial consisted of a 1-second delay (blank screen), a 1- or 2-second presentation of an image, another .5-second delay (blank screen), and were then prompted with a yes/no encoding question (i.e., “Is this an animal?), with unlimited time to respond. Participants studied each image once except during 4 sessions in which patients studied only 50 images due to poor performance in categorizing images during the encoding question.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAfter the encoding phase, participants completed a distractor task (approximately 15 minutes) to inhibit active rehearsal. They then completed an old/new recognition memory test with confidence ratings. During the retrieval phase (Figure 1b), 100 images were shown; 50 were previously shown during the learning phase (“old”; targets) and 50 were not presented before during learning (“new”; foils). Patients used a button box to respond on a 1-6 confidence scale (1=new, very sure; 2=new, sure; 3=new, guess; 4=old, guess; 5=old, sure; 6=old, very sure). Each recognition trial consisted of a 1-second delay (blank screen), a 1- or 2- second presentation of a target or foil item, another .5-second delay (blank screen), and were then prompted to submit a confidence rating (i.e., 1-6) with unlimited time to respond.\u003c/p\u003e\n\u003ch2\u003e\u003cem\u003eStatistical Analysis\u003c/em\u003e\u003c/h2\u003e\n\u003cp\u003eFor each single unit, and for each image presented during both the learning phase and the recognition phase, we analyzed spike counts that occurred during the pre-stimulus period (1000ms-200ms before the onset of the image) and the post-stimulus period (200ms-1000ms after the onset of each image). In addition to examining raw spike counts, for some analyses, the recordings were normalized as follows: for each image (\u003cem\u003ei\u003c/em\u003e) and each neuron (\u003cem\u003ej\u003c/em\u003e), spike counts during the pre-stimulus period were computed across all learning and recognition trials in that session, yielding the pre-stimulus mean (µ\u003csub\u003eprestim\u003c/sub\u003e\u003cem\u003e\u003csub\u003ej\u003c/sub\u003e\u003c/em\u003e) and standard deviation (σ\u003csub\u003eprestim\u003c/sub\u003e\u003cem\u003e\u003csub\u003ej\u003c/sub\u003e\u003c/em\u003e) for each neuron. Next, normalized post-stimulus spike counts were computed for every neuron’s response to every test image (N\u003csub\u003enorm\u003cem\u003eij\u003c/em\u003e\u003c/sub\u003e) using the following equation: N\u003csub\u003enorm\u003cem\u003eij\u003c/em\u003e\u003c/sub\u003e = (N\u003csub\u003epoststim\u003c/sub\u003e\u003cem\u003e\u003csub\u003eij\u003c/sub\u003e\u003c/em\u003e - µ\u003csub\u003eprestim\u003c/sub\u003e\u003cem\u003e\u003csub\u003ej\u003c/sub\u003e\u003c/em\u003e)/σ\u003csub\u003eprestim\u003c/sub\u003e\u003cem\u003e\u003csub\u003ej\u003c/sub\u003e\u003c/em\u003e, where N\u003csub\u003epoststim\u003c/sub\u003e\u003cem\u003e\u003csub\u003eij\u003c/sub\u003e\u003c/em\u003e represents the number of post-stimulus spike counts recorded for neuron \u003cem\u003ej\u003c/em\u003e in response to image \u003cem\u003ei\u003c/em\u003e. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;The generic recognition memory signal is, by definition, not item specific and instead consists of a statistically significant difference in the average firing rate to old items compared to new items on the recognition test (see relevant analyses in supplemental materials). Our focus here is on detecting a sparsely coded episodic memory signal, if it exists. A sparse episodic memory signal consists of a small fraction of neurons selectively responding to a small fraction of “old” items. Thus, a different approach is needed to detect this signal. Critically, testing for a neuron’s signal associated with a single old item can be performed only once. Testing the same neuron’s response to the same item multiple times would create new episodic memories of the test itself, which might influence each subsequent test. Multiple testing of neuron-item pairs might also simply identify a concept neuron (i.e., a semantic memory signal).\u003c/p\u003e\n\u003cp\u003eA single neuron exhibiting an item-specific response on a single test would not necessarily reflect a sparse episodic memory signal for that old item. For example, it might instead reflect a semantic memory signal, one that would have occurred even if the word were new, or it might simply reflect an artifactual noise signal that happened to coincide with the presentation of the test item. However, both semantic memory signals and artifactual noise signals would be expected to occur with equal probability to old items (targets) and new items (foils). Therefore, the prediction of interest applies to a comparison of the full distributions of item-by-neuron spike counts for targets vs. foils.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo identify the presence of the sparse coding signal, the normalized target-by-neuron and foil-by-neuron response spike count distributions were compared both visually and statistically, and for each medial temporal lobe region separately (hippocampus: 73,539 item-by-neuron recordings, amygdala: 99,944 item-by-neuron recordings). To compare them visually, empirical quantile-quantile (QQ) plots were generated. The goal of this method is to visually assess whether two datasets come from the same or similar distributions. A linear pattern indicates that the two distributions have the same shape even if they have different means and standard deviations (i.e., even if they differ in the first two moments). Non-linear patterns reveal differences in skewness (the third moment) or tail behavior (i.e., kurtosis, the fourth moment) between the two distributions.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eStatistically, the moments associated with the two empirical distributions were compared using bootstrap analyses. The empirical data consisted of N\u003csub\u003et\u003c/sub\u003e target-by-neuron normalized spike counts and N\u003csub\u003ef\u003c/sub\u003e foil-by-neuron normalized spike counts. The sparse coding signal would be evident as an interaction effect between the two MTL regions of interest (hippocampus and amygdala) and the skewness of the two distributions (target and foils) in which the difference in skewness between the target and foil distributions is greater in the hippocampus relative to the corresponding difference in the amygdala. \u0026nbsp; First, for each statistical moment (i.e., mean, standard deviation, skewness, and kurtosis), the observed interaction value was calculated (e.g., the difference in skewness values between the target and foil distributions for the amygdala was subtracted from the corresponding difference for the hippocampus).\u003c/p\u003e\n\u003cp\u003eNext, we combined the N\u003csub\u003et\u003c/sub\u003e and N\u003csub\u003ef\u003c/sub\u003e empirical recordings to conduct the bootstrapping procedure. For each of B\u003cem\u003e\u0026nbsp;=\u0026nbsp;\u003c/em\u003e10,000 bootstrap trials, a bootstrapped “target” distribution of N\u003csub\u003et\u003c/sub\u003e observations was created by randomly sampling with replacement from the combined empirical data, and a bootstrapped “foil” distribution of N\u003csub\u003ef\u003c/sub\u003e observations was created by randomly sampling with replacement from the combined empirical data. We then calculated the difference in the relevant statistical measure (e.g., skewness) between these two randomly sampled distributions. Repeating this process for B\u003cem\u003e\u0026nbsp;=\u0026nbsp;\u003c/em\u003e10,000 bootstrap trials yielded 10,000 bootstrap difference scores for the relevant statistic. The proportion of bootstrapped trials for which the absolute value of the interaction difference score was greater than or equal to the absolute value of the empirically observed interaction difference score determined the \u003cem\u003eP\u003c/em\u003e value. Significance was defined as \u003cem\u003eP\u0026nbsp;\u003c/em\u003e\u0026lt; .05 (two-tailed), unless otherwise specified. The bootstrap procedure was repeated for the interaction difference score for each statistical moment separately. To determine the direction of the interaction effects, the same bootstrap procedure was used to examine if the \u0026nbsp;difference in each statistical moment between the target and foil distributions was statistically reliable within the hippocampus and the amygdala separately.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAfter performing this omnibus analysis, we investigated whether the sparse coding signature during retrieval (i.e., a significant interaction in skewness values between the target and foil distributions and between the hippocampus and amygdala) was selectively associated with two specific circumstances that should theoretically give rise to it (i.e., excitable neurons at encoding and subsequent memory). First, we investigated whether the sparse coding signature was selectively present at retrieval for targets associated with excitable neurons during encoding. Within each session, the mean pre-stimulus (1000ms-200ms) and mean post-stimulus (200ms-1000ms) raw spike counts during the encoding phase (i.e., during list presentation) were calculated for each item. Items were then rank ordered by the raw mean spike count value and placed into “Low” or “High” groups using a median split, first for pre-stimulus values and again for post-stimulus values. Items that were at the median value pre- or post-stimulus were excluded from the primary analysis. This resulted in target items being categorized into one of four categories based on the level of firing pre- and post-stimulus presentation: 1) increase in firing: Low-High, 2) \u0026nbsp;consistently high firing: High-High, 3) decrease in firing: High-Low, and 4) consistently low firing: Low-Low. Items most strongly associated with heightened excitability were defined to be those in the Low-High category, and the question of interest was whether a difference in skewness between the target-by-neuron response distribution and foil-by-neuron response distribution would be limited to those items.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eSecond, we investigated whether the sparse coding signature was selectively associated with items that were successfully remembered (i.e., correct “old” decisions on the recognition test). To address this issue, after classifying targets into one of four categories based on their responsivity during encoding (e.g., Low-Low), the targets were further classified within each category as remembered or forgotten. The question of interest was whether a difference in skewness between the target-by-neuron response distribution and foil-by-neuron response distribution would be limited not only to targets most associated with excitable neurons (Low-High) but also to a further subset of those targets that were successfully retrieved during the recognition test.\u0026nbsp;\u003c/p\u003e"},{"header":"Results","content":"\u003ch2\u003e\u003cem\u003eBehavioral Performance\u003c/em\u003e\u003c/h2\u003e\n\u003cp\u003eOn average, participants performed above chance on the recognition memory test for old items, 67.7% correct, SEM = 0.2% (i.e., hit rate = .677) and for foil items, 73.9% correct, SEM = 0.2% (i.e., false alarm rate = 1 - .739 = .261). The average \u003cem\u003edʹ\u0026nbsp;\u003c/em\u003eacross sessions was 1.24 (min = .05, max = 2.47). Four sessions with negative \u003cem\u003edʹ\u003c/em\u003e values were excluded.\u003c/p\u003e\n\u003ch2\u003e\u003cem\u003eThe item-specific episodic memory signal was selectively detected in the hippocampus\u003c/em\u003e\u003c/h2\u003e\n\u003cp\u003eA sparse episodic memory signal consists of a small fraction of neurons selectively responding to a small fraction of \u0026ldquo;old\u0026rdquo; items (items studied before the recognition test). In the simplest case, a single neuron would exhibit a strong response to a single old item but not to any other old item nor to any new item. In other words, the neuron\u0026rsquo;s response would be (1) item specific, and (2) specific to that item only if it is old. We investigated this issue by comparing the full distributions of item-by-neuron spike counts for targets vs. foils.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eConsider a simplified concrete example. If recordings were made from 5 single neurons during the presentation of 20 items on the recognition test (10 old items and 10 new items), the two distributions of interest would consist of the 50 recordings to old items (5 neurons X 10 old items) and the 50 recordings to new items (5 neurons X 10 new items). The two distributions\u0026mdash;which we refer to as the target-by-neuron response distribution and the foil-by-neuron response distribution\u0026mdash;should be a similar shape when compared to one another. For example, a singular neuron would respond strongly to only one target item, but not to any of the other target items, nor to any of the foil items. Thus, the distributions of the target-by-neuron recordings and the foil-by-neuron recordings may be a similar shape, with the only difference being a very small percentage of the target-by-neuron recordings showing a differentially strong response. Statistically, the prediction is that the target-by-neuron response distribution should be significantly more skewed to the right compared to the foil-by-neuron response distribution. Moreover, this difference in skewness should reflect the influence of a very small percentage of target-by-neuron responses. Therefore, the difference in skewness should disappear after removing a small percentage of the highest values from both distributions.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo visualize the shapes of these distributions, we used empirical QQ plots (Chambers et al., 1983, see Methods). In this type of plot, the quantiles of one dataset are plotted on one axis, and the corresponding quantiles of the second dataset are plotted on the other axis. Here, we compared the quantiles of the foil-by-neuron normalized spike count distribution (x-axis) vs. the quantiles of the target-by-neuron normalized spike count distribution (y-axis), separately for the hippocampus and amygdala (Figure 2a \u0026amp; 2b). Deviations from linearity arise when the distributions differ in moments beyond the first (mean) and second (standard deviation). Of most interest here, an upward deflection away from the diagonal at the upper right of the QQ plot indicates a difference in the third moment (skewness). More specifically, it indicates that the target item distribution is more skewed to the right than the foil item distribution (a feature that might not be visually apparent in a plot of the frequency distributions). Note that points on a QQ plot are densely grouped and represent thousands of recordings that have similar values for both distributions, so there are many more points on the plot than it appears.\u003c/p\u003e\n\u003cp\u003eFor the hippocampus, an upward deflection is evident at higher x and y values, indicating that the target (i.e., old item) distribution is more positively skewed compared to the foil (i.e., new item) distribution. To investigate whether the apparent difference in skewness is due to a small percentage of observations in the target-by-neuron response distribution, we removed the top 0.25% of the highest-ranking recordings from both distributions. In a plot of the remaining 99.75% of recordings, the deflection was no longer apparent (Figure 2c \u0026amp; 2d), thus providing visual evidence of sparse coding in hippocampal neurons. For the amygdala, a less pronounced downward deflection was observed, indicating differentially higher spiking for a small proportion of neurons in the foil distribution. Similarly, after we removed the top 0.25% of the highest-ranking recordings from both distributions, the deflection towards the x-axis was no longer apparent.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eWe next examined whether the visual deflections were statistically significant, focusing mainly on the skewness measures of the two distributions (Figure 3; Table 1). First, and most importantly, we tested for an interaction between the two MTL regions of interest (hippocampus and amygdala) and the skewness of the two distributions (target and foils) using bootstrap tests (see Methods). Theoretically, the sparse coding signal would be evident as an interaction effect such that the magnitude of the target\u0026ndash;minus-foil difference in skewness would be larger in the hippocampus than in the amygdala. A one-tailed test was used based on the specific directional hypotheses predicted by Urgolites et al. (2022). The interaction was significant (\u003cem\u003eP =\u0026nbsp;\u003c/em\u003e.049), replicating the significant interaction reported by Urgolites et al. (2022).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1. \u0026nbsp;Statistical Results for the Statistical Moments of the Target vs. Foil Distributions at Retrieval\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"https://myfiles.space/user_files/122228_c8a1650c59388082/122228_custom_files/img1745393103.png\"\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eNote.\u0026nbsp;\u003c/em\u003eA significant interaction was defined as a greater difference in a given statistical moment (e.g., skewness) between the target-vs.-foil normalized spike count distributions in the hippocampus compared to the corresponding difference in the amygdala according to bootstrap tests (B = 10,000, \u003cem\u003eP\u003c/em\u003e \u0026lt; .05, one-tailed). The interactions for mean (\u003cem\u003eP\u003c/em\u003e = .047) and skewness (\u003cem\u003eP\u003c/em\u003e = .049) were significant. Differences in each statistical moment between the two distributions (target-vs.-foil) were also considered separately within each structure (hippocampus and amygdala) using bootstrap tests (B = 10,000, \u003cem\u003eP\u003c/em\u003e \u0026lt; .05, one-tailed). In the hippocampus, differences in the upper moments of the target and foil spike count distributions were not significant but trended in that direction (standard deviation: \u003cem\u003eP\u003c/em\u003e = .054, skewness: \u003cem\u003eP\u003c/em\u003e = .070, kurtosis: \u003cem\u003eP\u003c/em\u003e = .086). In the amygdala, the differences were not significant (standard deviation: \u003cem\u003eP\u003c/em\u003e = .344, skewness: \u003cem\u003eP\u003c/em\u003e = .254, kurtosis: \u003cem\u003eP =\u0026nbsp;\u003c/em\u003e.178). Finally, mean firing in the amygdala distinguished between target and foil items (\u003cem\u003eP\u003c/em\u003e = .019) but not in the hippocampus (\u003cem\u003eP\u003c/em\u003e = .334). ~ = \u003cem\u003eP\u003c/em\u003e \u0026lt; 0.10, * = \u003cem\u003eP\u003c/em\u003e \u0026lt; .05.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;Next, one-tailed bootstrap tests were performed to assess the directionality of the effects (Table 1). In the hippocampus, the \u0026nbsp; target-by-neuron response distribution was more positively skewed (4.17) than the foil-by-neuron response distribution (3.05), as also reported by Urgolites (2022), but the difference here was not quite significant (\u003cem\u003eP\u003c/em\u003e = .070). Conversely, in the amygdala, the skewness of the target-by-neuron response distribution (3.0) was slightly less than the skewness of the foil-by-neuron response distribution (3.28), but the difference did not approach significance (\u003cem\u003eP\u003c/em\u003e = .254). In the amygdala, the only significant difference was in the means of the distributions, with foil items eliciting significantly more firing on average than targets (targets = .14, foils = .16; \u003cem\u003eP\u003c/em\u003e = .019).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe theoretically important skewness interaction reported by (Urgolites et al., 2022) \u0026nbsp;was evident in this independent dataset as well, and it is consistent with an item-specific sparse coding signal being selectively observed in the hippocampus. Moreover, as described in more detail next, the interaction became more visually apparent and more statistically reliable when examining subsets of targets that, theoretically, should exhibit the same effects.\u003c/p\u003e\n\u003ch2\u003e\u003cem\u003eThe item-specific episodic memory signal was related to excitability at encoding\u003c/em\u003e\u003c/h2\u003e\n\u003cp\u003eNext, we tested a theory of neuronal allocation according to which episodic memories are assigned to neurons with heightened excitability at learning. We operationalized a test of this theory by repeating the same sparse coding analysis just described (visual tests using QQ plots, statistical tests based on skewness), but this time by partitioning retrieval data for targets\u0026nbsp;by the pattern of spiking\u0026nbsp;observed before and during item presentation at encoding.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eWithin session, each target was assigned to one of four categories determined by their relative firing pre- and post- stimulus presentation at encoding (see Methods, Low-Low, Low-High, High-High, High-Low). Based on the theoretical predictions of sparse coding and neuronal allocation, we expected the item-specific episodic memory signal would remain selective to the hippocampus and would further be selective to the subset of targets which were associated with a relative increase in excitability at encoding (low firing pre-stimulus and high firing post-stimulus item presentation at learning; Low-High). Additionally, we expected that the episodic memory signal would remain undetected in the amygdala (see Figures 2 \u0026amp; 3 and Table 1), regardless of spiking observed before and during item presentation at encoding.\u003c/p\u003e\n\u003cp\u003eWe repeated the analyses used to identify the sparse coding signal (see Methods and Section 1) based on the skewness of the target-by-neuron distributions and the foil-by-neuron distributions, separately for each category of relative spiking at encoding, beginning with a visual inspection of the relevant QQ plots (Figure 4). A sharp upward deflection of the QQ plot (indicative of greater skewness for the target distribution relative to the foil distribution) was present only for the targets that were categorized as Low-High at encoding (i.e., the excitable neurons) and was further selective to hippocampal neurons (Figure 4a). For the amygdala, by contrast, almost all points were densely grouped along the red diagonal line (Figure 4b), providing no evidence of a skewness difference for targets with heightened excitability at encoding. The upward deflection was not visually apparent on any of the remaining QQ plots, either in the hippocampus or the amygdala, for the subsets of targets in which spiking remained high (High-High: Figure 4c and 4d), spiking decreased (High-Low: Figure 4e and 4f), or remained low (Low-Low: Figure 4g and 4h) (Supplemental Figure 1).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eWe next tested the statistical significance of the visual trends in the QQ plots using bootstrapping tests (see Methods; B = 10,000, Bonferroni corrected: P \u0026lt; 0.0125, two-tailed). With regard to the measure of most interest, skewness, a significant interaction was observed between the MTL regions (hippocampus vs amygdala) and trial type (Table 2: targets labeled Low-High at encoding vs foil items; \u0026nbsp;\u003cem\u003eP\u003c/em\u003e = 0.008). For the hippocampus, the Low-High target-by-neuron response distribution was significantly more skewed than the foil-by-neuron response distribution (6.79 vs. 3.06, \u003cem\u003eP\u003c/em\u003e = .004). In contrast, in the amygdala, the skewness of the corresponding distributions did not differ significantly (3.27 vs. 3.28, \u003cem\u003eP\u003c/em\u003e = .989). No significant interactions for the difference in skewness values between targets and foils and the MTL regions were present for the other subsets of targets in which spiking decreased (High-Low), remained high (High-High), or remained low (Low-Low). Additionally, interaction tests performed on the other statistical moments (mean, standard deviation, and kurtosis) were not significant for any of the subsets of targets, including the targets with heightened excitability at encoding (Low-High). (Supplemental Table 1).\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eTable 2. Statistical Results for Skewness of the Target vs. Foil Distributions at Retrieval Partitioned by the Pattern of Target Firing at Encoding\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"https://myfiles.space/user_files/122228_c8a1650c59388082/122228_custom_files/img1745393321.png\"\u003e\u003cbr\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eNote.\u003c/em\u003e A significant interaction was defined as a greater difference in a given statistical moment \u0026nbsp;between the target-vs.-foil normalized spike count distributions in one region (e.g., the hippocampus) compared to the corresponding difference the other region (e.g., the amygdala) according to bootstrap tests (B = 10,000, Bonferroni corrected\u003cem\u003e, P\u003c/em\u003e \u0026lt; .0125). The interaction was significant for the test involving targets associated with excitable neurons at encoding (Low-High: \u003cem\u003eP\u003c/em\u003e =.008) but not for any other category of targets at encoding (High-High: \u003cem\u003eP =\u0026nbsp;\u003c/em\u003e.803, High-Low: \u003cem\u003eP =\u0026nbsp;\u003c/em\u003e.891, or Low-Low: \u003cem\u003eP =\u0026nbsp;\u003c/em\u003e.473). Within the hippocampus, a significant difference between target vs. foil skewness (theoretically indicative of the item-specific memory signal) was observed only for targets that were associated with excitability at encoding (Low-High: \u003cem\u003eP =\u0026nbsp;\u003c/em\u003e.004). Within the amygdala, no significant difference in skewness between the target and foil distributions was observed. Corresponding analyses for the other statistical moments are reported in Supplemental Table 1. ** = \u003cem\u003eP\u003c/em\u003e \u0026lt; 0.01\u003c/p\u003e\n\u003ch2\u003e\u003cem\u003eSubsequent memory was associated with the item-specific episodic memory signal in the hippocampus\u003c/em\u003e\u003c/h2\u003e\n\u003cp\u003eNext, QQ plots were generated separately to compare the distributions of the normalized spike counts of remembered targets to foil items (hippocampus: Figure 5a \u0026amp; amygdala: Figure 5b, hippocampus) and forgotten targets to foil items (hippocampus: Figure 5c \u0026amp; amygdala: Figure 5d) In the hippocampus, greater skewness of the target-by-neuron distribution is apparent in the QQ plots when the targets were subsequently remembered (Figure 5a) but not when they were forgotten (Figure 5c). Indeed, when the targets were forgotten, the QQ plot trend is in the opposite direction. No apparent nonlinear trends were evident in the QQ plots for the amygdala whether the targets were remembered or forgotten (Figure 5b and 5d). The non-linear deflection in the hippocampal QQ plot (Figure 5a) was no longer apparent after removing a small fraction of the target-by-neuron and foil-by-neuron recordings (Supplemental Figure 2), indicating relatively few high firing neurons contributed to the greater skewness of the remembered target-by-neuron distribution when compared to foil items.\u003c/p\u003e\n\u003cp\u003eStatistically, \u0026nbsp;bootstrapping tests determined the interaction between the skewness of the two distributions (remembered targets vs. foils) and brain regions (hippocampus vs. amygdala) was significant (Table 3, B = 10,000, \u003cem\u003eP\u003c/em\u003e = .027, two-tailed). Considering the hippocampus only, skewness was significantly greater for the remembered target distribution compared to the foil item distribution (4.82 vs. 3.05, \u003cem\u003eP\u003c/em\u003e = .036). The interaction test for a difference in kurtosis of the normalized spike count distributions between MTL regions (remembered targets vs. foil items) was significant (\u003cem\u003eP\u003c/em\u003e = .037) and marginally significant for standard deviation (\u003cem\u003eP\u003c/em\u003e = .083). Within each region considered separately, a significant difference in the remembered vs. foil item distributions was also selectively observed in the hippocampus for standard deviation (1.26 vs. 1.18, \u003cem\u003eP\u003c/em\u003e = .015) and a marginally significant difference was present for kurtosis (72.3 vs. 29.9, \u003cem\u003eP\u003c/em\u003e = .064). Similar effects in the hippocampus were not observed for forgotten targets, and no such effects were observed in the amygdala for either remembered (Table 3) or forgotten targets (Table 4).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 3. Analyses of the Statistical Moments of Remembered Target vs. Foil Distributions at Retrieval\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"https://myfiles.space/user_files/122228_c8a1650c59388082/122228_custom_files/img1745393380.png\"\u003e\u003cbr\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eNote.\u0026nbsp;\u003c/em\u003eA significant interaction was defined as a greater difference in a given statistical moment between the target-vs.-foil normalized spike count distributions in one region (e.g., the hippocampus) compared to the corresponding difference the other region (e.g., the amygdala) according to bootstrap tests (B = 10,000\u003cem\u003e, P\u003c/em\u003e \u0026lt; .05). The interaction between the two distributions (remembered targets vs. foils) and brain region (hippocampus vs. amygdala) was significant for skewness (\u003cem\u003eP\u003c/em\u003e = .027) and kurtosis (\u003cem\u003eP\u003c/em\u003e = .037). Within each region considered separately, significant differences were observed for standard deviation (\u003cem\u003eP\u003c/em\u003e = .015) and skewness (\u003cem\u003eP\u003c/em\u003e = .036) and was marginally significant for kurtosis (\u003cem\u003eP\u003c/em\u003e = .064). These effects were not present in the amygdala. ~ = \u003cem\u003eP\u003c/em\u003e \u0026lt; .10, * = \u003cem\u003eP\u003c/em\u003e \u0026lt; .05.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 4. Analyses of the Statistical Moments of Forgotten Target vs. Foil Distributions at Retrieval\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cimg src=\"https://myfiles.space/user_files/122228_c8a1650c59388082/122228_custom_files/img1745393414.png\"\u003e\u003c/strong\u003e\u003cbr\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eNote.\u0026nbsp;\u003c/em\u003eA significant interaction was defined as a greater difference in a given statistical moment between the target-vs.-foil normalized spike count distributions in one region (e.g., the hippocampus) compared to the corresponding difference the other region (e.g., the amygdala) according to bootstrap tests (B = 10,000\u003cem\u003e, P\u003c/em\u003e \u0026lt; .05). The interaction between a given statistical moment of the two distributions (forgotten targets vs. foils) and brain region (hippocampus vs. amygdala) was significant for the mean only (\u003cem\u003eP\u003c/em\u003e \u0026lt; .001). Within each region considered separately, \u0026nbsp;the difference between the mean of the forgotten vs. foils item distribution was significant in the amygdala (\u003cem\u003eP\u003c/em\u003e \u0026lt; .001) but not the hippocampus (\u003cem\u003eP\u003c/em\u003e = .470). ~ = \u003cem\u003eP\u003c/em\u003e \u0026lt; 0.10, *** = \u003cem\u003eP\u003c/em\u003e \u0026lt; .001.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;We next performed a direct comparison between the remembered and forgotten targets. The analyses just described compared the target-by-neuron normalized spike count distribution (y-axis of the QQ plot) and the foil-by-neuron normalized spike count distribution (x-axis of the QQ plot), separately for remembered targets and forgotten targets. In the next analysis, we examined the QQ plot with the target-by-neuron distribution for remembered targets plotted on the y-axis and the target-by-neuron distribution for forgotten targets on the x-axis (Figure 6). In other words, unlike the analyses described thus far, recordings to foil items were not included in this analysis. In the hippocampus, a difference in skewness was visibly apparent (Figure 6a). Conversely for amygdala recordings, almost all points fell on the diagonal line, and no deflection was observed in the QQ plot (Figure 6b). The non-linear deflection in the hippocampal QQ plot (Figure 6a) was no longer apparent after removing a small fraction of the target-by-neuron and foil-by-neuron recordings (Supplemental Figure 3), indicating relatively few high firing neurons contributed to the greater skewness of the remembered target-by-neuron distribution when compared to forgotten items.\u003c/p\u003e\n\u003cp\u003eA significant interaction was observed for skewness values between brain region (hippocampus vs. amygdala) compared to the status of the test item (Table 5; remembered targets compared to forgotten targets, \u003cem\u003eP\u003c/em\u003e = .034). Within the hippocampus, skewness was significantly greater for the remembered target distribution compared to the forgotten item distribution (4.82 vs. 2.30, \u003cem\u003eP\u003c/em\u003e = .022), but the corresponding comparison in the amygdala was not significant (2.94 vs. 3.15, \u003cem\u003eP\u003c/em\u003e = .732).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAlthough not relevant to the theoretical issues of interest here, it is worth noting that a significant interaction was observed between the MTL regions for the difference in the means for remembered compared to forgotten targets (Table 5, \u003cem\u003eP\u003c/em\u003e = .0015). Within the amygdala, mean firing was significantly greater for the remembered item distribution (.16) compared to the forgotten item distribution (.09), but the corresponding difference in the hippocampus was not significant. \u0026nbsp;Note that this effect in the amygdala is the generic memory signal that distinguished between general categories of stimuli (remembered versus forgotten), rather than representing a single episodic memory.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 5. Analyses of the Statistical Moments of Remembered Target vs. Forgotten Target Distributions at Retrieval\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cimg src=\"https://myfiles.space/user_files/122228_c8a1650c59388082/122228_custom_files/img1745393537.png\"\u003e\u003c/strong\u003e\u003cbr\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eNote.\u0026nbsp;\u003c/em\u003eA significant interaction reflects a greater difference in a statistical moment (e.g., skewness) of the spike count distributions for the remembered vs. forgotten targets in one region compared to the other region (hippocampus vs. amygdala) according to bootstrap tests (B\u003cem\u003e\u0026nbsp;=\u003c/em\u003e10,000, \u003cem\u003eP\u003c/em\u003e \u0026lt; 0.05). The interactions were significant for the mean (\u003cem\u003eP\u003c/em\u003e = .0015) and skewness ( P = .034), and marginally significant for kurtosis (\u003cem\u003eP\u003c/em\u003e = .059). Within each region considered separately, differences in the upper moments of the distributions were detected in the hippocampus (skewness: \u003cem\u003eP\u003c/em\u003e = .022, kurtosis: \u003cem\u003eP\u003c/em\u003e = .0501) but not the amygdala (skewness: \u003cem\u003eP\u003c/em\u003e = .732 and kurtosis: \u003cem\u003eP\u003c/em\u003e = .789). Conversely, mean firing was greater for remembered compared to forgotten items in the amygdala (\u003cem\u003eP\u003c/em\u003e \u0026lt; .001) but not in the hippocampus (\u003cem\u003eP\u003c/em\u003e = .156). Although the interaction was not significant for standard deviation, within each region, both the hippocampus exhibited a significant difference (\u003cem\u003eP\u003c/em\u003e = .020) \u0026nbsp;and the amygdala exhibited a marginally significant difference (\u003cem\u003eP\u003c/em\u003e = .055), with greater SD values for remembered compared to forgotten items. * = P \u0026lt; 0.05, ** = \u003cem\u003eP\u003c/em\u003e \u0026lt; 0.01, *** = \u003cem\u003eP\u003c/em\u003e \u0026lt; 0.001.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eThe item-specific episodic memory signal was selectively detected for remembered but not forgotten targets associated with excitability at encoding\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThus far, we have identified two theoretically motivated and distinct subsets of targets for which the target-by-neuron response distribution exhibits greater skewness than the foil-by-neuron response distribution: 1) targets with heightened excitability at encoding and 2) targets that were subsequently remembered. Thus, we posited that the item-specific episodic memory signal in the hippocampus (i.e., elevated skewness resulting from a small percentage of recordings) would be selectively identified for targets that were \u003cem\u003eboth\u003c/em\u003e excitable at encoding and subsequently remembered.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIndeed, as might be expected, the skewness difference (theoretically reflective of sparse coding) was visually identified only in the QQ plots that were selective to the hippocampus and involving targets that were (1) associated with increased excitability during encoding and (2) were also subsequently remembered (Figure 7a). If the targets that were associated with increased excitability during encoding were subsequently forgotten, no such effect was evident (Figure 7c). This pattern also did not emerge in the amygdala for targets that exhibited increased excitability during encoding, whether those targets were later remembered (Figure 7b) or forgotten (Figure 7d). Additionally, no other encoding level-by-subsequent memory combinations within the hippocampus or amygdala (Supplemental Figure 5) exhibited the skewness difference that we hypothesize arises from sparse coding.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe remarkable specificity of these findings is best appreciated by examining a bar graph (Figure 8) showing the four statistical moments of the target-by-neuron response distribution (mean, standard deviation, skewness, and kurtosis) broken down by brain region (amygdala and hippocampus) and excitability at encoding. \u0026nbsp;The four statistical moments are shown separately in the four plots arrayed from left to right, with remembered targets shown in the top row of four graphs (Figure 8a-d) and forgotten targets shown in the bottom row of four graphs (Figure 8e-h). Excitability at encoding is color-coded within each of the eight graphs, with High-High in red, High-Low in orange, Low-Low in green, and Low-High in blue. The moments for the foil-by-neuron response distribution are shown in gray. For the foil items, the value for a given moment (e.g., the mean) is the same for the remembered-target graphs on top and the forgotten-target graphs on the bottom (i.e., the remembered vs. forgotten distinction does not apply to foil items). Note in particular, the measures shown in column C (skewness), where the blue bar in the top right graph of that column (neural recordings from the hippocampus, involving excitable targets that were also subsequently remembered) conspicuously stands out among all other bars. Essentially the same pattern is also evident for kurtosis in column D and is also somewhat apparent for the standard deviation in column B.\u003c/p\u003e\n\u003cp\u003eIn summary, the most specific and theoretically relevant subset of targets, items that were later remembered and were also associated with heightened excitability at encoding, were the only targets with evidence of sparse coding. Moreover, this effect was evident only in the hippocampus.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eNeurocomputational models hold that relatively few neurons represent distinct episodic memories, even similar ones, in distinct, pattern-separated sparse neural assemblies (Marr, 1971; McClelland et al., 1995; Norman \u0026amp; O\u0026rsquo;Reilly, 2003; Treves \u0026amp; Rolls, 1994). The formation of such representations should avoid catastrophic forgetting and facilitate the later retrieval of events that occurred in a particular spatial and temporal context (i.e., it should facilitate the retrieval of episodic memories). In addition, more recent theories of neural allocation hold that, during encoding, excitable neurons are differentially allocated to sparsely represent episodic memories (Silva et al., 2009), an idea that has been supported by work with rodents (Kim et al., 2013; Silva et al., 1998). To test these mechanisms of sparse coding and neuronal allocation in humans, we analyzed an open-source dataset consisting of single-unit recordings from the MTL that were made while epilepsy patients were engaged in an old/new recognition task for a list of images (Chandravadia et al., 2020; Faraut et al., 2018).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eOur findings replicated previously reported evidence of an item-specific (pattern-separated) sparse code in humans that was also specific to the hippocampus (Urgolites et al., 2022; Wixted, 2014, 2018). Additionally, here, we found that the most specific and theoretically relevant subset of targets\u0026mdash;those that were associated with heightened excitability during encoding and were also subsequently remembered\u0026mdash;yielded the clearest evidence of a sparse, pattern-separated code. Critically, the presence of this sparse code was only evident in the hippocampus. Together, the results provide clear evidence that excitable neurons are preferentially allocated to create a sparse, pattern-separated episodic memory code in the hippocampus that, in turn, supports subsequent retrieval.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAn alternative view accepts the possibility that episodic memories may be represented as a pattern-separated code in rats and monkeys but rejects the idea that the same is true of humans (Quiroga, 2020). According to this perspective, single unit recording studies from the human MTL have thus far yielded no evidence in support of predictions made by longstanding neurocomputational models. The findings reported here serve as a rebuttal to this claim. Moreover, some of the work that gave rise to the idea that newly formed episodic memories are encoded via overlapping neural assemblies (Ison et al., 2015) do not actually support that perspective (see supplemental material). Related work has identified \u0026ldquo;episode-specific\u0026rdquo; neurons in the human hippocampus, which are neurons that encode the combination of elements within an individual episode (Kolibius et al., 2023).\u003c/p\u003e\n\u003cp\u003eThe episodic memory signal of interest here was item-specific, not generic. As noted earlier, a generic memory signal\u0026mdash;whereby a neuron fires differentially to the category of old items vs. new items\u0026mdash;has often been detected in the hippocampus , the amygdala, and multiple regions of the cortex (Rutishauser et al., 2008, 2010, 2015; Urgolites et al., 2022). This research has provided valuable insights, but our emphasis here is on the item-specific memory signal in the hippocampus that neurocomputational models have long proposed as the foundation of episodic memory. In accordance with those models and with later theoretical developments, our findings provide the first clear evidence that the human hippocampus selectively represents single episodic memories and facilitates remembering using a sparse pattern-separated code, which is preferentially allocated to excitable neurons during encoding.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAuthor contributions:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eC.W.T contributed data analysis, data interpretation, and manuscript writing. P.N.S. contributed data curation and statistical guidance at all stages of the analysis and manuscript writing. J.T.W. supervised the project and contributed to all aspects of the manuscript. All of the authors discussed the results at all stages of the project.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003cstrong\u003eCompeting interests:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eC.W.T., P.N.S., and J.T.W. have no competing interests.\u0026nbsp;\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eData Availability Statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe dataset analyzed during the current study is available in the following Open Science Framework repository: https://osf.io/hv7ja/ (Chandravadia et al., 2020)\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eCai, D. J., Aharoni, D., Shuman, T., Shobe, J., Biane, J., Song, W., Wei, B., Veshkini, M., La-Vu, M., Lou, J., Flores, S. E., Kim, I., Sano, Y., Zhou, M., Baumgaertel, K., Lavi, A., Kamata, M., Tuszynski, M., Mayford, M., \u0026hellip; Silva, A. J. (2016). A shared neural ensemble links distinct contextual memories encoded close in time. \u003cem\u003eNature\u003c/em\u003e, \u003cem\u003e534\u003c/em\u003e(7605), 115\u0026ndash;118. https://doi.org/10.1038/nature17955\u003c/li\u003e\n\u003cli\u003eChambers, J. M., Cleveland, W. S., Kleiner, B., \u0026amp; Tukey, P. (1983). \u003cem\u003eGraphical Methods of Data Analysis\u003c/em\u003e. Duxbury Press.\u003c/li\u003e\n\u003cli\u003eChandravadia, N., Liang, D., Schjetnan, A. G. P., Carlson, A., Faraut, M., Chung, J. M., Reed, C. M., Dichter, B., Maoz, U., Kalia, S. K., Valiante, T. A., Mamelak, A. N., \u0026amp; Rutishauser, U. (2020). A NWB-based dataset and processing pipeline of human single-neuron activity during a declarative memory task. \u003cem\u003eScientific Data\u003c/em\u003e, \u003cem\u003e7\u003c/em\u003e(1), 78. https://doi.org/10.1038/s41597-020-0415-9\u003c/li\u003e\n\u003cli\u003eFaraut, M. C. M., Carlson, A. A., Sullivan, S., Tudusciuc, O., Ross, I., Reed, C. M., Chung, J. M., Mamelak, A. N., \u0026amp; Rutishauser, U. (2018). Dataset of human medial temporal lobe single neuron activity during declarative memory encoding and recognition. \u003cem\u003eScientific Data\u003c/em\u003e, \u003cem\u003e5\u003c/em\u003e(1), 180010. https://doi.org/10.1038/sdata.2018.10\u003c/li\u003e\n\u003cli\u003eFrench, R. M. (1999). Catastrophic forgetting in connectionist networks. \u003cem\u003eTrends in Cognitive Sciences\u003c/em\u003e, \u003cem\u003e3\u003c/em\u003e(4), 128\u0026ndash;135. https://doi.org/10.1016/S1364-6613(99)01294-2\u003c/li\u003e\n\u003cli\u003eIson, M. J., Quiroga, R. Q., \u0026amp; Fried, I. (2015). Rapid encoding of new memories by individual neurons in the human brain. \u003cem\u003eNeuron\u003c/em\u003e, \u003cem\u003e87\u003c/em\u003e, 220\u0026ndash;230.\u003c/li\u003e\n\u003cli\u003eKim, J., Kwon, J.-T., Kim, H.-S., \u0026amp; Han, J.-H. (2013). CREB and neuronal selection for memory trace. \u003cem\u003eFrontiers in Neural Circuits\u003c/em\u003e, \u003cem\u003e7\u003c/em\u003e. https://doi.org/10.3389/fncir.2013.00044\u003c/li\u003e\n\u003cli\u003eKolibius, L. D., Roux, F., Parish, G., Ter Wal, M., Van Der Plas, M., Chelvarajah, R., Sawlani, V., Rollings, D. T., Lang, J. D., Gollwitzer, S., Walther, K., Hopfeng\u0026auml;rtner, R., Kreiselmeyer, G., Hamer, H., Staresina, B. P., Wimber, M., Bowman, H., \u0026amp; Hanslmayr, S. (2023). Hippocampal neurons code individual episodic memories in humans. \u003cem\u003eNature Human Behaviour\u003c/em\u003e, \u003cem\u003e7\u003c/em\u003e(11), 1968\u0026ndash;1979. https://doi.org/10.1038/s41562-023-01706-6\u003c/li\u003e\n\u003cli\u003eMarr, D. (1971). Simple memory: A theory for archicortex. Philos (R. S. Lond, Trans.). \u003cem\u003eB Biol. Sci\u003c/em\u003e, \u003cem\u003e262\u003c/em\u003e, 23\u0026ndash;81.\u003c/li\u003e\n\u003cli\u003eMcClelland, J., McNaughton, B., \u0026amp; O\u0026rsquo;Reilly, R. (1995). Why there are complementary learning systems in the hippocampus and neocortex: Insights from the successes and failures of connectionist models of learning and memory. \u003cem\u003ePSYCHOLOGICAL REVIEW\u003c/em\u003e, \u003cem\u003e102\u003c/em\u003e(3), 419\u0026ndash;457. https://doi.org/10.1037/0033-295X.102.3.419\u003c/li\u003e\n\u003cli\u003eNorman, K. A., \u0026amp; O\u0026rsquo;Reilly, R. C. (2003). Modeling hippocampal and neocortical contributions to recognition memory: A complementary-learning-systems approach. \u003cem\u003ePsychological Review\u003c/em\u003e, \u003cem\u003e110\u003c/em\u003e(4), 611\u0026ndash;646. https://doi.org/10.1037/0033-295X.110.4.611\u003c/li\u003e\n\u003cli\u003eQuiroga, R. Q. (2020). No pattern separation in the human hippocampus. \u003cem\u003eTrends Cogn. Sci\u003c/em\u003e, \u003cem\u003e24\u003c/em\u003e, 994\u0026ndash;1007.\u003c/li\u003e\n\u003cli\u003eRutishauser, U., Ross, I. B., Mamelak, A. N., \u0026amp; Schuman, E. M. (2010). Human memory strength is predicted by theta-frequency phase-locking of single neurons. \u003cem\u003eNature\u003c/em\u003e, \u003cem\u003e464\u003c/em\u003e(7290), 903\u0026ndash;907. https://doi.org/10.1038/nature08860\u003c/li\u003e\n\u003cli\u003eRutishauser, U., Schuman, E. M., \u0026amp; Mamelak, A. N. (2008). Activity of human hippocampal and amygdala neurons during retrieval of declarative memories. \u003cem\u003eProceedings of the National Academy of Sciences\u003c/em\u003e, \u003cem\u003e105\u003c/em\u003e(1), 329\u0026ndash;334. https://doi.org/10.1073/pnas.0706015105\u003c/li\u003e\n\u003cli\u003eRutishauser, U., Ye, S., Koroma, M., Tudusciuc, O., Ross, I. B., Chung, J. M., \u0026amp; Mamelak, A. N. (2015). Representation of retrieval confidence by single neurons in the human medial temporal lobe. \u003cem\u003eNature Neuroscience\u003c/em\u003e, \u003cem\u003e18\u003c/em\u003e(7), 1041\u0026ndash;1050. https://doi.org/10.1038/nn.4041\u003c/li\u003e\n\u003cli\u003eSilva, A. J., Kogan, J. H., Frankland, P. W., \u0026amp; Kida, S. (1998). CREB AND MEMORY. \u003cem\u003eAnnual Review of Neuroscience\u003c/em\u003e, \u003cem\u003e21\u003c/em\u003e(Volume 21, 1998), 127\u0026ndash;148. https://doi.org/10.1146/annurev.neuro.21.1.127\u003c/li\u003e\n\u003cli\u003eSilva, A. J., Zhou, Y., Rogerson, T., Shobe, J., \u0026amp; Balaji, J. (2009). Molecular and Cellular Approaches to Memory Allocation in Neural Circuits. \u003cem\u003eScience\u003c/em\u003e, \u003cem\u003e326\u003c/em\u003e(5951), 391\u0026ndash;395. https://doi.org/10.1126/science.1174519\u003c/li\u003e\n\u003cli\u003eSquire, L. R. (2004). Memory systems of the brain: A brief history and current perspective. \u003cem\u003eNeurobiology of Learning and Memory\u003c/em\u003e, \u003cem\u003e82\u003c/em\u003e(3), 171\u0026ndash;177. https://doi.org/10.1016/j.nlm.2004.06.005\u003c/li\u003e\n\u003cli\u003eTreves, A., \u0026amp; Rolls, E. T. (1994). Computational analysis of the role of the hippocampus in memory. \u003cem\u003eHippocampus\u003c/em\u003e, \u003cem\u003e4\u003c/em\u003e, 374\u0026ndash;391.\u003c/li\u003e\n\u003cli\u003eTulving, E. (1972). Episodic and semantic memory. In E. Tulving \u0026amp; W. Donaldson (Eds.), \u003cem\u003eOrganization of memory\u003c/em\u003e. Academic Press.\u003c/li\u003e\n\u003cli\u003eUrgolites, Z. J., Wixted, J. T., Goldinger, S. D., Papesh, M. H., Treiman, D. M., Squire, L. R., \u0026amp; Steinmetz, P. N. (2022). Two kinds of memory signals in neurons of the human hippocampus. \u003cem\u003eProceedings of the National Academy of Sciences\u003c/em\u003e, \u003cem\u003e119\u003c/em\u003e(19), e2115128119. https://doi.org/10.1073/pnas.2115128119\u003c/li\u003e\n\u003cli\u003eWillmore, B., \u0026amp; Tolhurst, D. J. (2001). Characterizing the sparseness of neural codes. \u003cem\u003eNetwork: Computation in Neural Systems\u003c/em\u003e, \u003cem\u003e12\u003c/em\u003e(3), 255. https://doi.org/10.1088/0954-898X/12/3/302\u003c/li\u003e\n\u003cli\u003eWixted, J. T. (2014). Sparse and distributed coding of episodic memory in neurons of the human hippocampus. \u003cem\u003eProc. Natl. Acad. Sci. U.S.A\u003c/em\u003e, \u003cem\u003e111\u003c/em\u003e, 9621\u0026ndash;9626.\u003c/li\u003e\n\u003cli\u003eWixted, J. T. (2018). Coding of episodic memory in the human hippocampus. \u003cem\u003eProc. Natl. Acad. Sci. U.S.A\u003c/em\u003e, \u003cem\u003e115\u003c/em\u003e, 1093\u0026ndash;1098.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":false,"highlight":"","institution":"University of California, San Diego","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"sparse coding, episodic memory, human hippocampus, single-unit recording, neuronal allocation, subsequent memory","lastPublishedDoi":"10.21203/rs.3.rs-5731906/v2","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5731906/v2","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"Neurocomputational models hold that individual episodic memories are represented by a sparse, pattern-separated coding scheme in the hippocampus. In addition, recent theories of neuronal allocation suggest that the assignment of individual neurons to a sparse code is non-random and is associated with intrinsic neural excitability. Here, utilizing an independent dataset of single-unit recordings from epilepsy patients, we demonstrate that a relatively small proportion of high-firing hippocampal neurons represent a single item within a recognition memory test. Critically, only items that were both remembered and showed heightened excitability during encoding were preferentially allocated to a sparse, pattern-separated code, one that was selectively present in the hippocampus. Our findings suggest that individual episodic memories are represented by a sparse distributed coding scheme and that neuronal excitability guides the preferential allocation of hippocampal neurons into sparse codes, which in turn supports subsequent retrieval.","manuscriptTitle":"Excitability at encoding determines sparse coding of remembered episodic memories in the human hippocampus","msid":"","msnumber":"","nonDraftVersions":[{"code":"","date":"2025-09-03 10:30:46","doi":"","editorialEvents":[{"type":"decision","content":"Revision requested","date":"2025-09-03T10:30:46+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-09-02T01:54:00+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-08-28T22:36:10+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"315752673990979275435160837191780624689","date":"2025-07-23T17:44:19+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"39269165411822002231853730961862098512","date":"2025-07-23T14:20:05+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-07-23T09:01:07+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-07-22T06:42:11+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2025-07-19T00:36:03+00:00","index":"","fulltext":""},{"type":"notPreprinted","content":""}],"status":"timeline","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}},{"code":2,"date":"2025-04-23 11:02:19","doi":"10.21203/rs.3.rs-5731906/v2","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-05-02T07:31:25+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-04-28T04:11:04+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-04-23T10:54:39+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"315752673990979275435160837191780624689","date":"2025-03-13T20:08:07+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"39269165411822002231853730961862098512","date":"2025-03-05T19:03:20+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-03-05T17:21:07+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-03-05T12:57:12+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-03-05T12:05:34+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-03-05T04:20:59+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2025-02-18T23:01:06+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"e2e9ca37-a5d1-41d2-b6af-3c2302e982fc","owner":[],"postedDate":"April 23rd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":52317970,"name":"Biological sciences/Neuroscience/Cognitive neuroscience"},{"id":52317971,"name":"Biological sciences/Psychology/Human behaviour"}],"tags":[],"updatedAt":"2025-11-24T16:07:39+00:00","versionOfRecord":{"articleIdentity":"rs-5731906","link":"https://doi.org/10.1038/s41598-025-21967-7","journal":{"identity":"scientific-reports","isVorOnly":false,"title":"Scientific Reports"},"publishedOn":"2025-11-18 15:58:53","publishedOnDateReadable":"November 18th, 2025"},"versionCreatedAt":"2025-04-23 11:02:19","video":"","vorDoi":"10.1038/s41598-025-21967-7","vorDoiUrl":"https://doi.org/10.1038/s41598-025-21967-7","workflowStages":[]},"version":"v2","identity":"rs-5731906","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5731906","identity":"rs-5731906","version":["v2"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}