Impact of working memory training on brain network integration and neurotransmitter systems: a resting-state fMRI

Abstract

Working memory training (WMT) has been demonstrated to enhance cognitive performance, yet the underlying neural mechanisms remain insufficiently understood. Brain network connectivity, particularly as measured by the participation coefficient (PC), offers a valuable framework for elucidating these neural changes. This study investigated the effects of WMT on brain network connectivity, utilizing PC as a primary assessment of network integration and segregation. The relationship between WMT-induced changes in PC and the density of specific neurotransmitter receptors was examined. Seventy-six healthy participants were randomly assigned to either a WMT group or a control group. After 8 wks of training, the WMT group exhibited significant cognitive improvements, especially in near and far transfer tasks. These behavioral improvements were accompanied by specific changes in brain connectivity, including a reduction in PC within the sensorimotor network and node-specific alterations in the left prefrontal cortex, temporo-occipital-parietal junction, and parietal operculum. Moreover, changes in PC were significantly correlated with the density of dopamine D2 receptors, mu-opioid receptors, and metabotropic glutamate receptor 5. These findings enhance our understanding of how WMT influences cognitive function and brain network connectivity, highlighting the potential for targeting specific networks and neurotransmitter systems in cognitive training interventions.

Introduction

Working memory (WM) is the cognitive capacity that enables the temporary storage and manipulation of information necessary for complex tasks such as language comprehension, learning, and reasoning (Baddeley 1992). Given its centrality to daily cognitive activities and its decline in numerous neurodegenerative and psychiatric disorders (Huntley and Howard 2010; Weicker et al. 2016), enhancing WM capacity has become a significant focus in clinical and cognitive neuroscience research (Klingberg 2010). Cognitive training interventions, particularly working memory training (WMT), have emerged as promising strategies to enhance WM performance (Klingberg 2010; Fellman et al. 2020). WMT typically involves structured exercises designed to improve the efficiency and capacity of the underlying neural circuits through repeated practice and adaptive difficulty adjustments (Chein and Morrison 2010). Numerous studies have demonstrated that WMT can lead to substantial improvements in WM tasks, along with transfer effects to related cognitive domains such as attention, executive control, and fluid intelligence (Jaeggi et al. 2010; Flegal et al. 2019). However, the neural mechanisms underlying these cognitive enhancements remain inadequately understood.

The advancements in neuroimaging techniques, particularly functional magnetic resonance imaging (fMRI), have facilitated the exploration of brain network connectivity changes associated with cognitive interventions such as WMT (Klingberg 2010; Barnes et al. 2016). Network neuroscience frameworks, such as graph theory, offer valuable tools for quantifying and characterizing the complex interactions within and between large-scale brain networks. The PC, a key metric from graph theory, measures the extent to which a brain region (node) is integrated across multiple functional networks (Guimera and Nunes Amaral 2005). High PC values indicate that a node functions as a hub, facilitating communication and information transfer across diverse neural systems, thereby supporting cognitive integration and flexibility (Bullmore and Sporns 2009; Bertolero et al. 2017). Changes in PC have been observed following cognitive interventions. For instance, networks characterized by high PC nodes, specifically within the sensorimotor network (SMN), ventral attention network, and frontoparietal network, were significantly associated with trained-domain improvements following a 6-wk visual speed of processing training in individuals with amnestic mild cognitive impairment (Chen et al. 2022). The increase in PC is thought to reflect the brain’s adaptation to the cognitive demands of WMT, facilitating more efficient and coordinated processing across neural networks (Pedersen et al. 2020). Therefore, enhancing brain network integration may present a promising target for developing effective cognitive training interventions.

Neurotransmitters, which act as chemical messengers facilitating information transmission between neurons, play a critical role in regulating synaptic transmission and plasticity by binding to postsynaptic receptors (Bear et al. 2020). Synaptic plasticity, a fundamental component for learning and memory (Daoudal and Debanne 2003; Kim and Linden 2007), is essential for cognitive functions such as WM. Different types of neurotransmitters exert distinct regulatory roles in synaptic plasticity, thereby influencing the functions of WM. WM relies on synaptic plasticity in the prefrontal cortex (PFC) and associated brain areas for the temporary storage and efficient processing of information (Miller and Cohen 2001). Previous research has found that striatal dopamine D2 receptor release can increase following a 5-wk WMT, coinciding with significant training-related behavioral improvements (Bäckman et al. 2011; Bäckman and Nyberg 2013). Additionally, changes in other neurotransmitters such as glutamate and GABA, which are essential for maintaining the brain’s excitatory and inhibitory balance, play a crucial role in the synaptic plasticity required for cognitive training effects (Bazzari and Parri 2019; Kolasinski et al. 2019; Zacharopoulos et al. 2021).

The potential link between functional network changes and neurotransmitter activity in WMT remains an important area of investigation. In our previous study, we combined positron emission tomography (PET) and single-photon emission computed tomography (SPECT) neurotransmitter profiles with fMRI data to explore the effects of WMT on neuroplasticity (Fang et al. 2024). We identified significant correlations between neuroimaging measures, such as amplitude of low-frequency fluctuation and gray matter volume, and the spatial distribution of neurotransmitters like serotonin and dopamine. The findings suggested that WMT-induced neuroplastic changes were accompanied by alterations in neurotransmitter activity patterns, offering valuable insights into the underlying neurochemical processes.

Understanding the intricate relationships between WMT, network connectivity, and neurotransmitter density is important for elucidating the neural mechanisms underlying cognitive enhancement. The present study aimed to investigate the effects of WMT on brain network connectivity, using PC as the primary indicator of network integration and segregation. Additionally, by combining PC metrics with neurotransmitter distributions, the interactions between network dynamics and neurotransmitter systems within the context of WMT were explored.

Methods and materials

Participants

The sample size was determined using G^*Power software (Faul et al. 2009), with a statistical power (1 − β) of 0.80, an alpha level of 0.05, and effect size (Cohen’s d = 0.6). This calculation indicated that a total sample size of 72 would provide sufficient power to detect group differences in matched pairs. The study included 76 healthy participants, who were recruited through online advertisements and university mailing lists. To be eligible for inclusion, participants had to meet the following criteria: no history of neurological or psychiatric disorders, normal or corrected-to-normal vision, and no use of medications that could affect cognitive function. Exclusion criteria included a history of substance abuse, prior engagement in cognitive training programs, or any medical conditions that would contraindicate MRI scanning (eg the presence of metal implants or pacemakers). An independent researcher who was not involved in the data collection or analysis performed the randomization process. Participants were randomly assigned to either the WMT group or the control group in a 1:1 ratio, using a computer-generated random number sequence. Both the participants and the researchers conducting the cognitive assessments were blinded to group allocation. This blinding procedure was implemented to minimize any potential bias in the outcome measures.

WMT protocol

Adaptive running memory task

A computerized adaptive running memory task was employed to train WM (Zhao et al. 2013), which incorporated three categories of memory items: animals, letters, and locations. The running memory task is designed to measure the updating process (Klingberg 2010). In each trial, participants sequentially view various stimuli presented centrally on a computer screen, with the number of items per trial ranging from 5 to 11. Participants were asked to remember the last three stimuli presented in each sequence. To optimize WM capacity, participants were required to refresh their memory continuously through repeated practice. The training consisted of 30 trials, organized into six blocks of five trials each. Initially, each stimulus appeared for 1,750 ms. If a participant correctly answered at least three of five trials within a block, the stimulus presentation duration for the subsequent block was reduced by 100 ms. The duration of each training session was determined by the time required to complete the final block, with no fixed time limit imposed. Participants typically started with about 30 min of training per day, which gradually decreased to around 20 min by the final session. Adherence to training protocol was monitored through the software, which logged participant data. The training program consisted of five weekly sessions, organized as three consecutive training days followed by one rest day, then two additional training days followed by another rest day, over a period of 8 wks, for a total of 40 training sessions.

Control task

The control group performed a simple memory task designed to match the WMT in terms of engagement and duration, but without the adaptive component. Each trial began with a fixation point displayed at the center of the screen for 300 ms, followed by an image of an animal that remained visible for 1,750 ms. Participants completed 10 min of this simple control task each day (Hou et al. 2023).

Training procedure

Sessions were conducted on computers in a controlled lab environment to ensure consistency. Each session lasted approximately 30 min, with task difficulty automatically adjusted based on the participant’s performance.

Cognitive assessments

Evaluation of near-transfer effects

Near-transfer effects were assessed through three specific tests closely related to WM tasks: the update function test (UFT), inhibition function test (IFT), and switching function test (SFT). The UFT requires participants to memorize and recall a sequence of letters, evaluating their ability to update information in WM The IFT utilizes a stroop task, where participants focus on the numerical values of differently sized Arabic numerals, testing their ability to inhibit automatic responses. The SFT involves task-shifting, where participants must alternate their attention between color-coded numerical significance and parity. Detailed descriptions of these tasks are provided in the prior study (Ma et al. 2017).

Evaluation of far-transfer effects

The assessment of far-transfer effects was conducted using the visuospatial sketchpad test (VST) and the phonological loop test (PLT), both of which measure skills pertinent to WMT. In the VST, participants are required to memorize and recall the locations of black squares on a grid to evaluate their visuospatial memory and attention. The PLT assesses phonological processing by requiring participants to retain and recognize a sequence of capital letters, followed by identifying the corresponding lowercase letter. Further details on these tasks can be found in the previous study (Ma et al. 2017).

Data acquisition and Preprocessing

Resting-state fMRI images were acquired using a 3.0 T MRI scanner with the following parameters: repetition time (TR) of 2,000 ms, echo time (TE) of 40 ms, flip angle of 90°, and slice thickness of 3.5 mm. A total of 210 volumes were collected. T1-weighted anatomical images were acquired using a magnetization-prepared rapid gradient echo sequence with a TR of 8.1 ms, TE of 3.2 ms, and a voxel size of 1 mm³. Resting-state fMRI data were preprocessed using the DPABI toolbox (http://rfmri.org/dpabi) integrated with SPM12 (https://www.fil.ion.ucl.ac.uk/spm/software/spm12). For each participant, the first ten functional volumes were removed. The remaining volumes were corrected for slice acquisition timing and head motion. T1-weighted structural images were aligned to the average functional images and segmented into gray matter, white matter, and cerebrospinal fluid (CSF). Subsequently, the images were normalized to Montreal Neurological Institute space using the DARTEL method. We regressed out signals from white matter, CSF, the global signal, and head motion parameters to mitigate potential confounding effects. Linear detrending and temporal filtering were applied within a frequency range of 0.01 to 0.1 Hz. Finally, spatial smoothing was conducted with a 6-mm full-width at half-maximum Gaussian kernel to facilitate subsequent statistical analyses.

PC calculation

Brain networks were defined using the Scaefer300 atlas (Schaefer et al. 2018), with each region of interest analyzed for its PC to measure its integration across different networks. PC was calculated for each region pre- and posttraining using the NetworkX package (https://networkx.org/) in Python (Hagberg et al. 2008). This analysis focused on changes in node-level PC and network integration within several key brain networks.

Spatial correlation with neurotransmitter density

After identifying PC changes following WMT, the spatial relationships between the PC changes and neurotransmitter density distributions were examined using the JuSpace toolbox (Dukart et al. 2021) (https://github.com/juryxy/JuSpace). The neurotransmitter maps focus on serotonin receptors (5-HT1a, 5-HT1b, and 5-HT2a), dopamine receptors (D1 and D2), Gamma-aminobutyric acid receptor (GABAa), μ opioid receptor (Mu), metabotropic glutamate receptor 5 (mGluR5), N-methyl-d-aspartate receptor, dopamine transporters, serotonin reuptake transporter, vesicular acetylcholine transporter, noradrenaline transporter, and F-DOPA (dopamine synthesis capacity). Correlations were calculated based on the Scaffer300 atlas, with adjustment made for spatial autocorrelation of local gray matter probability. Statistical significance was determined using 5,000 permutations and corrected for multiple comparisons across neurotransmitter maps using the false discovery rate (FDR) method.

Statistical analysis

Demographic and cognitive evaluation data were analyzed using the Statistical Package for the Social Sciences (SPSS, version 21, Chicago, IL) software. Group differences in demographic variables and cognitive performance were assessed using Chi-square tests for categorical data and t-tests (independent samples and paired samples) for continuous data, depending on the nature of the comparison.

For node-specific PC values, independent sample t-tests were used to compare differences between the WMT and control groups, while paired sample t-tests were used to evaluate within-group pre- and posttraining differences. Given the large number of nodes analyzed, a false positive correction with a threshold of P < 1/300 was applied to control for multiple comparisons (Bassett et al. 2009). Network-level PC comparisons followed a similar approach, with initial paired sample t-tests used to assess within-group changes (pre- vs. posttraining) and independent sample ttests used for between-group comparisons (WMT vs. control). Network PC values were further corrected for multiple comparisons using the FDR method, with a threshold of (P < 0.05).

Results

Demographic and cognitive measures

The analysis of demographic variables revealed no significant differences between the WMT group and control group in terms of sex (χ² = 0.033, P = 0.856), age (t = 0.075, P = 0.941), or education years (t = 0.056, P = 0.855) (Table 1).

Posttraining assessments in the WMT group exhibited significant improvements in certain cognitive domains on both near transfer and far transfer tests compared to baseline measurements. Specifically, there was a notable reduction in reaction times (RTs) for tests assessing near transfer effects, including the UFT (t = 2.575, P = 0.013, q = 0.028) and the SFT (t = 2.170, P = 0.035, q = 0.037). Additionally, a significant reduction in RT was observed in the PLT (t = 2.507, P = 0.015, q = 0.016), which assessed far transfer effects. However, no significant differences in accuracy were observed in pre- and posttraining within each group. Furthermore, the control group showed no significant changes in either RT or accuracy between pre- and postassessments. The details are displayed in Fig. 1.

PC changes within networks

In the analysis of network-level PC changes between pre- and posttraining sessions, significant differences were observed in the WMT group (Fig. 2). Specifically, the WMT group exhibited a significant reduction in PC within the SMN following the intervention. In contrast, the control group showed no significant changes in PC across the networks between the pre- and posttest measurements.

PC changes of nodes

There is no difference in the comparison of node-specific PC between the WMT and control groups at pretest. In the comparison of node-specific PC between the WMT and control groups at posttest, significant differences were observed in specific brain networks (Fig. 3). Notably, a reduction in PC was found in the frontoparietal network, specifically in the left prefrontal cortex (PFC_6). Conversely, an increase in PC was observed in two nodes within the ventral attention network, including the temporo-occipital-parietal junction (TempOccPar_3) and the parietal operculum (ParOper_2). No significant differences were observed in the PC of these nodes when comparing pre- and posttest measurements within each group individually.

Associations with neurotransmitter density

No significant pre- to posttest changes in PC values were observed in the control group. The primary focus of this study is on the PC changes observed in the WMT group and their associations with neurotransmitter density. The analysis of correlations between PC changes (t-values) following WMT and neurotransmitter density revealed significant associations with specific neurotransmitter systems (Fig. 4). Notably, changes in PC were significantly correlated with the density of D2, MU, and mGluR5 receptors.

Discussion

The present study investigates the changes in brain network connectivity induced by WMT, using the PC as the primary indicator of network integration and segregation. Our findings demonstrated that WMT not only resulted in significant cognitive improvements, particularly in tasks requiring near and far transfer effects, but also induced specific alterations in brain network connectivity. These changes were most pronounced in the reduction of PC within the SMN and node-specific changes within the frontoparietal and ventral attentional network. Moreover, we found significant correlations between these PC changes and neurotransmitter density, particularly with D2, MU, and mGluR5. These results suggest that WMT facilitates cognitive enhancement by modulating brain network connectivity, with these neuroplastic changes closely linked to key neurotransmitter systems. The findings provide valuable insights into the neural mechanisms through which WMT exerts its effects on cognitive function.

Behavioral improvement

The observed cognitive improvements in the WMT group, particularly in tasks requiring near and far transfer effects, align with previous research indicating the effectiveness of WMT in enhancing cognitive performance (Chein and Morrison 2010; Schneiders et al. 2011). The significant reduction in RT on tasks such as the SFT and the PLT suggests that WMT leads to more efficient cognitive processing. These behavioral findings are consistent with the idea that WMT enhances the brain’s ability to update, manipulate, and maintain information, which are core components of WM (Zhao et al. 2013; Constantinidis and Klingberg 2016; Pappa et al. 2020). The absence of significant changes in the control group further underscores the specificity of the WMT’s effects, confirming that the cognitive gains observed were directly attributable to the training intervention.

Network-level PC changes

At the network level, our findings revealed a significant reduction in PC within the SMN following WMT. The SMN is essential for processing and integrating sensory and motor information, which supports the coordination of cognitive processes critical for WM tasks (Buchsbaum and D'Esposito 2019). Previous studies have reported that involvement of the motor system in WM is inversely related to individual WM capacity. This suggests that improvements in motion processing efficiency may be related to enhancements in WM performance (Marvel et al. 2019). In line with this, the reduction of PC within the SMN may reflect a decrease in redundant cross-network interactions, enabling the SMN to focus more effectively on its primary functions following WM training. These changes contribute to improved cognitive control mechanisms, optimizing the allocation of neural resources and facilitating better performance in WM tasks.

Node-specific PC changes

The node-specific analysis revealed more granular insights into how WMT affects brain connectivity. A decrease in PC was observed in the FPN, particularly in the left prefrontal cortex (PFC_6). The PFC is integral to executive functions, including cognitive control and decision-making, with the left PFC playing a pivotal role in the manipulation and maintenance of information (Leung et al. 2023; Zhou et al. 2023). The decrease in PC within the FPN suggests enhanced segregation within this executive control network, allowing for more focused and effective cognitive processing.

Conversely, increases in PC were observed in two nodes within the SAN, including the temporo-occipital-parietal junction (TempOccPar_3) and the parietal operculum (ParOper_2). The temporo-occipital-parietal junction (TPJ) is a key hub within the SAN, involved in detecting and integrating significant stimuli from both the external environment and internal cognitive processes. It plays an essential role in shifting attention and allocating cognitive resources (Corbetta and Shulman 2002; Seeley et al. 2007; Carter and Huettel 2013), particularly for tasks that require quick adaptation to novel information. The parietal operculum, located within the lateral sulcus and covering the insula, is crucial for sensory processing and somatosensory perception in integrating complex stimuli (Maule et al. 2015). These increases suggest that the TPJ and parietal operculum have become more central within the brain’s network architecture, enhancing their ability to facilitate attentional control and process salient stimuli. The results reflect an adaptive mechanism by which WMT optimizes network integration and segregation, prioritizing cognitive tasks and minimizing interference from irrelevant information. This finding is consistent with previous research that cognitive training can promote functional reorganization, neural efficiency, and adaptability (Metzler-Baddeley et al. 2016; Thompson et al. 2016).

Neurotransmitter receptor correlations

The relationships between multiple neurotransmitter systems and PC change were found for dopamine D2, MU, and mGluR5. Dopamine D2 receptors are crucial in modulating executive functions and cognitive flexibility, playing a pivotal role in WM and attention. Dopamine influences WM capacity by regulating the stability and flexibility of neural representations in the PFC (Fallon et al. 2013). The positive correlation of PC changes with D2 receptor density suggests that enhanced dopaminergic signaling may facilitate greater network integration within key cognitive regions (Wang et al. 2004; Dodds et al. 2009; Bäckman and Nyberg 2013). This supports the notion that WMT induces neuroplastic changes that optimize dopaminergic signaling pathways, thereby enhancing WM performance (Bäckman and Nyberg 2013; D'Esposito and Postle 2015; Bäckman et al. 2017). mGluR5, known for its role in synaptic plasticity and excitatory neurotransmission, is essential for the maintenance and manipulation of information within WM (Gillick and Zirpel 2012). Similarly, the positive correlation with mGluR5 density underscores the role of glutamatergic signaling in synaptic plasticity and network reorganization. This suggests that WMT may enhance cognitive function through excitatory neurotransmission mechanisms (Homayoun et al. 2004; Lepannetier et al. 2018).

Furthermore, the involvement of MU receptors, typically associated with reward processing and pain modulation (Lueptow et al. 2018; van Steenbergen et al. 2019), plays a key role in motivation and cognitive control. MU receptors in the limbic system are believed to enhance motivation for rewards, thereby increasing the effectiveness of cognitive training by boosting motivational drive (van Steenbergen et al. 2019). These findings collectively suggest that WMT not only induces functional reorganization of brain networks but also engages specific neurochemical pathways that underpin these neural adaptations.

Conclusion

This study demonstrates that WMT enhanced cognitive functions, accompanied by specific changes in brain network connectivity and neurochemical modulation. Reductions in PC within the SMN, along with node-specific changes in the FPN and SAN, suggest a reorganization of brain networks linked to cognitive improvements. The correlations between PC changes and neurotransmitter receptor densities highlight the neurochemical underpinnings underlying these neural adaptations. These findings provide valuable insights into how WMT influences brain function and cognitive performance.

Author contributions

Chaozong Ma (Conceptualization, Writing—original draft), Yijun Li (Conceptualization, Data curation), Yuntao Gao (Conceptualization, Formal analysis), Xinxin Lin (Investigation), Yilin Hou (Data curation, Formal analysis), Wei He (Investigation, Validation), Yuanqiang Zhu (Conceptualization), Jun Jiang (Methodology), Yuanjun Xie (Formal analysis, Methodology, Supervision, Writing—original draft, Writing—review & editing), Peng Fang (Data curation, Funding acquisition, Supervision, Writing—review & editing), and all authors contributed, read, and approved the final version.

Funding

This work was funded by the National Natural Science Foundation of China (No. 32471081, 61806210), Shaanxi Provincial Natural Science Basic Research Program (2024JC-YBQN-0209, S2024-JC-QN-2303), School-level project (No. 2023KXKT058), National Key Laboratory of Unmanned Aerial Vehicle Technology (No. WR202420-2), and Clinical Medicine and X Research Center Project (LHJJ24XL03).

Conflict of interest statement

There were no conflicts of interest in this work.

Ethics statement

This study was approved by the Institutional Review Board of First Affiliated Hospital of Fourth Military Medical University (KY20183115-1). Prior to participating in the study, each participant provided written informed consent.

References

Author notes