Skip to main content
NCBI home page
Wiley Open Access Collection logoLink to Wiley Open Access Collection
. 2025 Dec 22;67(2):606–619. doi: 10.1111/epi.18705

Network hubs supporting memory encoding: Implications for cognitive preservation in epilepsy surgery

Barbora Matouskova 1,2, Jan Cimbalnik 3, Pavel Daniel 2,4, Martin Kojan 2,4, Robert Roman 4, Lenka Jurkovicova 2,4, Martin Pail 1,2,3, Milan Brazdil 2,4, Michal Kucewicz 5,6, Petr Klimes 1,✉
PMCID: PMC12927698  PMID: 41427594

Abstract

Objective

Memory impairments are common in neurological disorders, especially drug‐resistant epilepsy, where cognitive deficits such as accelerated long‐term forgetting may occur before surgery due to interictal/ictal discharges, antiseizure medications, and disease progression. Postsurgical memory decline affects 40%–60% of patients. Understanding the neural mechanisms of memory encoding is essential for identifying critical network hubs and guiding surgical interventions that preserve cognitive function.

Methods

We analyzed intracranial electroencephalographic recordings from epilepsy patients (n = 15) performing a verbal memory encoding task to identify key structures involved in successful memory formation. Functional connectivity was computed using multiple features across several frequency bands. Temporal dynamics of connectivity were examined to characterize pre‐encoding and encoding phases.

Results

The middle temporal gyrus (MTG) consistently emerged as a verbal memory hub across features and frequency bands. The inferior frontal gyrus (IFG) was identified as a beta‐band‐specific hub via phase synchrony. Temporal analyses revealed that MTG exhibited stable and widespread connectivity differences across all encoding phases, suggesting a role in pre‐encoding network priming. In contrast, IFG displayed a transient decrease in connectivity and spatial extent during early encoding, indicating a more focal and dynamic contribution.

Significance

These findings suggest that successful memory encoding relies on widespread, frequency‐specific connectivity patterns centered on MTG, complemented by transient, localized IFG activity. Identifying such hubs may inform surgical planning by highlighting regions critical for cognitive outcomes, ultimately supporting strategies to preserve memory in epilepsy surgery.

Keywords: free recall task, functional connectivity, inferior frontal gyrus, intracranial EEG, middle temporal gyrus

Key points.

  • Middle temporal gyrus is a verbal memory hub distinguishing recalled from forgotten words across features and frequencies.

  • Inferior frontal gyrus is a beta‐band hub with a transient drop in connectivity during early encoding of recalled words.

  • Memory encoding relies not only on poststimulus activity but also on prestimulus connectivity, which plays a significant role during encoding.

1. INTRODUCTION

Cognitive impairments are a prominent feature of drug‐resistant epilepsy, with memory dysfunction among the most commonly reported complaints. 1 Long‐term cognitive deficits—including accelerated forgetting and inefficient encoding—can arise from interictal epileptiform discharges, chronic seizure activity, and the cognitive side effects of antiseizure medications. 2 , 3 , 4 Although temporal lobe resections remain an effective treatment for seizure control, they carry a substantial risk of postoperative memory decline, particularly when the resection involves structures critical for memory processing. 5 , 6

The field of epileptology has increasingly recognized epilepsy as a network disorder, with studies highlighting how seizures are driven by dynamic network reconfigurations and how the brain's intrinsic functional architecture can predict its electrical responses. 7 , 8 , 9 As such, identifying and preserving essential memory‐related networks during presurgical planning is of clinical importance. This requires a deeper understanding of the neural mechanisms supporting successful memory encoding in the epileptic brain, particularly the dynamics of large‐scale brain networks that the disease or its treatment may disrupt.

Previous studies highlight the importance of the hippocampus (HIP), medial temporal lobe, prefrontal cortex, and amygdala in memory consolidation, noting that their impairment can lead to significant memory deficits. 10 , 11 , 12 , 13 , 14 However, the functional role and synchronization patterns of these key structures remain insufficiently understood.

Previous studies have shown that successful memory encoding is associated with widespread theta (3–8 Hz) synchronization and high‐frequency gamma (30–100 Hz) desynchronization, suggesting that interregional communication is supported by these activities. 15 Also in the HIP, successful word encoding involves global theta connectivity and decreased gamma connectivity, suggesting that theta oscillations coordinate neural networks for memory. 16 Castelhano et al. 17 found spectral and connectivity differences between dorsal and ventral HIP during various cognitive tasks, and Topçu et al. 18 reported low‐theta activity differences in the left anterior prefrontal cortex during early stages of verbal encoding.

Direct brain stimulation studies have explored the causal role of oscillatory activity. Using intracranial electroencephalography (iEEG), Kucewicz et al. 19 showed that high‐gamma activity evoked by word presentation could be modulated by stimulation, particularly during impaired encoding. In follow‐up work, they reported consistent high‐gamma responses (65–115 Hz) during successful encoding across 16 cortical areas, with increased power during and reductions before and after the word presentation. 20

Closed‐loop stimulation during free recall tasks further implicates distributed network activity. State‐dependent lateral temporal cortex stimulation, guided by machine learning classifiers trained on spectral features across multiple brain regions, enhanced memory by rescuing poor encoding states. 21 , 22 Its effectiveness depends on anatomical and functional factors, with the strongest benefits when stimulation targets regions near white matter with strong low‐frequency (5–13 Hz) connectivity to memory‐related areas. 23

The success of neuromodulation in well‐connected lateral temporal cortex sites 23 suggests that the principles of network integration are key to memory. This led us to a two‐part scientific objective. Our primary hypothesis was that successful memory encoding depends on a sparse set of critical hubs that coordinate communication across large‐scale brain networks. We hypothesized that a systematic, whole‐brain analysis would identify a limited number of regions exhibiting significant connectivity differences between remembered and forgotten words. Second, we aimed to characterize the network properties of any region identified by this data‐driven approach.

2. MATERIALS AND METHODS

2.1. Patients

Our dataset comprised patients with drug‐resistant epilepsy who underwent presurgical iEEG monitoring at St. Anne's University Hospital in Brno from January 2022 until February 2024. During this period, 25 patients with drug‐resistant epilepsy underwent presurgical stereo‐EEG (SEEG) implantation. Of these, 19 consented to participate in our study. One patient did not complete the study, resulting in no recording of the task, and was excluded. Another three patients were excluded due to poor data quality, one because of microphone issues that prevented free recall task annotation and two due to overall poor signal quality. Our final cohort included 15 adult participants (five females and 10 males) with a mean age of 33.3 ± 6.2 years. Additional details are provided in Table S1.

2.2. Task

Our dataset was part of a larger study in which subjects participated in up to five cognitive tasks probing verbal memory and eye movements, with tracking of pupil size and gaze position on the task display screen. Each session began with a smooth pursuit task, followed by a free recall task, a saccadic and anti‐saccadic paradigm, and paired‐associate learning, and concluded with a word screening task. The words used in all verbal tasks were presented in the patients' native languages, either Czech or Slovak. All tasks were recorded with patients comfortably seated in their hospital beds, with one experimental task session conducted per day.

For this paper, we focused exclusively on the free recall task. This task consisted of 15 trials, each divided into four distinct phases: a countdown phase, an encoding phase where 12 words (common nouns) were presented sequentially (each word displayed for 1.5 s with a 1‐s interval between words), a distractor phase with simple arithmetic equations, and a free recall phase where patients were asked to recall in 30 s as many words as possible from the encoding phase in any order (Figure 1A). If a patient completed multiple free recall sessions, only the first one was used.

FIGURE 1.

FIGURE 1

Open in a new tab

Methodological steps used in this study to identify connectivity differences during encoding of recalled and forgotten words. (A) Free recall task used in this study. This task consisted of countdown, encoding, distractor, and free recall phase presented in 15 trials. Patients were asked to encode as many words as possible during the encoding part and recall them during the free recall phase in any order. (B) Connectivity maps displaying the median feature values for all possible channel pairs, shown separately for recalled words (left) and forgotten words (right). (C) Difference matrix representing the difference of median feature values, derived by subtracting the connectivity matrices for recalled and forgotten words from panel B. (D) Significant connections with statistically significant differences (red dots) in connectivity between recalled and forgotten words, defined by a Mann–Whitney U‐test p‐value of <.01 and a medium or large Cliff delta effect size of >.33. The red rectangle represents all possible connections for one of the given structures that are later used to compute metrics to determine verbal memory hubs. SEEG, stereoelectroencephalographic.

The number of words recalled by each patient ranged from 23 (12.8%) to 75 (41.7%), highlighting the variability in recall performance across the cohort. The median number of recalled words was 49, corresponding to approximately 27.2% of the encoded words. This indicates a range of recall abilities within the cohort, which may reflect differences in cognitive performance, attention, or memory processes. Table S2 provides a detailed overview of the number of recalled and forgotten words for each patient, along with the corresponding percentages of recalled and forgotten words. The language used during the task was based on each patient's native language, with Czech or Slovak being the respective languages.

2.3. Recordings

Electrophysiological signals were acquired from intracranially implanted electrodes at a sampling frequency of 5 kHz. These signals were obtained using both standard and hybrid clinical penetrating depth electrodes (AdTech, DIXI Medical), with electrode contacts spaced 5–10 mm apart. All SEEG signals were processed in a unipolar configuration, using the average of all contacts confirmed to be within the brain as the reference. Postimplantation, the precise localization of the electrodes was determined using the Automated Anatomical Labeling Atlas 3. 24

2.4. Electrode implantation

Table S3 provides a summary of electrode coverage for individual subjects. It lists the number of contacts per subject and the spatial distribution of contacts in terms of distance. These values reflect the coverage density and variability of the intracranial recordings and provide context for interpreting intersubject comparisons and connectivity metrics.

Sixteen of the implanted brain structures met the criteria for inclusion in further analysis, having been covered in at least eight patients, corresponding to half of the cohort. These structures represent key regions involved in various cognitive and functional processes and include the amygdala, cingulate gyrus, fusiform gyrus (FFG), HIP, inferior frontal gyrus (IFG), inferior temporal gyrus (ITG), insula (INS), middle frontal gyrus (MFG), middle temporal gyrus (MTG), orbitofrontal cortex (OFC), parahippocampal gyrus, postcentral gyrus, superior frontal gyrus (SFG), superior temporal gyrus (STG), supramarginal gyrus, and temporal pole. Their widespread distribution across cortical and subcortical areas provides a robust foundation for investigating global connectivity patterns in memory processing. Table S4 summarizes the number of patients with electrode coverage in each region, reflecting the distribution of intracranial recordings across cortical and subcortical areas.

2.5. Functional connectivity computing

We computed several bivariate connectivity metrics for each recalled and forgotten word using our EPYCOM Python library. These included the following: coherence, which captures frequency‐specific linear coupling between signals; linear correlation, which quantifies time‐domain similarity in amplitude fluctuations; phase synchrony, which measures the consistency of phase relationships regardless of amplitude; and relative entropy, which reflects information asymmetry between signals. For a detailed explanation of the features, please refer to the EPYCOM library 25 or the feature documentation. The features were calculated across standard frequency bands—delta (1–4 Hz), theta (4–8 Hz), alpha (8–12 Hz), beta (12–20 Hz), low gamma (20–45 Hz), high gamma (55–80 Hz), ripples (80–250 Hz), fast ripples (250–600 Hz)—and for the raw unfiltered recordings, where there was no calculation for phase synchrony, which requires filtering into a specific frequency band.

Event‐related feature computation focused on a 3‐s window, spanning from 1 s before the word presentation to 2 s after the onset on the screen. This interval was subdivided into 21 windows of 500 ms, with a 100‐ms step. Features were computed for each window to track their evolution during word memorizing.

From our list of channels, all channels identified by experts as part of the seizure onset zone were excluded. Features were then calculated for all possible pairs of remaining channels.

2.6. Connectivity maps

For each patient, feature, frequency band, and time window, we computed separate connectivity maps for recalled and forgotten words (Figure 1B). These maps were estimated by taking the median values of the features for all recalled and forgotten words, allowing us to capture distinct connectivity patterns associated with memory recall and forgetting. We then used these connectivity maps to derive a difference matrix (Figure 1C), calculated as the difference between the median feature values from connectivity maps for recalled and forgotten words. This difference matrix allowed us to identify channels with varying connectivity patterns, highlighting connections associated with the differential processing of recalled and forgotten words.

To more accurately assess the connections demonstrating differing connectivity during word encoding for recalled versus forgotten words, we conducted a nonparametric unpaired Mann–Whitney U‐test for each connection individually, using values across all trials. Additionally, we calculated Cliff delta to evaluate the effect size for each connection, categorizing the effect size as negligible, small, medium, or large based on the thresholds of .147, .33, and .474, 26 respectively. We defined significant connections as connections exhibiting differential connectivity during encoding of recalled and forgotten words if they had a p‐value of <.01 and a medium or large effect size of >.33 (Figure 1D).

2.7. Verbal memory hubs for connectivity analysis

For each patient, connectivity feature, frequency band, and time window, we identified brain structures exhibiting the most pronounced differences in global connectivity between the encoding of recalled and forgotten words. To achieve this, we first calculated the proportion of significant connections within each structure, considering a connection to belong to a structure if at least one of its channels was located within that structure (Figure 1D, red rectangle). Additionally, we computed the median distance between channels with significant connections, the median p‐value, and the U‐statistic of significant connections within each structure. Analyses were restricted to brain structures in which at least half of the patients had electrode coverage.

To examine the overall significant structures, we determined the median proportion of significant connections and their median distance within each structure across all patients for each connectivity feature, frequency band, and time window. To identify the most relevant structures distinguishing between recalled and forgotten words and determine verbal memory hubs for each feature and frequency band, we performed an independent two‐sample permutation test for each brain structure. This test compared the distribution of a given structure's median proportion of significant connections across time windows against the corresponding distribution from all other structures combined. The test statistic was the difference in medians between these two groups. For each test, we generated a null distribution by performing 10 000 resamples where structure labels were randomly permuted between the two groups. This procedure directly tests the null hypothesis that a structure's median proportion of significant connections is drawn from the same underlying distribution as that of other structures. A one‐tailed p‐value was calculated to specifically assess whether a given structure exhibited a significantly higher proportion of connections, consistent with our hub hypothesis.

Structures with a p‐value of <.05 in this test were defined as verbal memory hubs, that is, regions that, for a specific connectivity feature and frequency band, exhibited a significantly higher proportion of connections with all other channels showing differences in connectivity between recalled and forgotten words compared to other structures.

To focus subsequent analyses on the most informative aspects of the data, we selected a single connectivity feature and corresponding frequency bands based on their overall ability to identify distinct verbal memory hubs. Specifically, we determined the most relevant feature as the one associated with the highest number of different structures identified as verbal memory hubs across all frequency bands. For each of these verbal memory hubs, we then included only the frequency bands in which the structure was classified as a verbal memory hub. All further analyses of individual hub connections and their temporal profiles were restricted to this selected connectivity feature and the corresponding frequency bands.

Having restricted further analysis to the relevant connectivity feature and selected frequency bands, we examined the spatial characteristics of significant connections for verbal memory hubs. To evaluate whether some hubs exhibited more spatially focal or more widespread connectivity patterns than others, we performed Mann–Whitney U‐tests on the distributions of median distances for all time windows between verbal memory hubs. This allowed us to assess whether the spatial extent of significant connections differed significantly across verbal memory hubs for each connectivity feature and frequency band.

To investigate temporal dynamics, we segmented the time window into three distinct segments: (1) a pre‐encoding segment (750–50 ms before word presentation), (2) an early encoding segment (50–950 ms after word presentation), and (3) a late encoding segment (1050–1750 ms after word presentation). These time points correspond to the midpoints of the time windows used for connectivity analysis. For each segment, we computed the median proportion of significant connections for each connectivity feature, frequency band, and structure, allowing us to examine changes in brain connectivity before and during word encoding. Additionally, Mann–Whitney U‐tests were performed to compare the pre‐encoding segment with both the early and late encoding segments, enabling the identification of significant changes in connectivity associated with word encoding.

2.8. Individual connections of verbal memory hubs

To identify brain structures with robust connectivity to verbal memory hubs, we quantified their connection strength at the structure level, focusing on the specific feature and frequency bands where a verbal memory hub was identified. The procedure began with the set of significant channel‐to‐channel connections from our initial analysis (Figure 1D). We then focused on a specific target structure and a specific hub. For each patient, we calculated a proportion for each target structure and time window to represent its connectivity strength with the hub. This proportion was explicitly defined with a numerator, representing the number of electrode channels within the target structure that had at least one significant connection to any channel within the hub, and a denominator, representing the total number of electrode channels implanted within that same target structure. We then averaged these proportions across all patients to create a group‐level measure. To obtain a single robust value for the target structure, we computed the median of these group‐averaged proportions across all time windows. This allowed us to identify specific brain regions exhibiting robust and temporally consistent memory‐related connectivity with verbal memory hubs.

Next, we investigated the temporal dynamics of these connections with a parallel analysis. The procedure began with the same set of significant channel‐to‐channel connections. For each patient, we again calculated the proportion of channels within each target structure that showed differential connectivity for recalled versus forgotten words. This proportion was then averaged across all patients for each structure and time window to create a group‐level measure. Subsequently, we computed the median of these average proportions across the three previously defined time segments: (1) pre‐encoding, (2) early encoding, and (3) late encoding. This approach allowed us to assess the overall level of differential connectivity within each structure across the three distinct segments of the encoding process.

3. RESULTS

3.1. Implantation distribution

Electrodes were implanted in various brain structures based on clinical requirements for seizure localization and functional mapping. The variability in electrode placement aligns with individual clinical evaluations while ensuring comprehensive sampling of key brain regions involved in cognition and seizure propagation. The distribution of implanted intracranial contacts across anatomical structures is shown in Figure 2. The spatial spread of these electrodes is detailed in Figure S1. Overall, there was substantial variability in the number of contacts per structure across patients. The MTG, HIP, and IFG had the highest median number of contacts. The prominence of the MTG is consistent with the high prevalence of temporal lobe epilepsy, where this region is frequently implicated. Similarly, the HIP was extensively sampled, which aligns with its frequent involvement in epilepsy and the common occurrence of hippocampal sclerosis. The IFG was also heavily targeted, reflecting its well‐established role in language and executive functions, which are often evaluated during presurgical mapping to minimize postoperative deficits.

FIGURE 2.

FIGURE 2

Open in a new tab

Number of contacts implanted in patients within the used structures. Box plots illustrate the number of implanted contacts in brain structures for our patient cohort. The central lines represent the median; the boxes span the interquartile range (IQR) from the first to the third quartile. Whiskers extend to the most extreme data points within 1.5 IQRs of the quartiles, and values beyond this range (outliers) are shown as hollow circles. This visualization provides a clear summary of the central tendency, spread, and presence of extreme values for each implanted structure. AMYG, amygdala; CG, cingulate gyrus; FFG, fusiform gyrus; HIP, hippocampus; IFG, inferior frontal gyrus; INS, insula; ITG, inferior temporal gyrus; MFG, middle frontal gyrus; MTG, middle temporal gyrus; OFC, orbitofrontal cortex; PHG, parahippocampal gyrus; PoCG, postcentral gyrus; SFG, superior frontal gyrus; SMG, supramarginal gyrus; STG, superior temporal gyrus; TPO, temporal pole.

3.2. Verbal memory hubs for connectivity analysis

Our analysis identified the MTG as the most prominent verbal memory hub across all connectivity features (Figure 3). Specifically, for both coherence and relative entropy, the MTG displayed a substantial proportion of significant connections across a wide range of frequency bands, including delta, theta, alpha, beta, low gamma, high gamma, and ripples for coherence, and delta, theta, alpha, beta, low gamma, high gamma, and fast ripples for relative entropy. Although linear correlation also identified MTG as a verbal memory hub, this was limited to the gamma and ripple bands, and phase synchrony highlighted MTG in the alpha, beta, and gamma bands. Notably, phase synchrony uniquely highlighted the IFG as an additional verbal memory hub, restricted to the beta band.

FIGURE 3.

FIGURE 3

Open in a new tab

Verbal memory hubs across features and frequency bands. The box plots illustrate the median proportion of significant connections across patients. Each point represents the median proportion for a specific time window. Brain regions identified as hubs are marked with an asterisk (*). A region was classified as a hub if the median proportion of its significant connections was statistically higher than the median of other regions, as determined by a one‐tailed permutation test with 10 000 resamples. The central lines represent the median; the boxes span the interquartile range (IQR) from the first to the third quartile. Whiskers extend to the most extreme data points within 1.5 IQRs of the quartiles; outliers are not shown.

Because phase synchrony is the only feature identifying multiple structures as verbal memory hubs and MTG appears consistently as a hub across all features, we focused further analysis on phase synchrony, which we considered as the most relevant one. Detailed results for phase synchrony in the alpha, beta, and gamma bands are illustrated in Figure 4.

FIGURE 4.

FIGURE 4

Open in a new tab

Verbal memory hubs identified by the phase synchrony feature, including only frequency bands in which at least one verbal memory hub was detected. The box plots show the median percentage of significant connections across patients. Each point on the plots represents the median value for a specific time window. Brain regions that were classified as hubs are marked with an asterisk (*). A region was identified as a hub if its median percentage of significant connections was statistically higher than the median of other regions, a determination made using a one‐tailed permutation test with 10 000 resamples. The central line represents the median; the boxes span the interquartile range (IQR) from the first to the third quartile. Whiskers extend to the most extreme data points within 1.5 IQRs of the quartiles, and values beyond this range (outliers) are shown as hollow circles.

3.3. Individual connections of verbal memory hubs analysis

To further investigate the functional relevance of the MTG and IFG as verbal memory hubs, we examined their individual connections with other brain structures using the phase synchrony feature, as it yielded the highest number of detected verbal memory hubs across all frequency bands. We focused on alpha, beta, and gamma bands, where these hubs showed significant results. We quantified the proportion of channels in other brain structures exhibiting robust memory‐related connectivity with the MTG and IFG hubs. In Figure 5, MTG connections are displayed in black and IFG in gray.

FIGURE 5.

FIGURE 5

Open in a new tab

Individual connections of the middle temporal gyrus (black) and inferior frontal gyrus (gray) across the entire encoding phase. The figure displays the connections of the middle temporal gyrus and inferior frontal gyrus, which were identified as verbal memory hubs. The width of each line represents the strength of connectivity, quantified by the proportion of channels in a target structure that showed significant memory‐related activity, averaged across patients and time windows. AMYG, amygdala; CG, cingulate gyrus; FFG, fusiform gyrus; HIP, hippocampus; IFG, inferior frontal gyrus; INS, insula; ITG, inferior temporal gyrus; LING, lingual gyrus; MFG, middle frontal gyrus; MTG, middle temporal gyrus; OFC, orbitofrontal cortex; PHG, parahippocampal gyrus; PoCG, postcentral gyrus; SFG, superior frontal gyrus; STG, superior temporal gyrus; TPO, temporal pole.

MTG displayed broad and consistent memory‐related connectivity across all frequency bands, particularly in the gamma range, showing significant interactions with regions involving temporal (ITG, STG, FFG), limbic (HIP, INS), and frontal regions (MFG, SFG, OFC). IFG, identified as a verbal memory hub specifically in the beta band, exhibited more focal connectivity patterns involving MTG, STG, MFG, and INS. To support this observation, we compared the median distances of significant connections for MTG and IFG. The Mann–Whitney U‐test revealed a significant difference (p < .05), with IFG exhibiting shorter median connection distances than MTG, indicating a more spatially focal connectivity profile. These results highlight the MTG as a dominant hub of memory‐related network reconfiguration, with IFG contributing more selectively in beta band, particularly to structures implicated in language and executive processes.

3.4. Temporal dynamics of verbal memory hubs

To examine how memory‐related connectivity evolves over time, we analyzed temporal dynamics in verbal memory hubs identified in the phase synchrony feature across three distinct time windows: pre‐encoding, early encoding, and late encoding.

In the MTG, no significant differences in the proportion of significant connections were observed between the pre‐encoding segment and either the early or late encoding segments across any of the studied frequency bands. In contrast, the IFG in the beta band showed a significant decrease in the proportion of significant connections during the early encoding segment compared to the pre‐encoding window (Mann–Whitney U‐test, p < .05; Figure 6).

FIGURE 6.

FIGURE 6

Open in a new tab

Temporal dynamics of local connections of the middle temporal gyrus (black) and inferior frontal gyrus (gray). Local connections of the middle temporal gyrus and inferior frontal gyrus were determined as a verbal memory hub by the phase synchrony feature. The width of each line represents the strength of connectivity, quantified by the proportion of channels in a target structure that showed significant memory‐related activity, averaged across patients and time windows. Line widths are normalized separately for each frequency band. AMYG, amygdala; CG, cingulate gyrus; FFG, fusiform gyrus; HIP, hippocampus; IFG, inferior frontal gyrus; INS, insula; ITG, inferior temporal gyrus; LING, lingual gyrus; MFG, middle frontal gyrus; MTG, middle temporal gyrus; OFC, orbitofrontal cortex; PHG, parahippocampal gyrus; PoCG, postcentral gyrus; SFG, superior frontal gyrus; SMG, supramarginal gyrus; STG, superior temporal gyrus; TPO, temporal pole.

We further investigated the spatial characteristics of these connections across time windows by analyzing the distances between significantly connected channels. Consistent with the connectivity proportion results, only the IFG in the beta band exhibited a significant decrease in the median distance during the early encoding segment compared to the pre‐encoding window (p < .05), suggesting that the remaining significant connections during this time were more spatially localized. No other structure–frequency combinations showed significant changes in connection distance across the temporal windows.

These results suggest that memory performance is likely influenced not only by connectivity during stimulus presentation but also by the ongoing state of the brain prior to encoding. The absence of abrupt shifts between time windows, particularly in the MTG, supports the idea that successful memory formation may depend on broader temporal dynamics rather than isolated events at the moment of encoding.

4. DISCUSSION

Our study provides compelling evidence for the involvement of the MTG and IFG as key verbal memory hubs that differentiate successful from unsuccessful word encoding in a free recall task. The MTG consistently emerged as a hub across multiple connectivity features and frequency bands, suggesting a broad role in coordinating memory‐related networks. Its stable connectivity across pre‐encoding, early encoding, and late encoding segments implies a persistent brain state that facilitates encoding not only during stimulus presentation, but also beforehand. This broad involvement aligns with the MTG's established functions in semantic memory, language processing, and multimodal sensory integration. 27 , 28 , 29 , 30

Conversely, the IFG was identified as a verbal memory hub specifically through phase synchrony in the beta band. Moreover, the IFG exhibited a significant decrease in the proportion of significant connections during the early encoding segment, accompanied by a reduction in the spatial distance between connected channels, reflecting predominantly intra‐IFG connectivity during this period. This points to a more dynamic role, possibly reflecting executive control and semantic processing during early word processing. 31 , 32 , 33 , 34 , 35

These temporal dynamics highlight a critical aspect of memory encoding; the subsequent recall of a word is not solely determined by the brain's immediate response during word presentation but also by the state prior to word presentation, particularly within the MTG and IFG. This underscores the importance of considering the brain's prestimulus state as a crucial factor influencing memory encoding outcomes, in addition to the neural activity elicited during and after stimulus presentation. 36 , 37 , 38 , 39

Although the absolute median proportions of significant connections within these hubs appear numerically small, the robustness of our findings is underscored by their derivation from a rigorous two‐tiered aggregation process: first across individuals, then across time windows. This methodological approach mitigates the potential influence of individual variability and transient fluctuations, thereby highlighting consistent and reliable patterns of differential connectivity across our cohort.

However, interpreting proportional connectivity measures depends on electrode coverage. In less sampled regions, even a single significant connection may result in a high proportion, potentially inflating the apparent importance of that region due to chance. By contrast, denser sampled regions provide more stable and interpretable estimates. In our dataset, the key regions identified, particularly the MTG and IFG, are generally well sampled across participants, which supports the reliability of the observed effects (Figure 2).

Consequently, despite the seemingly low absolute proportions, the statistical significance established through permutation testing, coupled with the consistent identification of these hubs across multiple connectivity features and frequency bands (particularly for the MTG), strongly supports their critical role in distinguishing whether a word will be subsequently recalled or forgotten, as illustrated in Figure 3.

Our findings offer new insights into the dynamic network architecture of memory encoding, raising important hypotheses for both neuromodulatory therapies and surgical planning in epilepsy. The stable and widespread connectivity of the MTG suggests that this region may play a central role in sustaining the baseline network states that support successful memory encoding. This observation posits the MTG as a candidate region whose functional integrity could be important for postsurgical cognitive outcomes, a hypothesis that warrants investigation in longitudinal studies. In a clinical context, this suggests that presurgical functional mapping could one day incorporate such connectivity profiles to better delineate surgical boundaries around critical memory‐encoding tissue. In contrast, the IFG showed transient, phase‐specific connectivity changes, indicating that its contribution to memory may be tightly linked to pre‐encoding dynamics. More broadly, our results suggest that tailoring interventions based not only on spatial location but also on a region's temporal connectivity profile could be a valuable future avenue for research. Integrating these presurgical functional connectivity patterns with postoperative outcomes is a critical next step. Such studies must also systematically investigate how these network dynamics differ based on resection in the language‐dominant versus nondominant hemisphere, a critical factor in predicting verbal memory outcomes. This work is necessary to determine whether such data can ultimately help clinicians in their efforts to balance seizure control with the preservation of cognitive function.

Our study has several limitations inherent to intracranial EEG research. A primary limitation is the uneven spatial sampling resulting from clinically driven electrode implantations, which could introduce sampling bias. We sought to mitigate this in two ways. First, rather than analyzing raw connectivity values, which can be artificially high between nearby electrodes due to shared signal sources, we focused on the statistical difference in connectivity between subsequently recalled and forgotten words. Second, to account for the varying number of electrodes across regions, our primary metric was the proportion of significant channels within a structure, not the absolute count. Further limitations include clinical heterogeneity in our cohort (e.g., epilepsy type, seizure onset zones, and their lateralization) and the use of two closely related languages (Czech and Slovak), which are potential sources of variance.

5. CONCLUSIONS

In conclusion, our study (n = 15) emphasizes that whether a word will be subsequently recalled or forgotten is not solely determined by the brain's immediate response to the stimulus. Rather, the preexisting functional connectivity state, particularly within the MTG, appears to be a critical determinant. The dynamic changes observed in the IFG further underscore the temporally evolving nature of the neural networks supporting memory encoding. These insights advance our understanding of the neural mechanisms underlying memory and hold potential implications for clinical practice, particularly for developing targeted interventions aimed at enhancing memory function and surgical planning by considering the brain's state both before and during the encoding process.

AUTHOR CONTRIBUTIONS

Conception and design of the study: Barbora Matouskova, Jan Cimbalnik, Michal Kucewicz, and Petr Klimes. Acquisition and analysis of the data: Barbora Matouskova, Jan Cimbalnik, Pavel Daniel, Martin Kojan, Robert Roman, Lenka Jurkovicova, Martin Pail, Milan Brazdil, and Petr Klimes. Drafting of the manuscript and preparing the figures: Barbora Matouskova, Jan Cimbalnik, Michal Kucewicz, and Petr Klimes.

CONFLICT OF INTEREST STATEMENT

None of the authors has any conflict of interest to disclose. We confirm that we have read the Journal's position on issues involved in ethical publication and affirm that this report is consistent with those guidelines.

ETHICS STATEMENT

The study was approved by the Ethics Review Board of St. Anne's University Hospital in Brno (Approval no. 4G/2022, January 12, 2022).

CONSENT

All patients provided written informed consent prior to participation in the study.

Supporting information

TABLE S1.

TABLE S2.

TABLE S3.

TABLE S4.

FIGURE S1.

EPI-67-606-s001.docx (101.5KB, docx)

ACKNOWLEDGMENTS

This work was supported by the National Science Center, Poland under the Weave‐UNISONO call in the Weave program "NeuroWeave" 2021/0 3/Y/NZ4/00082 and the Czech Science Foundation under the Weave‐UNISONO call in the Weave program 22‐28594K.

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.

REFERENCES

  • 1. Novak A, Vizjak K, Rakusa M. Cognitive impairment in people with epilepsy. J Clin Med. 2022;11(1):267. 10.3390/jcm11010267 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2. Elger CE, Helmstaedter C, Kurthen M. Chronic epilepsy and cognition. Lancet Neurol. 2004;3(11):663–672. 10.1016/S1474-4422(04)00906-8 [DOI] [PubMed] [Google Scholar]
  • 3. Helmstaedter C, Witt J‐A. Epilepsy and cognition – a bidirectional relationship. Seizure. 2017;49:83–89. 10.1016/j.seizure.2017.02.017 [DOI] [PubMed] [Google Scholar]
  • 4. Hermann B, Seidenberg M, Lee E‐J, Chan F, Rutecki P. Cognitive phenotypes in temporal lobe epilepsy. J Int Neuropsychol Soc. 2007;13:12–20. 10.1017/S135561770707004X [DOI] [PubMed] [Google Scholar]
  • 5. Baxendale S, Thompson PJ, Sander JW. Neuropsychological outcomes in epilepsy surgery patients with unilateral hippocampal sclerosis and good preoperative memory function. Epilepsia. 2013;54(9):e131–e134. 10.1111/epi.12319 [DOI] [PubMed] [Google Scholar]
  • 6. Rausch R, Kraemer S, Pietras CJ, Le M, Vickrey BG, Passaro EA. Early and late cognitive changes following temporal lobe surgery for epilepsy. Neurology. 2003;60(6):951–959. 10.1212/01.WNL.0000048203.23766.A1 [DOI] [PubMed] [Google Scholar]
  • 7. Keller CJ, Bickel S, Entz L, Ulbert I, Milham MP, Kelly C, et al. Intrinsic functional architecture predicts electrically evoked responses in the human brain. Proc Natl Acad Sci USA. 2011;108(25):10308–10313. 10.1073/pnas.1019750108 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8. Khambhati AN, Davis KA, Oommen BS, Chen SH, Lucas TH, Litt B, et al. Dynamic network drivers of seizure generation, propagation and termination in human neocortical epilepsy. PLoS Comput Biol. 2015;11(12):e1004608. 10.1371/journal.pcbi.1004608 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9. Machegger L, Bosque Varela P, Crespo Pimentel B, Kuchukhidze G. Connectomes in temporal lobe and idiopathic generalized epilepsies. J Clin Med. 2025;14(11):3744. 10.3390/jcm14113744 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10. Huijgen J, Samson S. The hippocampus: a central node in a large‐scale brain network for memory. Rev Neurol. 2015;171(3):204–216. 10.1016/j.neurol.2015.01.557 [DOI] [PubMed] [Google Scholar]
  • 11. Fletcher PC, Frith CD, Grasby PM, Shallice T, Frackowiak RSJ, Dolan RJ. Brain Systems for Encoding and Retrieval of auditory—verbal memory. An in vivo study in humans. Brain. 1995;118(2):401–416. 10.1093/brain/118.2.401 [DOI] [PubMed] [Google Scholar]
  • 12. McGaugh JL, Cahill L, Roozendaal B. Involvement of the amygdala in memory storage: interaction with other brain systems. Proc Natl Acad Sci USA. 1996;93(24):13508–13514. 10.1073/pnas.93.24.13508 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13. Richardson MP. Pre‐operative verbal memory fMRI predicts post‐operative memory decline after left temporal lobe resection. Brain. 2004;127(11):2419–2426. 10.1093/brain/awh293 [DOI] [PubMed] [Google Scholar]
  • 14. Solomon EA, Stein JM, Das S, Gorniak R, Sperling MR, Worrell G, et al. Dynamic theta networks in the human medial temporal lobe support episodic memory. Curr Biol. 2019;29(7):1100–1111.e4. 10.1016/j.cub.2019.02.020 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15. Solomon EA, Kragel JE, Sperling MR, Sharan A, Worrell G, Kucewicz M, et al. Widespread theta synchrony and high‐frequency desynchronization underlies enhanced cognition. Nat Commun. 2017;8(1):1704. 10.1038/s41467-017-01763-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16. Johnson EL, Kam JWY, Tzovara A, Knight RT. Insights into human cognition from intracranial EEG: a review of audition, memory, internal cognition, and causality. J Neural Eng. 2020;17(5):051001. 10.1088/1741-2552/abb7a5 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17. Castelhano J, Duarte I, Bernardino I, Pelle F, Francione S, Sales F, et al. Intracranial recordings in humans reveal specific hippocampal spectral and dorsal vs. ventral connectivity signatures during visual, attention and memory tasks. Sci Rep. 2022;12(1):3488. 10.1038/s41598-022-07225-0 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18. Topçu Ç, Marks VS, Saboo KV, Lech M, Nejedly P, Kremen V, et al. Hotspot of human verbal memory encoding in the left anterior prefrontal cortex. EBioMedicine. 2022;82:104135. 10.1016/j.ebiom.2022.104135 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19. Kucewicz MT, Berry BM, Kremen V, Miller LR, Khadjevand F, Ezzyat Y, et al. Electrical stimulation modulates high γ activity and human memory performance. eNeuro. 2018;5(1):ENEURO.0369‐17.2018. 10.1523/ENEURO.0369-17.2018 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20. Kucewicz MT, Saboo K, Berry BM, Kremen V, Miller LR, Khadjevand F, et al. Human verbal memory encoding is hierarchically distributed in a continuous processing stream. eNeuro. 2019;6(1):ENEURO.0214‐18.2018. 10.1523/ENEURO.0214-18.2018 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21. Ezzyat Y, Kragel JE, Burke JF, Levy DF, Lyalenko A, Wanda P, et al. Direct brain stimulation modulates encoding states and memory performance in humans. Curr Biol. 2017;27(9):1251–1258. 10.1016/j.cub.2017.03.028 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22. Ezzyat Y, Wanda PA, Levy DF, Kadel A, Aka A, Pedisich I, et al. Closed‐loop stimulation of temporal cortex rescues functional networks and improves memory. Nat Commun. 2018;9(1):365. 10.1038/s41467-017-02753-0 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23. Ezzyat Y, Kragel JE, Solomon EA, Lega BC, Aronson JP, Jobst BC, et al. Functional and anatomical connectivity predict brain Stimulation's mnemonic effects. Cereb Cortex. 2024;34(1):bhad427. 10.1093/cercor/bhad427 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24. Rolls ET, Huang C‐C, Lin C‐P, Feng J, Joliot M. Automated anatomical labelling atlas 3. Neuroimage. 2020;206:116189. 10.1016/j.neuroimage.2019.116189 [DOI] [PubMed] [Google Scholar]
  • 25. Cimbalnik J, SweetVlad. ICRC‐BME/Epycom: EPYCOM‐Beta. Zenodo . 2020. 10.5281/ZENODO.4030569 [DOI]
  • 26. Cliff N. Dominance statistics: ordinal analyses to answer ordinal questions. Psychol Bull. 1993;114(3):494–509. 10.1037/0033-2909.114.3.494 [DOI] [Google Scholar]
  • 27. Cabeza R, Nyberg L. Imaging cognition II: an empirical review of 275 PET and fMRI studies. J Cogn Neurosci. 2000;12(1):1–47. 10.1162/08989290051137585 [DOI] [PubMed] [Google Scholar]
  • 28. Chao LL, Haxby JV, Martin A. Attribute‐based neural substrates in temporal cortex for perceiving and knowing about objects. Nat Neurosci. 1999;2(10):913–919. 10.1038/13217 [DOI] [PubMed] [Google Scholar]
  • 29. Rolls ET. The temporal cortex. In: Rolls ET, editor. Brain computations and connectivity, 2nd ed. Oxford University Press; 2023. 10.1093/oso/9780198887911.003.0008 [DOI] [Google Scholar]
  • 30. Tranel D, Damasio H, Damasio AR. A neural basis for the retrieval of conceptual knowledge. Neuropsychologia. 1997;35(10):1319–1327. 10.1016/S0028-3932(97)00085-7 [DOI] [PubMed] [Google Scholar]
  • 31. Fedorenko E, Duncan J, Kanwisher N. Broad domain generality in focal regions of frontal and parietal cortex. Proc Natl Acad Sci USA. 2013;110(41):16616–16621. 10.1073/pnas.1315235110 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32. Feredoes E, Tononi G, Postle BR. Direct evidence for a prefrontal contribution to the control of proactive interference in verbal working memory. Proc Natl Acad Sci USA. 2006;103(51):19530–19534. 10.1073/pnas.0604509103 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33. Gabrieli JDE, Poldrack RA, Desmond JE. The role of left prefrontal cortex in language and memory. Proc Natl Acad Sci USA. 1998;95(3):906–913. 10.1073/pnas.95.3.906 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 34. Jonides J, Nee DE. Brain mechanisms of proactive interference in working memory. Neuroscience. 2006;139(1):181–193. 10.1016/j.neuroscience.2005.06.042 [DOI] [PubMed] [Google Scholar]
  • 35. Ralph MAL, Jefferies E, Patterson K, Rogers TT. The neural and computational bases of semantic cognition. Nat Rev Neurosci. 2017;18(1):42–55. 10.1038/nrn.2016.150 [DOI] [PubMed] [Google Scholar]
  • 36. Gruber MJ, Otten LJ. Voluntary control over Prestimulus activity related to encoding. J Neurosci. 2010;30(29):9793–9800. 10.1523/JNEUROSCI.0915-10.2010 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37. Haque RU, Wittig JH, Damera SR, Inati SK, Zaghloul KA. Cortical low‐frequency power and progressive phase synchrony precede successful memory encoding. J Neurosci. 2015;35(40):13577–13586. 10.1523/JNEUROSCI.0687-15.2015 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 38. Jun S, Kim JS, Chung CK. Prediction of successful memory encoding based on lateral temporal cortical gamma power. Front Neurosci. 2021;15:517316. 10.3389/fnins.2021.517316 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 39. Otten LJ, Quayle AH, Akram S, Ditewig TA, Rugg MD. Brain activity before an event predicts later recollection. Nat Neurosci. 2006;9(4):489–491. 10.1038/nn1663 [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

TABLE S1.

TABLE S2.

TABLE S3.

TABLE S4.

FIGURE S1.

EPI-67-606-s001.docx (101.5KB, docx)

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.

Articles from Epilepsia are provided here courtesy of Wiley

RESOURCES