Introduction

Episodic memory, often described as an "autobiographical" memory, refers to the recollection of specific events that occurred at a particular time and place1. This unique form of memory enables us to remember personal experiences with rich contextual details and background information, setting it apart from other forms of memory such as factual memory. The emphasis on context and background in episodic memory contributes to its ability to provide a comprehensive understanding of past events2. The study of episodic memory has increasingly focused on dynamic brain networks, which have provided valuable insights into the spatiotemporal structures of the brain, particularly those associated with memory performance and cognitive-behavioral processes3,4. Dynamic brain networks, a class of brain network models based on temporal data, offer distinct advantages over static models in terms of accuracy, computational efficiency, and adaptability5. By employing the analysis of the temporal evolution of community affiliations within brain networks, dynamic functional networks can describe the brain’s dynamic integration, coordination, and response to both internal and external stimuli across various timescales6.

Research has demonstrated that memory representations pass through multiple stages of transformations to achieve successful long-term memory formation and recall7. The dynamic nature of episodic memory involves the formation and retrieval of complex associations among multiple elements such as people, objects, locations, and emotions8. Understanding the dynamic properties of the brain in episodic memory can provide insights into the brain-behavior relationships involved in memory processes. For example, Cory et al. assessed specific changes in interregional dFC patterns across memory retrieval processes and found that they primarily involved a ventral frontal to temporal-parietal network shift associated with search and initial reactivation of episodic memory traces9. Additionally, Ray et al. explored the dynamic reorganization of the frontal-parietal network during cognitive control and episodic memory. The study indicated that these dynamic changes are crucial for the successful execution of cognitive control and episodic memory3. Furthermore, Solomon et al. studied the role of dynamic theta networks in the human hippocampal formation in supporting episodic memory. They found that during encoding, the connectivity patterns dynamically changed, starting from subregions of the hippocampus and eventually converging at the left entorhinal cortex. In the retrieval-related networks, the involvement of the inferior temporal cortex and CA1 significantly increased, reflecting a major reorganization of the encoding network4. These findings provide new insights into how the brain processes and stores memories and highlight the importance of studying dynamic functional networks for a deeper understanding of episodic memory.

Meanwhile, network neuroscience investigates dynamic network reconfiguration to explore the evolution of community structure over time, by analyzing the association between community organization and network dynamics. Recruitment and integration are two statistical metrics used to quantify network reconfiguration. These metrics provide probabilistic measures of changes in community structure over time. they were employed to investigate interaction preferences of brain networks in healthy individuals during task performance10. Our previous study was to apply dynamic functional connectivity analysis to functional magnetic resonance imaging data during episodic memory and quantify integration and recruitment metrics to reveal abnormal dynamic reconfiguration of brain networks in SCZ11. Additionally, studies have shown that excessive activation of the dorsolateral prefrontal cortex may lead to imbalanced resource allocation during integration processes, thereby affecting memory performance12. Ester et al. found that appropriately activating the Visual network and Somatomotor network during recruitment processes can enhance effective information processing, thereby improving subsequent memory performance13. Thus, we hypothesize that dynamic network reconfiguration can indicate the performance of episodic memory. However, the relationship between dynamic network reconfiguration pattern and episodic memory performance remains unknown.

In this study, we conducted an analysis on data from healthy adults involved in an episodic memory task, namely the paired-associate memory task14. This task is particularly well-suited for investigating the encoding and retrieval processes of episodic memory, as it requires participants to remember and later retrieve the associations between pairs of objects, closely mimicking the kind of memory processes that occur in daily life. Additionally, the task is structured in such a way that encoding and retrieval are distinct phases. This separation allows for a clear and independent assessment of each cognitive process, providing a more nuanced understanding of how these processes contribute to episodic memory function. We constructed time-dependent networks for each individual and employed a community detection algorithm to partition the brain network into distinct regions. Integration and recruitment metrics were then calculated separately to examine whether there was the correlation between episodic memory scores and the reconstruction of brain functional connectivity, and predict memory performance.

Materials and methods

Participants

The data used in this study were acquired from the publicly available OpenfMRI database15. Specifically, the dataset originated from the University of California, Los Angeles (UCLA) Consortium for Neuropsychiatric Phenomics LA5c Study and was obtained with approval from the UCLA Institutional Review Board. The release of the data included a statement confirming that all participants provided written informed consent after receiving a comprehensive explanation14. This investigation involved a cohort of 81 healthy participants (Age (mean ± SD) 32.35 ± 7.29 years, 42 male and 39 female, Memory score (mean ± SD) 3.18 ± 0.50). All methods to the dataset were performed in accordance with the relevant guidelines and regulations.

Paired-associate memory task

In this task, two scans were performed to assess declarative memory encoding and retrieval, respectively. The first scan consisted of a block of 64 encoding trials. Forty of the encoding trials were "memory" trials and 24 were "control" trials. During memory trials, first, two words appeared for 1 s, one on each side of the screen. Then, line drawings of two objects that matched the words appeared above the words and they were presented together for 3 more seconds. One of the objects was drawn in black and white, and one object drawn using a single color (e.g. orange). For control trials, pairs of scrambled stimuli, one black and white and one colored, appeared for 2 s. The participant indicated by button press which side the colored object was on. Participants were instructed to remember the objects and the relationship between the objects. The trial types were intermixed, and the ITI between trials was jittered. The encoding run, in total, was 8.07 min long14.

During the retrieval scan there were 104 total trials: 24 control trials, 40 correct trials, and 40 incorrect trials. The retrieval task required the participants to look at a pair of objects and rate their confidence in their memory of the pairing. There were 4 possible response options ranging from "Sure correct" to "Sure incorrect", allowing the responses to be analyzed as a spectrum or binarized into yes/no type responses. During control trials, on one side of the screen was one of the four retrieval confidence response options "sure correct", "maybe correct", "maybe incorrect", or "sure incorrect". On the other side of the screen was "xxxx". Participants were asked to press the button (ranging from 1 to 4) that corresponded to the response option displayed (for example, a response of 1 if "sure correct" appeared). In the 40 correct trials, items were shown paired as they had been at encoding. During the 40 incorrect trials items were shown paired differently than they were at encoding; some objects are the same, but were just paired incorrectly. The retrieval scan was 8.93 min long14. The above method descriptions for paired-associate memory task were adapted from a previous study by Poldrack et al.14. Finally, memory performance was evaluated in terms of retrieval accuracy.

Memory score calculation

The participants were graded according to their retrieval accuracy of rating confidence in the retrieval trials. There were four possible response options: "completely right" (selecting "sure correct" for the "correct" trials or "sure incorrect" for the "incorrect" trials), "mostly right," "mostly wrong," and "completely wrong." These response types corresponded to scores of 4, 3, 2, and 1, respectively. Subsequently, the scores for each trial were summed, and the average score was calculated to represent the memory performance.

Data acquisition

In our study, fMRI data were obtained using 3T Siemens Trio scanners located at the Ahmanson-Lovelace Brain Mapping Center and the Staglin Center for Cognitive Neuroscience at UCLA14. The fMRI data were collected using a T2-weighted echoplanar imaging (EPI) sequence with specific parameters: slice thickness of 4 mm, 34 slices, repetition time (TR) of 2 s, echo time (TE) of 30 ms, a flip angle of \({90}^{^\circ }\), a matrix size of 64 × 64, field of view (FOV) of 192 mm, and oblique slice orientation. Additionally, a high-resolution anatomical scan using magnetization-prepared rapid acquisition with gradient echo (MPRAGE) was performed, with parameters including TR of 1.9 s, TE of 2.26 ms, FOV of 250 mm, matrix of 256 × 256, slice thickness of 1 mm, and 176 slices. The above method descriptions for data acquisition were adapted from a previous study by Poldrack et al.14.

Preprocessing

Preprocessing of functional data used the Data Processing & Analysis for Brain Imaging (DPABI, http://rfmri.org/dpabi) toolbox. Initially, slice timing and head motion correction were conducted, followed by the removal of covariates. We excluded participants with max head motion exceeding 3.0 mm and 3.0° based on head motion criteria16. This investigation involved a cohort of 81 participants after exclusion. The resulting data were then registered to the standard Montreal Neurological Institute (MNI) space and smoothed using a Gaussian kernel with a full width at half maximum of 4 mm. Subsequently, the images underwent bandpass filtering with a frequency range of 0.01–0.25 Hz. Finally, a functionally defined atlas17 containing 360 regions of interest was employed to extract the time-series data. To observe the correlation between dynamic metrics and memory score at module level during encoding and retrieval tasks, inspired by Yeo et al., we divided 360 regions of interest into 7 subnetworks18, namely Visual (Vi), Somatomotor (SM), Dorsal Attention (DA), Ventral Attention (VA), Limbic (LB), Frontoparietal (FP), and Default mode (DM).

Multi-layer network construction

We implemented a sliding window approach to analyze the fMRI time-series data. Based on previous research, a time window ranging from 30 to 60 s has been suggested as reasonable for studying dynamic functional connectivity using fMRI19. Additionally, Jones and van Kesteren et al. identified reliable communities using a window size of approximately 30 s20,21. Therefore, we utilized 30-s windows with a step length of 2 s (equivalent to 1 TR), resulting in a total of 228 windows (242 time points) during the encoding task and 254 windows (268 time points) during the retrieval task. At each window, inter-regional pairwise Pearson correlations were computed, which obtained 360 × 360 correlation matric for each participant.

Dynamic community detection

For the partitioning of brain communities, we employed the generalized Louvain (GenLouvain) method, as proposed by Ref.10. Shown in Fig. 1. This algorithm demonstrates superior performance compared to other community detection methods by optimizing the multi-layer modular quality function Q for community division. The quality function is described as follows:

Fig. 1
figure 1

Experimental procedure. Schematic illustration of the current analytical framework. The Power 360 atlas was utilized to extract time-series data from fMRI signals. A sliding window approach with a window length of 30 s and a window step of 1 TR (Repetition Time) was employed to construct a multilayer network. The functional connections within each layer were estimated using Pearson correlation. Dynamic community detection was performed for each scan, and the community identities were tracked for each region of interest over time using Module Allegiance. Integration and recruitment metrics were subsequently computed.

Full size image
$$Q=\frac{1}{2\mu }{\sum }_{ijlr}\left\{\left({A}_{ijl}-{\gamma }_{l}{P}_{ijl}\right){\delta }_{lr}+{\delta }_{ij}\left.{\omega }_{jlr}\right\}\delta \left({g}_{il},{g}_{jr}\right)\right.$$

In the formula, \({\text{A}}_{\text{ijl}}\) represents the edge between region i and j in layer l. the element \({\text{P}}_{\text{ijl}}\) gives the components of the corresponding matrix for a null model,\({\upgamma }_{\text{l}}\) is the structural resolution parameter of slice l, the quantity \({\text{g}}_{\text{jl}}\) gives the community assignment of node i in slice l, the quantity \({\text{g}}_{\text{jr}}\) gives the community assignment of node j in slice r,the parameter \({\upomega }_{\text{jlr}}\) is the connection strength between node j in slice l and node j in slice r, the total edge weight in the network is \(\upmu = \frac{1}{2}\sum\nolimits_{{{\text{jr}}}} {{\text{K}}_{{{\text{jr}}}} }\), the strength of node j in slice l is Kjl = kjl + cjl, the intraslice strength of node j in slice l is \({\text{k}}_{\text{j}}\text{l}\), and the interslice strength of node j in slice l is \({\text{c}}_{\text{jl}}={\sum }_{\text{r}}{\upomega }_{\text{jlr}}\). We employ the Newman-Girvan null model within each layer by using \({\text{P}}_{{{\text{ijl}}}} = {\text{k}}_{{{\text{il}}}} {\text{k}}_{{{\text{jl}}}} /{\text{2m}}_{{\text{l}}}\) where \({\text{m}}_{{\text{l}}} = \frac{1}{2}\sum\nolimits_{{{\text{ij}}}} {{\text{A}}_{{{\text{ijl}}}} }\) is the total edge weight in slice l22. The free parameters are the structural resolution parameters, \({\upgamma }_{\text{l}}\), and the interslice coupling parameters, \({\upomega }_{\text{jlr}}\), here assumed to be constant (\({\upgamma }_{\text{l}}=\upgamma\), \(\forall \text{l}\) and \({\upomega }_{\text{jlr}}=\upomega\),\(\forall \text{j}\) and \(\forall \text{l}\ne \text{ r}\),meaning that node j in slice l connects to node j in every slice \(\text{r}\ne \text{l}\) with weight \(\upomega\)). We set γ and ω to their default values (= 1) following prior studies23. The above method descriptions for dynamic community detection were adapted from a previous study by Mattar et al.10.

Due to the inherent stochastic nature of the algorithm, there will be slight variations in the output with each run24,25,26,27. Therefore, we performed the algorithm 50 times, calculated quantitative metrics for each run, and subsequently averaged these metrics to obtain a final assessment.

Module allegiance

The module allegiance matrix quantifies the proportion of layers that assign two nodes to the same community10. It is represented by a square matrix with dimensions 360 × 360, where each element of the matrix indicates the frequency with which two nodes co-occur in the same community across different time windows. The matrix can be mathematically described using the following formula:

$${P}_{ij}=\frac{1}{OT}\sum_{o=1}^{O}\sum_{t=1}^{T}{a}_{i,j}^{t,o}$$

where O is the number of repetitions performed by the GenLouvain algorithm, and T is the number of layers in the multi-layer network, \({a}_{i,j}^{t,o}\) is 1 if nodes i and j belong to the same community, 0 otherwise.

Recruitment and integration

Recruitment and integration metrics can be derived from the module allegiance matrix10 as shown in Fig. 1. Recruitment refers to the probability that a region is assigned to the same community as the nodes within its own module during community detection. The recruitment metric of region i in module S can be described as follows:

$${R}_{i}^{S}=\frac{1}{{n}_{S}}{\sum }_{j\epsilon S}{P}_{ij}$$

In the equation, \({n}_{S}\) represents the number of regions in module S. Regions with high recruitment are more likely to be assigned to communities within their own modules.

Integration is defined as the probability that a region is assigned to the same community as the nodes in other modules during community detection. The integration metric of region i in module S can be described as follows:

$${\text{I}}_{\text{i}}^{\text{S}}=\frac{1}{\text{N}-{\text{n}}_{\text{S}}}{\sum }_{\text{j}\notin \text{S}}{\text{P}}_{\text{ij}}$$

where N is the total number of brain regions, regions with high integration tend to appear in communities dominated by other modules.

Statistical analysis

Paired t-tests with controlling the factors of age, sex and education years were conducted for recruitment and integration indicators during encoding and retrieval tasks of each pair of modules, including the whole brain (WB) (Bonferroni corrections). To reveal the relationship between the dynamic reconfiguration of functional networks and memory performance in the episodic memory task, Partial correlation analysis was performed to detect the relationship between dynamic network metrics and memory performance controlling the factors of age, sex and education years. To accurately identify which correlation results are significant, we performed FDR correction on the correlation results.

Correlation analysis was used to explore the potential link between dynamic functional network reconfiguration and memory performance. To examine how well this link can be utilized to predict memory performance, we constructed a model that can predict individual memory scores based on the dynamic reconfiguration of functional network employing the leave-one-subject-out cross-validation (LOOCV) strategy. Firstly, we perform feature selection by calculating the correlation between dynamic reconfiguration of functional network and memory scores using the training set. The features with a significant correlation (p-value < 0.001, FDR correction) are retained. Secondly, we construct a predictive model using multiple regression approach to fit the selected features with the memory scores from the training set. Finally, we apply this model to the unseen test data to generate memory scores. Throughout this process, each participant is used as a test sample once. Once all the LOOCV folds are completed, we evaluate the predictive performance by examining the correlation values between the predicted and observed memory scores. Inspired by Poldrack et al., we introduce the sum of squared errors (SSE) as another metric for evaluating the performance of predictive models28. Finally, we performed a ranking test (repeated 10,000 times) to assess the statistical significance of the observed behavioral scores. In each iteration, prior to the regression analysis, the memory scores obtained from the participants were randomly shuffled. This allows us to determine whether the predicted performance was achieved by chance.

Results

The dynamic network metrics in whole brain and subnetwork

Figure 2 shows the distribution of dynamic metrics at WB and module level during encoding and retrieval tasks and Fig. 3 further indicate the comparison result of each pair of modules, including WB. We obtained similar results in encoding and retrieval tasks of integration in the comparison of integration between subnetworks and WB. The integration of Vi, FP and DM were higher than WB while SM, DA and VA were lower than WB (Fig. 3A,B). However, in the comparative analysis of recruitment, we observed inconsistent results in encoding and retrieval tasks. In encoding task, the recruitment of DA, VA and FP were higher than WB while Vi, LB, DM were lower than WB (Fig. 3C). In retrieval task, we only observed the recruitment of FP were higher than WB. Vi, SM and DM were lower than WB (Fig. 3D).

Fig. 2
figure 2

The distribution of dynamic metrics at whole brain and module level during encoding and retrieval tasks. (A) integration at encoding task; (B) integration at retrieval task; (C) recruitment at encoding task; (D) recruitment at retrieval task. Vi Visual, SM Somatomotor, DA Dorsal Attention, VA Ventral Attention, LB Limbic, FP Frontoparietal, DM Default mode, WB whole brain.

Full size image
Fig. 3
figure 3

The comparison of dynamic measures at whole-brain and module levels during encoding and retrieval tasks. (A) integration at encoding task; (B) integration at retrieval task; (C) recruitment at encoding task; (D) recruitment at retrieval task. Vi Visual, SM Somatomotor, DA Dorsal Attention, VA Ventral Attention, LB Limbic, FP Frontoparietal, DM Default mode, WB whole brain. (***p < 0.001, **p < 0.01, *p < 0.05, Bonferroni correction).

Full size image

The correlations between dynamic network metrics and memory score

Figure 4 shows the correlation between dynamic metrics and memory score at whole brain during encoding and retrieval tasks. During encoding tasks, a notable negative correlation was observed between the integration of the entire brain and memory performance (r = − 0.320, P = 0.004, Fig. 4). However, the recruitment during the encoding task and the integration and recruitment during the retrieval task are not significantly correlated with overall memory performance (Table 1). The negative correlation is also observed between integration and memory scores within each network in the encoding task (The dynamic measurement of the network is obtained from the average measurement of the brain region belonging to the network.) (Table 1). Particularly in the FP network, the negative correlation is the strongest (r = − 0.364, P = 0.0008). The recruitment and memory scores did not show significant correlations across networks (Table 1). In the retrieval task, the recruitment was significantly positively correlated with memory scores in the Vi, SM, and VA networks, while no significant correlations were observed in other networks. On the other hand, the integration was not significantly correlated with memory performance in any of the networks (Table 1).

Fig. 4
figure 4

correlation between dynamic metrics and memory score at whole brain during encoding and retrieval tasks. (A) Integration at encoding task; (B) integration at retrieval task; (C) recruitment at encoding task; (D) recruitment at retrieval task. (FDR correction).

Full size image
Table 1 Correlations between dynamic metrics and static functional connectivity measures and memory score at whole brain and module level during encoding and retrieval tasks, respectively.
Full size table

To verify the importance of the dynamic properties of functional connectivity, we conducted a comparison of how strongly static functional connectivity measures correlate with memory scores (Table 1). In both the encoding and retrieval tasks, neither integration nor recruitment showed significant correlations with memory performance at whole brain (Table 1). In the encoding task, only the integration of the SM and VA networks showed a significant negative correlation with memory scores (Table 1). In the retrieval task, only the recruitment of the Vi network showed a significant positive correlation with memory scores (Table 1).

The correlations in nodal level regional dynamic reconfiguration metrics

We investigated correlations between dynamic metrics and memory score at node level during encoding and retrieval tasks. The results indicated significant negative correlation is observed between integration and memory scores within 15.0% of the nodes during encoding task (Fig. 5A, Table 2), including Right Inferior Frontal Junction, Right anterior mid-cingulate cortex, Right posterior middle frontal gyrus, Left Inferior Frontal Junction Posterior and so on in the FP network; Right Dorsal Visual Stream, Left Posterior Inferotemporal, Left Fusiform Face Area and so on in the Vi network; Right Lateral Intraparietal Cortex Dorsal, Right Posterior Parietal Cortex, Posterior Parietal Cortex and so on in the DA network; Left Pars Opercularis, Right Posterior Medial Gyrus and so on in the DM network; Right Frontal Operculum, Left Frontal Operculum Pars Triangularis and so on in the VA network (Fig. 5A, Table 2). Although no significant correlation is found between recruitment and memory scores at the network level, approximately 9.17% of the nodes show a significant relationship. Among these nodes, 60.6% showed a significant positive correlation between recruitment and memory scores. Most of them are located in the VA, SM, and DA networks, such as the Right Precentral Gyrus, Right Anterior Cingulate cortex, Right Primary Somatosensory Cortex, Right Posterior Cingulate Cortex, Right Prefrontal Cortex, and so on. A small portion of the positive correlation nodes are located in the Default network, such as the Right Medial Prefrontal Cortex. Additionally, 39.4% of the nodes demonstrated a significant negative correlation between recruitment and memory scores. These nodes are mainly located in the DM and LB networks, such as the Right Subgenual Anterior Cingulate Cortex, Left Transverse Gyral, and Left Posteromedial Orbitofrontal Cortex. A small portion of the negative correlation nodes are located in the Somatomotor network, such as Lateral Middle Belt and Lateral Lateral Belt (Fig. 5C, Table 2).

Fig. 5
figure 5

Significant correlations between dynamic metrics and memory score at node level during encoding and retrieval tasks. (A) Integration at encoding task; (B) integration at retrieval task; (C) recruitment at encoding task; (D) recruitment at retrieval task (p < 0.05, FDR correction). The figure was generated by the Connectome Workbench 1.5.0.

Full size image
Table 2 Significant correlations between dynamic metrics and memory score at node level during encoding and retrieval tasks.
Full size table

In the right superior temporal sulcus ventral part and Right Posterior Parietal Cortex of the DM Network, the Right Visual Motion Cortex of the Vi Network and the Left Inferior Precentral Gyrus of the LB Network, the integration shown a significant negative correlation with memory scores during the retrieval task (Fig. 5B, Table 2). The Left Progranular Cortex in the DA network and the Right Transverse Gyral Ventral in the LB network show a significant negative correlation with recruitment and memory performance. Conversely, the Right Posterior Inferior Temporal, Left Middle Temporal and Right Middle Temporal in the Vi network, as well as the Right Frontal Operculum and the Right Superior Central Executive Function in the VA network and the Left Operculum in the SM network, exhibit a significant positive correlation with recruitment and memory performance (Fig. 5D, Table 2).

Prediction analysis for memory score

The results demonstrate that the dynamic reconfiguration of the functional network can be utilized to predict memory scores as shown in Fig. 6. True and predicted memory score were significantly correlated (integration at encoding task: \(\text{r}=0.425,\text{ p}=7.61\times {10}^{-5},\text{ SSE}=16.3\); integration at retrieval task: \(\text{r}=0.575,\text{ p}=1.93\times {10}^{-8},\text{ SSE}=13.5\); recruitment at encoding task: \(\text{r}=0.697,\text{ p}=4.81\times {10}^{-13},\text{ SSE}=10.2\); recruitment at retrieval task: \(\text{r}=0.664,\text{ p}=1.36\times {10}^{-11},\text{ SSE}=11.2\)).

Fig. 6
figure 6

Dynamic reconfiguration of functional network predict memory score. (A) Integration at encoding task; (B) integration at retrieval task; (C) recruitment at encoding task; and (D) recruitment at retrieval task.

Full size image

The influence of window length and step size on the relationship between dynamic network metrics and memory performance

To investigate the impact of different window lengths and step sizes on the relationship between dynamic network metrics and memory performance, we calculated the correlations between dynamic network metrics and memory score under different window lengths and step sizes. Given that the recruitment during the encoding task and the integration and recruitment during the retrieval task did not exhibit significant correlations with memory performance at the whole brain, we instead present the correlations between the integration of the entire brain and memory scores under different window lengths and step sizes during the encoding task (Fig. 7). With the step size fixed at 2 s, when the window length ranged from 30 to 60 s, there was a significant and consistent negative correlation between the integration of the whole brain and memory scores during the encoding task (p < 0.01, Fig. 7A). This is in agreement with previous findings that a time window ranging from 30 to 60 s has been suggested as reasonable for studying dynamic functional connectivity using fMRI16. When the window length was held constant at 30 s, the correlation between the integration of the whole brain and memory scores decreased slightly with the increase in step size during the encoding task, but remained significantly negative across all step sizes tested (p < 0.01, Fig. 7B). The results suggest that different step sizes have minimal impact on the correlation between integration and memory scores.

Fig. 7
figure 7

The correlations between dynamic network metrics and memory score under different window lengths and step sizes. (A) The correlations between the integration of the entire brain and memory scores under different window lengths during encoding task (step size = 2s); (B) The correlations between the integration of the entire brain and memory scores under different step sizes during encoding tasks (window length = 30 s). (**p < 0.01, *p < 0.05, FDR correction).

Full size image

To investigate the potential influence of variable encoding/retrieval time across different windows due to the randomly distributed control trials and jittered experimental design, we implemented a detailed analytical approach. More details can be found in the supplementary materials. Our findings suggest that the distribution of encoding/retrieval trials across windows does not significantly impact our metrics as shown in Figs. S1 and S2.

Discussion

Our findings indicate a significant correlation between integration and memory performance across the entire brain during encoding, with a specific emphasis on the frontoparietal regions. During retrieval, recruitment becomes significantly linked to memory performance in the visual regions, somatomotor regions, and ventral attention regions. At the same time, utilizing the dynamic reconstruction of functional networks enables the prediction of memory scores, and there was a significant correlation between the true and predicted memory scores.

Over integration leads to poor memory performance during encoding task

In encoding tasks, a significant negative correlation between integration and memory scores is observed across networks, particularly in the FP network. Research indicates that the dorsolateral prefrontal cortex (including the dorsolateral prefrontal cortex, the rostral prefrontal cortex, and the ventrolateral prefrontal cortex) plays a crucial role in cognitive control and executive function, especially in the process of information integration29. According to relevant literature, excessive activation of the dorsolateral prefrontal cortex may lead to imbalanced resource allocation during the integration process, thereby impacting memory performance. A study has pointed out a negative correlation between high activation levels of the dorsolateral prefrontal cortex and poorer memory performance12. This could be attributed to the additional cognitive resources required for the integration process in encoding tasks, with the excessive activation of the dorsolateral prefrontal cortex leading to an irrational distribution of resources and consequently weakening memory capacity. Furthermore, Research has shown that interference and inhibition are the two main factors leading to forgetting30. In paired association learning, participants may initially learn a set of stimulus response word pairs, and then during testing, they may recall one of the words as a response to the word paired with it during encoding31. This type of learning can lead to active forgetting, i.e. active interference, as it may affect the participant's ability to recall target information32. This may be due to excessive integration causing interference or competition, leading to the interference or confusion of encoded information and subsequently affecting memory performance. This could be a potential reason why integration performs poorly in predicting memory scores compared to what was expected during recruitment. In summary, the negative correlation between integration and memory scores in encoding tasks may be attributed to the effects of excessive activation of the dorsolateral prefrontal cortex, resulting in imbalanced resource allocation, interference, and information confusion.

Recruitment enhances memory performance during retrieval task

Recruitment and memory scores exhibit a significant positive correlation in the Visual network, Somatomotor network, and Ventral Attention network during retrieval tasks. Research suggests that these networks are closely associated with cognitive processes such as perception, attention, and working memory, thus their activity levels may have a positive impact on memory performance33,34,35. Additionally, studies indicate that appropriate activation of these networks during recruitment processes can enhance effective information processing, leading to improved subsequent memory performance. For instance, research has found that visual details are maintained in memory through the spatial global recruitment during retrieval task13. Similarly, another study demonstrated a relationship between effective regulation of the Ventral Attention network and better memory performance36. Furthermore, the normal functioning of these networks plays a crucial role in maintaining attention and inhibiting interference, which is essential for accurately recalling target information during retrieval tasks. For example, the Somatomotor network, associated with motor control and action planning, can aid in filtering and manipulating relevant memory information, thereby enhancing retrieval accuracy37. In conclusion, the significant positive correlation between recruitment and memory scores during retrieval tasks may be attributed to the appropriate activation of the Visual network, Somatomotor network, and Ventral Attention network, which facilitates information processing, improves memory performance, and helps sustain attention and inhibit interference.

The nodal recruitment influencing memory performance during encoding task

During the encoding phase, nodes that show a significant positive correlation with memory scores are primarily located in the VA, SM, and DA networks. These networks are involved in the brain's processing of external stimuli and attentional control, which are closely related to information encoding and memory formation38. Therefore, the recruitment of these networks during the encoding phase may better reflect changes in memory performance. In contrast, nodes that exhibit a significant negative correlation with memory scores are mainly located in the LB and DM networks. The LB and DM networks are closely associated with higher cognitive processes such as emotion regulation, self-evaluation, and cognitive control. Excessive activity in the LB and DM networks may interfere with attention and cognitive control, leading to competition for resources with other networks, thereby affecting memory encoding and storage. Therefore, an excessive recruitment of nodes in the DM and LB networks during the encoding phase may result in decreased memory performance39. It appears that during the encoding phase, the recruitment of specific brain networks is linked to subsequent memory performance, with different networks showing either positive or negative correlations. This suggests a complex interplay between network activity and memory encoding, shedding light on the underlying mechanisms of memory formation. The presence of positive and negative relationships between recruitment and memory scores within each network leads to the coexistence of nodes that exhibit significant positive correlations and those that demonstrate significant negative correlations between recruitment and memory scores. As a result, there is no significant overall correlation in this state.

Limitation

There are several limitations in this study. The limited sample size may not provide sufficient statistical leverage to detect meaningful differences or relationships in the data, potentially missing important neural patterns associated with memory performance. By overcoming the limitation of a small sample size, future research can provide a more robust and comprehensive understanding of the neural mechanisms underlying episodic memory performance. Additionally, the study neglected the potential significance of inter-layer interactions. The brain's functional networks are highly complex and dynamic, with various layers interacting and communicating to coordinate cognitive processes. During episodic memory tasks, it is crucial to consider how information flows across different layers, as such flow might mediate and facilitate the encoding, storage, and retrieval of memories. Future study can consider both intra-layer and inter-layer connectivity to gain a more comprehensive understanding of the dynamic reconfiguration of the functional network during episodic memory tasks. Furthermore, our findings are primarily based on between-subject individual differences analysis. However, we did not shed light on the relationship between brain dynamics and dynamic memory encoding and retrieval processes as well as their interactions with attention within subjects. Future research could better understand these intra-subject changes by longitudinally tracking the same group of participants or by employing more complex experimental designs, such as event-related potential (ERP) studies, to capture the dynamic changes within individuals.

Conclusions

This study uses a multi-layer network model to explore the dynamic reconstruction of the brain's functional network during scenario memory tasks. The results indicate a correlation between individuals' memory scores and the reconstruction of functional connections. This is consistent with our hypothesis. We found a significant negative correlation between individuals' integration measures during the encoding phase and their memory scores, while the recruitment coefficients during the retrieval phase showed a significant enhancement. Our findings hold crucial significance in uncovering key factors involved in cognitive processes and providing deeper understanding of the mechanisms underlying cognitive functioning.