Dynamic causal modeling of low-density resting-state EEG in long-term meditation practitioners

Abstract

Meditation is a complex cognitive practice associated with significant neurophysiological changes, particularly in long-term practitioners. These individuals represent an ideal human model for investigating neural changes associated to their consistent, frequent, and sustained cognitive engagement. However, in Western societies, long-term practitioners are relatively rare compared to Eastern monastic communities. In this study, we leverage a collaboration with a unique monastic population, the Monks and Geshes of the Tibetan University of Sera Jey in India, to examine the long-term effects of meditation on resting-state effective brain connectivity. Specifically, we hypothesize that different levels of meditation experience modulate intrinsic connections in two resting-state brain networks: the default mode network (DMN) and the salience network (SN). To test this, we apply dynamic causal modeling for EEG to analyze effective connectivity and validate our hypothesis. Our results reveal that long-term meditation practice can alter connectivity within the DMN and SN, with distinct patterns of modulation based on meditator experience. Experienced meditators appear to exhibit enhanced self-referential processing in the DMN and reduced reactivity in the SN, supporting the notion that meditation refines attentional control and internal awareness. These findings provide new insights into the neurophysiological mechanisms underlying long-term meditation and highlight the role of monastic practitioners as an invaluable model for studying experience-related modifications in the human brain.

Introduction

Meditation refers to a set of highly differentiated practices that vary in focus, from relaxation¹, to broader goals such as emotion regulation² and heightened sense of well-being³. To bring structure to this vast field, an initial theoretical framework has been established, drawing from both traditional meditation texts and modern neuroscientific perspectives³. This framework categorizes standard meditation techniques into two broad types: focused attention (FA) and open monitoring (OM). FA entails the voluntary focusing of attention on a chosen object, such as the breath. On the other hand, OM involves non-reactive monitoring of the content of experience from moment to moment³. The goal of this classification was to provide a clear operational definition of meditation, facilitating research on the neurophysiological processes involved in meditation and its long-term effects. Building on this foundation, new classifications encompassing a wider range of meditation techniques have been proposed^4–6.

In the tradition of Tibetan Buddhism—which, starting from the 9th century AD, derives from the Indian tradition that, through an unbroken lineage, dates back to the 5th century BC - two main types of meditation are identified: Concentrative and Analytical^7,8. Concentrative meditation is similar to FA, where focus is placed on an object such as the repetition of a mantra, the breath, or a visualization. Through this practice, a state of calm and concentrated mind emerges, characterized by the suspension of conceptual thought. In contrast, analytical meditation, begins with a period of concentration on an object to calm the mind, after which attention is directed toward the conceptual analysis of a specific topic of the Dharma–that is, Buddhist teaching and philosophy.

The effects of meditation can be divided into two domains: trait effects, referring to the lasting changes in the person, and the state effects, which pertain to the immediate experiences during the meditation practice². Long-term trait effects are believed to have therapeutic benefits at the physical, cognitive, emotional, and psychological levels⁹, including a deepened sense of calmness, increased sense of comfort and well-being, heightened awareness of the sensory field, and a shift in the relationship to thoughts, feelings, and experience of self¹⁰. Additionally, long-term practitioners provide an ideal human model for investigating brain plasticity given their consistent, frequent, and sustained cognitive engagement¹¹.

In this context, a recent trend in meditation research, based on the collaboration between Western universities and monastic institutions^12–16, has the potential to provide valuable new insights into the characterization and study of meditation’s long-term effects. More specifically, these homogeneous communities (e.g., monasteries) offer a unique research setting where scientists can work with a diverse group of meditators sharing a taxonomy crystallized over centuries¹². Compared to Western societies, these communities include practitioners at different levels, from beginners to intermediates to advanced meditators with tens of thousands of hours of experience, all within their natural environment.

The effects of meditation on brain function have been extensively investigated in numerous functional neuroimaging studies, primarily focused on state effects¹⁰. A growing consensus is emerging regarding the activity and connectivity of specific cortical networks, primarily studied using functional magnetic resonance imaging (fMRI) and electroencephalography (EEG)^10,17–19. In this context, the so-called Default Mode Network (DMN) assumes great importance. This functional brain network, observed at rest, includes the medial prefrontal cortex, the posterior cingulate cortex, the posterior inferior parietal lobule, and the precuneus, and is implicated in verbal thinking concerning the self, autobiographical memory, introspection and mind-wandering^20,21. Notably, some types of meditation (e.g., concentrative meditation) have been observed to induce activations and deactivations of DMN nodes, which are associated with cognitive and attentional control. These findings suggest that meditation can reduce spontaneous thoughts about past and future events, leading to a state of consciousness focused on the present moment¹⁸. This effect may have particular relevance in the context of mood-related conditions, where persistent patterns of internally directed thought, such as rumination, are commonly observed²². In such cases, meditation-based interventions have been integrated into therapeutic protocols, with growing evidence supporting their potential to complement traditional treatments, including in populations with treatment-resistant symptoms²³ and in the prevention of relapse²⁴.

Brain networks can be characterized in terms of their activity and, more comprehensively, of their connectivity. Regarding connectivity, several frameworks have been proposed^25,26, focusing on both functional and effective connectivity. Functional connectivity studies the statistical interdependencies among nodes in a network, while effective connectivity refers to the directed influence that one neural system exerts over another²⁷. Among these, dynamic causal modeling (DCM) represents a preferred method for testing and validating hypotheses regarding brain effective connectivity in a specific network²⁸. DCM are biophysically inspired spatiotemporal models designed to answer questions about the architecture underlying neuronal dynamics and to make inferences about key neuronal parameters^29,30.

In this study, we leverage a collaboration between the University of Pisa (Pisa, Italy) and the Monastic University of Sera Jey (Karnataka, Mysore district, India) to investigate how long-term meditation practice reshapes resting-state effective brain connectivity. Unlike previous studies, which often focus on heterogeneous samples of meditators with varying backgrounds and training styles, our study benefits from a rare, homogeneous population of expert practitioners: the monks and Geshes of the Sera Jey Monastery. This setting provides an unprecedented opportunity to explore how meditation experience, accumulated through years of rigorous monastic training, modulates intrinsic brain networks. To this end, we apply DCM to analyze the effective brain connectivity of two key resting-state networks: the DMN and the salience network (SN). The SN plays a crucial role in detecting and filtering relevant stimuli, switching between internal and external attention, and regulating emotional and cognitive processes^31,32, i.e., functions that are systematically trained through meditative practice. Previous fMRI studies have reported significant changes in the functional connectivity of both the DMN and SN after short-term meditation programs³³. Notably, they found a positive correlation between intra-network connectivity in the SN and self-reported mindfulness levels, emphasizing the functional relevance of these networks in meditative states. Here, we combine the undisputed temporal resolution of EEG with the high physiological plausibility of DCM to dig further into the dynamics of effective connectivity during resting-state. Specifically, our central hypothesis is that increasing levels of meditative expertise are associated to stable, experience-dependent modifications in the extrinsic connections within the DMN and SN, reflecting long-term modifications. To test this, we develop a hierarchical Bayesian model that tracks variations in effective connectivity as a function of meditation experience, extending prior research on EEG resting-state connectivity in the general population³⁴. By integrating a dataset of resting-state EEG recordings from Tibetan monks across different levels of expertise, our study provides novel, neuroscientific insights into how meditation systematically reorganizes large-scale brain networks. This work represents one of the first attempts to examine resting-state connectivity dynamics in a highly controlled and culturally consistent population of lifelong meditation practitioners, offering a rare window into the long-term effects of sustained contemplative training on brain function.

Materials and methods

Participants

Twenty-three healthy Tibetan male Geshes and monks from the Sera Jey Monastery participated in the experiment. Volunteers gave their written informed consent to participate in the experiment. The study was conducted according with the guidelines of the Declaration of Helsinki and approved by the ethical committee of the University of Pisa (Review N. 26/2023). We divided participants into three classes based on their level of experience. Nine participants were considered beginners -B- (i.e., <1 year of experience and 20/30min per day of practice), six intermediates -I- (i.e., <10 years of experience and 60/90min per day of practice), and eight were advanced -A- (i.e., full-time meditators with at least 6 months of retreat). A summary of volunteers’ demographics is reported in Table 1.

Table 1.

Volunteers’ demographics.

ID	Category	Age (years)	Retreat periods (years)	Education
1	A	54	5	Geshe
2	A	64	7	Geshe
3	A	51	13	Geshe
4	A	55	5	Geshe
5	A	50	2	Monk
6	A	40	0.5	Geshe
7	I	78	N/A	Monk
8	A	78	9	Monk
9	I	37	N/A	Monk
10	A	53	11	Geshe
11	B	36	N/A	Monk
12	I	34	N/A	Monk
13	B	39	N/A	Monk
14	B	36	N/A	Monk
15	I	41	N/A	Monk
16	B	29	N/A	Monk
17	I	50	N/A	Monk
18	B	39	N/A	Monk
19	B	39	N/A	Monk
20	B	31	N/A	Monk
21	B	36	N/A	Monk
22	I	46	N/A	Monk
23	B	34	N/A	Monk

For each volunteer ID, we report his meditation experience category, age, period of retreat and education. Category can be among beginner (B), intermediate (I) and advanced (A). N/A: not applicable.

Experimental protocol

Before the experiment, participants were asked to fill in a demographic questionnaire reporting their level of meditation experience, the type of meditation usually practiced, the reference system (tantra or sutra) and whether they have been on retreat or not. The questionnaire is the same used in¹².

Then, the experiment began and participants were asked to perform analytical and/or concentrative meditation sessions, with the only constraint of performing an initial five minutes of eyes-closed resting state as in¹². We focused the analysis on the central 3 minutes of the resting period, driven by the empirical need to exclude transitional dynamics. Indeed, the duration of the resting phase varied because participants began meditation at their own pace. Therefore, we excluded the final portion to ensure that our window did not include the cognitive transition into the focused-attention state. Similarly, we were concerned about the initial minute of resting state, as a previous EEG study using DCM to investigate resting-state dynamics showed that connectivity follows a transient dynamic during the first part of the resting period³⁴.

At the end of the session, participants underwent a debriefing about meditation.

EEG signal acquisition and preprocessing

We acquired EEG signals using a portable 19-channel device from EBNEURO (EBNeuro BE PLUS LTM, Florence, Italy) at the sampling frequency of 512Hz. Channels were placed on the scalp according to the 10-20 International System. Then, channels’ impedance was checked and kept below 25k for the entire duration of the experiment.

We filtered EEG signals with a low-pass antialiasing filter at the cutoff frequency of 45Hz, and we performed a downsampling at the sampling frequency of 100Hz. Subsequently, we applied a zero-phase high-pass filter with a cutoff frequency of 1Hz to improve data stationarity. To identify and address bad channels, we applied a three-step procedure that removed channels being flat for more than 5 seconds, channels with excessive line noise (i.e., more than 5 standard deviations relative to all channels), and channels having a correlation lower than 0.7 to their reconstructed version based on nearby channels³⁵. Removed channels were subsequently recovered using spherical spline interpolation, and we then re-referenced the data to the average of all channels³⁶. Finally, EEG data was decomposed into a series of maximally independent components in time using independent component analysis (ICA) via the AMICA algorithm³⁷. Each independent component was inspected based on its scalp map, spectrum, and time course; only components reflecting brain activity were retained. The number of ICs retained over subjects was 11 ± 3 (mean ± standard deviation).

All datasets underwent a final quality check applying the Artifact Subspace Reconstruction (ASR) algorithm³⁵. ASR performs a spatial filtering that removes EEG artifacts based on the variance of principal components (PCs). Clean segments of data are used to estimate PC-specific variance thresholds, and PCs exceeding these thresholds in the uncleaned data are removed before reconstructing the signal. We adopted a conservative cut-off parameter of 30, which effectively removes the majority of artifacts while preserving relevant neural information³⁸. The average amount of data to be rejected according to the ASR across subjects was 0.34% (min: 0%; max: 3.19%), indicating that no significant residual artifacts were left after preprocessing.

Dynamic causal modeling

We imported preprocessed EEG data into SPM12 (Wellcome Trust Centre for Human Neuroimaging; www.fil.ion.ucl.ac.uk/spm/software/spm12). We adopted DCM for cross-spectral densities (CSD) to investigate brain connectivity among cortical regions³⁹. This framework explains observed CSDs by combining a physiologically plausible generative model of interacting cortical regions, and an electromagnetic forward model, which maps such generated (hidden) cortical activity to the observable channel space³⁰. The activity of each region in the network is explained by a neural mass model, which models the interaction among three interconnected neuronal subpopulations: inhibitory interneurons, excitatory spiny stellate cells and pyramidal neurons⁴⁰. Regions are coupled to each other through extrinsic forward (F), backward (B) and lateral (L) connections to resemble the hierarchical organization of the cortex⁴¹.

Accordingly, the activity of each cortical region in the network is described by a set of parameters , including the strength of extrinsic connections (i.e., F, B, L), conduction delays D, intrinsic connections strength (i.e., between subpopulations; ) and excitatory/inhibitory post-synaptic potentials H. Finally, the activity of each region’s subpopulations is projected to the scalp channels through an observer equation, which consists of an electromagnetic forward model accounting for field spread effects. More specifically, this forward model relies on a lead field matrix L, parameterized by sources’ spatial position and orientation , to describe how the electromagnetic field generated by each cortical source propagates to the electrodes. The DCM model is inverted using variational Bayesian methods, obtaining a posterior density estimate of each parameter ⁴². These estimates are then adopted to make inferences about a priori hypotheses on cortical connectivity.

Our goal was to investigate the effect of meditation experience on two canonical resting state networks: the DMN and the SN. Particularly, we aimed at highlighting whether connectivity strength and its fluctuations depended on the meditator level of experience. We reduced the EEG channels data to its first 6 principal components, as a trade-off between computational complexity and variance retained^43,44. This choice led to an explained variance of (mean ± standard deviation). Then, we estimated CSDs in the (4-45) Hz frequency range, using a Bayesian multivariate autoregressive model of order 8 (see⁴⁵ for details). We adopted the event-related-potential (ERP) neural mass model to describe cortical source dynamics⁴⁶. On the other hand, we modeled sources’ spatial activity through equivalent current dipoles (the ECD option in SPM12). Passive volume conduction effects of such dipoles were described through a boundary element model (BEM) of the head, based on the standard head model template from Montreal Neurological Institute (MNI; Montreal, Canada).

Network specification

DCM for CSD explains steady-state dynamics through the interactions among a small number of brain regions⁴⁷. The network can be determined either a priori from the literature^34,47–49 or by applying source reconstruction techniques^47,50. In this work, we defined both the DMN and SN using the guidance from previous literature. Particularly, each network was defined in terms of (x,y,z) position in the MNI space for each node, as well as their extrinsic connections (i.e., forward, backward, lateral)^34,51. A schematic representation of the modeled networks is depicted in Fig. 1. For the DMN, we specified four nodes: left/right lateral parietal area (l/rLP; MNI coordinates: − 46, − 66, 30; 49, − 63, 33), Precuneus (Prec; MNI coordinates: 0, − 58, 0), and medial prefrontal cortex (mPFC; MNI coordinates: − 1, 54, 27). For the SN, we specified five nodes: left and right lateral parietal area (l/rLP; MNI coordinates: 62, − 45, 30; − 62, − 45, 30), left and right anterior prefrontal cortex (l/raPFC; MNI coordinates: − 35, 45, 30; 32, 45, 30), and dorsal anterior cingulate cortex (dACC; MNI coordinates: 0, 21, 36). We chose these nodes based on the works of^34,51. Furthermore, we specified both networks as fully-connected, with forward, backward and lateral connections according to^34,41,52. We treated each node in the networks as an equivalent current dipole (’ECD’ option in SPM12) in the cortex.

Fig. 1.

Effective connectivity models for the default mode network (DMN; left) and the salience network (SN; right). Extrinsic connections were specified according with the hierarchical organization of the cortex: i.e., forward (blue), backward (red), and lateral (gray).

DCM connectivity analysis

We built a three-levels hierarchical model of connectivity exploiting the PEB framework^53,54 as follows:

1

As depicted in Fig. 2, we segmented single-subject EEG data into 1s-long windows without overlap, and we fitted a DCM model with parameters (one model for DMN and SN, respectively) for each subject i and window j. Variance that was not accounted for by the DCM model is represented by the residual white Gaussian noise term . The stationary condition of brain activity is at the basis of DCM for CSDs^39,55. Accordingly, for each subject, we conducted the Kwiatkowski–Phillips–Schmidt–Shin (KPSS) test against the alternative hypothesis of non-stationary time series for each channel, at a significance level of 0.05. False positives due to multiple comparison testing were corrected through the false-discovery-rate (FDR) procedure under dependence⁵⁶. The analysis confirmed data stationarity for each channel and time window of all the participants.

Fig. 2.

Illustration of the connectivity analysis using the parametric empirical Bayes (PEB) framework. A three-level hierarchical model estimates connectivity dynamics: (I) two distinct dynamic causal models (DCM) for cross-spectral density (CSD) (for the default mode (DMN) and salience networks (SN)) are fitted to consecutive 1s-long EEG windows of each subject; (II) subject-specific general linear models (GLM) estimate mean connectivity strength and temporal fluctuations using discrete-time cosine (DTC) basis functions; and (III) group-level analysis evaluates these effects in relation to meditation experience.

For each subject, we then concatenated together DCM estimates related to connections strength across windows, and we modeled between-window differences through a general linear model (GLM; level II in Eq.1). Here, following previous findings on fMRI and EEG resting-state effective connectivity^34,57, we described connectivity dynamics as systematic fluctuations over a constant term, i.e., the baseline mean connectivity. Fluctuations (i.e., between-window effects; GLM regressors) were modeled as a set of five orthogonal discrete-time cosinusoidal (DTC) basis functions with a frequency of 1 cycle/min, 1.5 cycle/min, 2 cycle/min, 2.5 cycle/min and 3 cycle/min:

2

where are the 180 time windows, and ^34,57. These regressors were included in as columns of the between-window design matrix in the GLM model. The Kronecker tensor product of with the identity matrix ensures that hypothesized effects would be modeled for all the M connections of the network. Additionally, we included a constant vector of in the first column of , and we mean-centered with respect to it. Accordingly, we obtained a model where the first regressor (the constant term) would correspond to the mean baseline connectivity, and the between-window effects would add to or subtract from it. The zero-mean white Gaussian residuals model random effects not captured by the factors of interest.

For each subject, the between-window level provided a set of parameters consisting of the posterior density estimates (expected effect size and covariance) for each regressor specified in (i.e., the mean connectivity and its fluctuation over time). To investigate whether these effects were consistent across subjects, and to further test for the hypothesis that meditation experience plays a role on such dynamics, we grouped together these estimates across subjects to specify a second between-subject GLM (level III in Eq.1). Here, alongside the constant term, we included in the between-subject design matrix the level of meditators experience. Specifically, experience level was encoded as 0,1,2 for B,I, and A participants respectively. The experience regressor was mean-centered with respect to the constant term so that the first term of would model consistent effects of between-window dynamics across subjects (i.e., mean connectivity and DTC components), whereas the second component of would model the covariance of between-window effects with experience. As for level II, random effects of the GLM at level III were represented by the zero-mean white Gaussian noise term .

We performed an iterative greedy search on the estimated group parameters to determine their significance. Specifically, we applied Bayesian Model Reduction (BMR) to estimate nested between-subject PEB models, both with and without the effect of a given regressor on a subset of network connections. Each nested model was associated with a posterior probability (Pp) representing its ability to explain the observed data⁵³. We then performed Bayesian Model Averaging (BMA) on the models resulted from the last iteration of the greedy-search procedure. BMA performs an average of the parameter posterior densities across explored models, weighted for their Pp. As a result, we obtained a set of group-level parameters that are no longer dependent on any specific model assumption. Finally, we applied a threshold to the BMA results, removing parameters with a Pp of being modulated by group factors lower than 0.95. This thresholding is based on the free energy of the models estimated during the BMR procedure (see⁵⁴ for more details).

Results

In Figs. 3 and 4 we report the results of the hierarchical connectivity analysis on the DMN and SN, respectively. Results are obtained from the BMA on the last iteration of the greedy-search procedure. Accordingly, among all possible effects on the network connections, only those that maximized the evidence for the best combination of parameters describing observed data is considered. Among these, connections exhibiting a significant average group or experience-related effect (Pp>0.95) are considered to be of interest. Posterior effect sizes here assume the meaning of scaling factors on the prior connectivity strength. Therefore, a positive effect size for the average effect on a given connection would indicate either a higher overall strength, in the case of mean connectivity, or a higher amplitude of strength fluctuation around such mean value, in the case of DTC components. Conversely, a positive effect size for experience would indicate a linear increase of either the connection strength with meditation experience, in the case of mean connectivity, or its fluctuation amplitude in the case of DTC components. These variations are always expressed with respect to prior connectivity values. For each connection, we report the significant posterior effect sizes (expected value and 90% credibility interval) related to the average group connectivity strength and its fluctuation over time as described by DTC1, DTC2, DTC3, DTC4 and DTC5 components (i.e., average effect). Moreover, we report significant effect sizes of meditation experience on these connectivity dynamics (i.e., experience effect).

Fig. 3.

Connectivity analysis results on the forward (F), backward (B) and lateral (L) connections of the default mode network (DMN). The average (i.e., irrespective of meditation experience; Average) mean connectivity strength and its temporal fluctuation according with the three discrete-time cosine (DTC) functions, as well as their interaction with the meditation experience (i.e., beginner, intermediate, advanced; Experience) is reported for each connection in terms of posterior effect size (expected value ± 90% credibility interval). Depicted effects are those resulting from the last iteration of Bayesian model reduction (BMR) and Bayesian model averaging (BMA). Significant effects (i.e., Pp>0.95) are marked by an *, and are graphically resumed in the cortical representation of the network at the bottom.

Fig. 4.

Connectivity analysis results on the forward (F), backward (B) and lateral (L) connections of the salience network (SN). The average (i.e., irrespective of meditation experience; Average) mean connectivity strength and its temporal fluctuation according with the three discrete-time cosine (DTC) functions, as well as their interaction with the meditation experience (i.e., beginner, intermediate, advanced; Experience) is reported for each connection in terms of posterior effect size (expected value ± 90% credibility interval). Depicted effects are those resulting from the last iteration of Bayesian model reduction (BMR) and Bayesian model averaging (BMA). Significant effects (i.e., Pp>0.95) are marked by an *, and are graphically resumed in the cortical representation of the network at the bottom..

For the DMN, we found a negative effect size for the average group forward and backward connections’ strength (Average-Mean of Fig. 3), indicating that over time such connections exhibited a reduced baseline strength with respect to their prior value, irrespective of meditation experience. On the other hand, lateral connections showed a positive effect size, indicating a higher baseline strength with respect to their prior value and irrespective of experience. The forward connection from lLP to Prec did not show a mean strength different from zero. This means that such connection did not contribute to the description of the observed dynamics at the group level. We found a positive effect size for the DTC2 component on both the forward connection from rLP to mPFC and the backward connection from mPFC to Prec (Average-DTC2 of Fig. 3). This means that the strength of such connections varied over time according with a cosinusoidal trend with a period of 1.5 cycle/min around their mean value, irrespective of the meditation experience. Moreover, we found a positive effect size for the DTC3 component (Average-DTC3 of Fig. 3) on the backward connection from Prec to rLP, indicating that its strength fluctuated over time with a period of 1.5 cycle/min and, again, irrespective of meditation experience. Of note, the amplitude of such fluctuations is determined by the amplitude of the DTC component, scaled by the respective posterior effect size. Concerning experience, we did not find a significant effect on the mean connectivity. Accordingly, mean group DMN connectivity strength did not differ based on meditation experience. Conversely, we found an interesting positive effect size for the interaction between experience and the DTC3 component on the backward connection from Prec to lLP (Experience-DTC3 of Fig. 3). This result indicates that the oscillation amplitude of the Prec-lLP connection increases with the level of experience, such that advanced meditators exhibit higher oscillations of this connection strength compared to beginners.

Regarding the results for SN depicted in Fig. 4, we observed a negative effect size for both the average strength of forward and backward connections with respect to their prior value, irrespective of meditation experience. Conversely, we observed a positive effect size for the average strength of lateral connections, always with respect to their prior values and irrespective of experience (Average-Mean; Fig. 4). The backward connection from mdACC to laPFC showed a null effect size, meaning that it did not give a significant contribution to the description of the observed dynamics. Interestingly, we found a negative effect size for experience on the mean strength of the lateral connection from lLP to rLP. This indicates that such connection becomes weaker and more inhibited as the level of meditation experience increases.

Considering temporal dynamics of SN connectivity, we did not find a significant fluctuation as described by DTC components on any connection, irrespective of experience. Nevertheless, we observed a significant negative interaction between experience and the DTC3 component on the backward connection from raPFC to rLP. Therefore, although showing no systematic fluctuation across the entire group of subjects, the raPFC-to-rLP connection strength manifested a tendency to oscillate with an amplitude that is inversely correlated with meditation experience.

Discussion

In this work, we analyzed EEG data of Tibetan Buddhist monks recorded at rest to investigate the hypothesis that meditation experience affects resting-state effective connectivity. Moreover, upon recent characterization of brain connectivity during rest^34,57–60, we investigated whether meditators would show systematic connectivity fluctuations that are conserved over subjects, and the potential role meditation experience plays on such dynamics. Accordingly, we followed the approach presented in^34,57 to build a statistical hierarchical model of connectivity using DCM and PEB. More specifically, for each subject we segmented 3min of resting-state into 1s-long non-overlapping epochs, and two distinct DCM for CSD (i.e., one for the DMN and one for the SN, respectively) were fitted to the data from each window. Then, single-subject between-window differences were modeled by means of a Bayesian linear model as fluctuations around a constant term (i.e., the baseline average connectivity). Such fluctuations were specified as a set of 5 DTC components^34,57. Finally, we modeled between-window differences across subjects through another Bayesian linear model. This allowed us to investigate the temporal trajectories of connectivity that were conserved over subjects, as well as potential differences of such fluctuations related to meditation experience. This hierarchical model was estimated using the PEB framework^53,54. Our results highlight connection-specific fluctuations during resting-state that are consistently shared by meditators, and confirm our initial hypothesis about the key role played by meditation experience on such patterns. A detailed discussion about key findings is reported in the remainder of this section.

The observed time-varying dynamics in the DMN, particularly the forward connection from the rLP to the mPFC, and the backward connections from the mPFC to the Prec and from the Prec back to the rLP, provide novel and valuable insights into the trait modifications of connectivity associated to long-term practice of meditation. These nodes are typically related to self-referential processing, with distinct but complementary roles in constructing and maintaining the sense of self^61–63. The mPFC is crucial for introspective thought and emotional evaluation, often acting as a hub for processing information about the self in relation to personal experiences⁶⁴. Its dynamic interaction with the Prec, a region implicated in integrating self-related spatial and contextual information^65,66, highlights a coordinated mechanism underlying the monitoring of internal states. The rLP, which integrates external contextual information with internal self-referential processes^67,68, may form a crucial link, enabling the transition between externally and internally focused attention^69,70. The observed fluctuations in connectivity strength over time, as described by the set of DTC components, suggest that these interactions are not static but adaptively modulate in response to the demands of maintaining self-awareness^58,59. Specifically, the forward connection from the rLP to the mPFC may represent the transfer of externally contextualized information to the mPFC for higher-order self-referential processing⁶⁷, while the backward connections involving the Prec may suggest a top-down feedback mechanism to maintain coherence in self-related representations^71,72. Interestingly, these fluctuations occurred irrespective of meditation experience, suggesting that the dynamic interplay among these regions may represent a fundamental feature of resting-state self-awareness.

The backward connection from the Prec to the lLP showed a more unique pattern. Indeed, although its strength did not show a consistent fluctuation at the group level, it showed an increasing tendency to oscillate as meditation experience increased. This finding suggests that the feedback from the Prec to the lLP becomes increasingly refined - or regulated - with greater meditative expertise^17,20. Given the lLP’s function in contextualizing Prec’s information with external perspectives^67,68, this result may reflect an enhanced capacity for experienced meditators to dynamically align internal self-awareness with external contextual cues. For experienced meditators, it is plausible that such dynamics reflect an enhanced ability to regulate and refine self-referential processes, a hypothesis supported by existing studies on the effects of meditation on DMN connectivity^19,73,74. These findings underscore the DMN’s role in dynamically sustaining self-awareness and point to its potential modulation through meditative practices.

As opposed to the DMN, SN is primarily responsible for detecting and filtering salient stimuli, both from external environments and internal states, and prioritizing these stimuli for further cognitive processing⁷⁵. The raPFC and rLP are key nodes of the SN, involved in cognitive control, attention, and the detection of emotional and motivational salience^70,75. In the case of the SN, the inverse relationship between the backward connection from the raPFC to rLP and meditation experience may suggest that as meditation experience increases, there is a reduced need for the network to react to salience in the environment. This could reflect an increased regulation of attention or cognitive control, where experienced meditators become less reactive to externally or internally salient stimuli. From a neurophysiological and computational perspective, the forward, backward, and lateral connections estimated through DCM correspond to distinct hierarchical processes within cortical networks^29,30. Forward connections convey ascending sensory or contextual signals, backward connections implement descending predictions and top-down modulation, and lateral connections mediate coordination between regions at the same hierarchical level. Within this framework, the observed reduction in backward connectivity from the raPFC to rLP in the SN with increasing meditation experience may reflect a decreased top-down drive to respond to salience, consistent with attenuated attentional reactivity and enhanced regulatory stability. This shift in the SN dynamics can be interpreted through the lens of the dual-mechanism model of attention⁷⁰, where meditation appears to bias the brain away from stimulus-driven bottom-up capture and toward more efficient top-down control. In the context of the SN, the raPFC-rLP connection typically facilitates the orientation of attention toward behaviorally salient or unexpected stimuli⁷⁵. Our finding of reduced connectivity here suggests that long-term practice may de-automate this orienting response, effectively raising the threshold for what the brain deems a salient distraction. This aligns with the focused attention meditation framework, where practitioners cultivate the ability to monitor the focus of attention and voluntarily disengage from distractors³, thereby reducing the metabolic and cognitive load associated with constant environmental monitoring⁷⁶.

In contrast, enhanced backward connectivity within the DMN (e.g., mPFC to Precuneus) could be interpreted as a reinforcement of top-down integration supporting coherent self-referential representations. This mapping may provide a possible interpretation linking the theoretical model of meditation-related attentional and self-processing refinement and the estimated DCM parameters. Indeed, recent findings suggested that advanced meditators display increased attention directed toward interoceptive activity and/or a single mental object, accompanied by the suppression of distracting stimuli from the external environment⁷⁷. Rather than engaging in automatic or habitual responses, experienced meditators may exhibit more volitional control over their attention, aligning with theories that meditation improves attentional regulation and mindful awareness^74,78. This inverse relationship contrasts with the DMN, where meditation experience tends to enhance connectivity and coherence, particularly in regions related to self-referential processes. In the case of the SN, a reduction in the network’s reactivity could be possibly seen as a refinement in attention regulation. In other words, as meditators become more adept at maintaining a stable internal state, they may become less ”reactive” to distractions, allowing them to focus more effectively on the present moment or on self-referential thought, as mediated by the DMN. These results complement and extend those reported by a previous fMRI study, where a comparison of resting-state SN functional connectivity before and after a short meditation program highlighted a positive correlation between dACC-lLP connectivity and self-reported awareness levels³³.

The effect of meditation experience on the baseline connectivity between the lateral parietal regions in the SN further supports the notion that meditation refines brain network dynamics^74,78. While less experienced meditators exhibit connectivity patterns similar to those of typical subjects, reflecting a more automatic processing of salient stimuli, more experienced meditators show a shift towards more selective connectivity. This change suggests that, with increased experience, meditators become more adept at regulating attention and less reactive to irrelevant stimuli, promoting a stable internal state. In typical subjects, homologous lateral parietal connections are a hallmark feature of the SN, facilitating the integration of external and internal information⁷⁵. This modulation in the SN connectivity aligns with the changes observed in the DMN, where meditation experience may enhance self-referential processing and internal focus, highlighting the broader functional effects of meditation on attentional and self-awareness networks^74,78. Importantly, the complementary patterns observed in the DMN (reinforcement of top-down integration) and the SN (reduction in reactive drive) suggest a functional reorganization of the so-called ”Triple Network” architecture⁷⁹. In this framework, the SN acts as a dynamic switch. In this light, we suggest that a more regulated SN in experienced meditators may facilitate a more stable engagement of the DMN during rest, thus preventing the frequent, involuntary attentional shifts - often termed ”mind-wandering” - that characterize non-meditative states⁸⁰. This suggests that meditation-induced traits are not confined to a single network but emerge from the refined coordination between systems responsible for internal representation and those responsible for external orientation.

In both networks, the effect size associated with group baseline (i.e., average) connectivity was negative for forward and backward connections, whereas it was positive for lateral connections. These findings align with previous studies that applied this methodology to EEG resting-state data of typical subjects³⁴, and are partially supported by previous DCM fMRI studies⁸¹. Notably, we found no significant connection from lLP to Prec in the DMN, nor from dACC to laPFC in the SN. While the former result is corroborated by a previous fMRI DCM study⁸², there does not seem to be a consensus regarding the expected baseline connectivity in resting-state networks across fMRI and EEG studies using DCM^{34,57,82–85}.

It is important to note that the present findings were derived from eyes-closed resting-state recordings obtained prior to any explicit meditative practice. As such, they are intended to capture stable, trait-level aspects of brain organization associated with long-term meditation experience. Nonetheless, we cannot entirely exclude the possibility that residual state-related effects (such as sustained attentional or interoceptive engagement) may have influenced resting connectivity, given the proximity of the resting condition to participants’ habitual meditative routines. Future studies could test the stability and generalizability of these patterns by comparing eyes-closed and eyes-open rest, as well as pre- and post-meditation states, to better disentangle trait-based connectivity from transient state carryover.

A further limitation concerns the interpretation of the resting-state condition. It could be argued that this period may reflect not only trait-related neural dynamics but also an anticipatory state toward the upcoming practice. However, given the short duration of the resting-state period (5 minutes) and the exclusion of the initial and final minutes from the analyses, we consider the impact of such anticipatory effects to be limited. Likewise, although no explicit measures of vigilance or arousal were collected, the brevity of the resting period makes substantial vigilance decrements unlikely. In this regard, previous work has demonstrated the feasibility of estimating dynamic EEG connectivity using DCM within 1-minute resting-state windows³⁴, while more recent evidence indicates that 2–3 minutes of good-quality resting-state data are sufficient for reliable characterization of ongoing neural dynamics⁸⁶. The present analysis window exceeds the minimum duration typically adopted in the field. Nonetheless, future studies could extend the resting-state recording to longer periods and incorporate physiological vigilance markers to further ensure stable and trait-like resting-state estimates.

A potential limitation of our study is the absence of a meditation-naïve control group, which prevents us from ascribing the observed effects exclusively to meditation practice as opposed to broader aspects of monastic lifestyle. Nonetheless, the focus of our analysis on comparing groups with widely different levels of practice, and the finding that meditation experience is largely decoupled from age (=0.31, p = 0.14), suggests that the dynamic features we observed are more likely linked to practice accumulation. Importantly, in traditional Tibetan monastic universities, meditation is not an institutionalized or formally scheduled activity, but rather a personal and self-directed practice. As such, the amount of time spent in the monastery does not necessarily reflect the amount of meditation experience accumulated by each practitioner, further supporting a practice-specific rather than lifestyle-based interpretation of our results. Furthermore, the robust connectivity fluctuations we observed contrast with the findings reported in previous EEG-DCM studies on non-meditating populations [ref], providing additional circumstantial evidence favoring a practice-specific interpretation. Future studies combining meditation-naïve controls and longitudinal designs will be essential to further disentangle lifestyle and practice-related effects and to establish causality. While future work may explore continuous metrics of experience, our categorical assignment based on duration and retreat status was essential to mitigate the risk of recall bias associated with estimating lifetime hours. This approach also avoids the use of Western trait mindfulness scales, whose cultural validity and accurate measurement of constructs are actively debated when applied to traditional Tibetan monastic communities⁸⁷.

Our final dataset, while unique, comprises a relatively small sample (N = 23) divided into three unequally sized groups based on meditation experience: beginners (n = 9), intermediate (n = 6) and advanced (n = 8). This imbalance is a major limitation when testing for interaction effects, as the smallest group size determines the overall statistical power and reduces it. To mitigate these issues, we exploited a hierarchical PEB approach. Compared to other frequentist approaches, PEB ensures that the group-level estimates are informed by both the individual subject-level data and the broader group distribution, offering more reliable conclusions even in the presence of unequal group sizes^54,88. Yet, we acknowledge that this remains a limitation of our current study. Accordingly, future work will build on this approach by increasing the sample size and balancing the groups to further improve the precision of the group-level comparisons and to increase the statistical power for detecting complex interaction effects.

In this work, we modeled connectivity dynamics as fluctuations around a constant term using five DTC components at frequencies of 1 cycle/min, 1.5 cycles/min, 2 cycles/min , 2.5 cycles/min and 3 cycles/min (i.e., 0.0167–0.05 Hz). These frequency values are known to reflect characteristic oscillations of resting-state dynamics^89,90 and their choice is supported by previous EEG and fMRI works carried out on typical subjects^34,57,84. It is nonetheless important to acknowledge that a broader range of components may also be relevant. Indeed, previous fMRI studies have used as many as 400 DTC components to investigate connectivity oscillations between 0.0078 Hz and 0.1 Hz⁸⁴. Yet, an empirical trade-off between the complete characterization of observed dynamics and model complexity exists^54,57. Particularly, modeling a higher number of components would lead to a drastic rise in the number of parameters to estimate. Given the limited number of subjects involved in our study, expanding the number of DTC components would result in an increase of uncertainty and dilution effects on DCM and PEB parameters estimates, making observed effects unreliable⁵⁴. Accordingly, we focused on a smaller set of key components³⁴.

A limitation of our network model is the omission of the anterior insula, a canonical hub of the salience network^51,91. While this region is routinely included in fMRI studies⁹², its depth relative to the scalp renders its contribution to EEG signals less reliable, leading to low signal-to-noise ratio and high uncertainty in DCM parameter estimates. For this reason, we adopted a more conservative network specification excluding the insula, in line with previous EEG-DCM work³⁴.

Another concern may regard the number of EEG channels used to carry out our analysis. Indeed, given the relatively low spatial resolution of the setup, effective connectivity estimates at the source level should be interpreted with caution, as they may be influenced by source inversion inaccuracies^93,94. However, DCM can be viewed as a source reconstruction technique in which connectivity parameters about cortical dynamics are biophysically constrained using a Bayesian statistical approach^29,30,52. As a result, DCM is suited to investigate dynamics on a small and a priori established set of cortical nodes - e.g., by means of their (x,y,z) position on the gray matter -, thus mitigating the problem of low-resolution EEG recordings associated to source localization uncertainty^30,52. Moreover, DCM does not directly rely on channel data to fit connectivity parameters. Rather, it performs a reduction of the channels’ mixture to a small number of orthogonal eigenmodes (typically in the range of 3-10 eigenmodes), and adopts such components to carry out model inversion³⁰. This decomposition is performed to drastically reduce computational complexity while enhancing the robustness of estimates by reducing artifact components^29,30,52. However, it also has the potential benefit of limiting the negative impact of a low number of electrodes on connectivity estimates.

Finally, a limitation that should be considered is related to the localization of cortical equivalent current dipoles during DCM model inversion. Indeed, we cannot exclude that the adoption of a standard template (i.e., the MNI template) for each subject may introduce approximations to the solution of the EEG inverse problem^95,96. Accordingly, future studies may consider the adoption of subject-specific head models as those obtained with MRI for enhancing the reliability of the inverted brain locations.

Conclusion

In this study, we provide evidence that long-term meditation practice is associated with enduring modifications in resting-state brain connectivity. By applying dynamic causal modeling to EEG data collected from a rare and culturally homogeneous population of Tibetan monastic practitioners, we were able to detect both trait-level changes and connectivity fluctuations modulated by meditative expertise. Our findings highlight distinct experience-dependent patterns within the default mode and salience networks, suggesting enhanced self-referential processing and reduced reactivity, respectively, in advanced practitioners. These results offer valuable insights into the neurocognitive mechanisms underlying contemplative training. Moreover, they emphasize the relevance of monastic communities as an exceptional model for investigating long-term brain modifications. Future research will benefit from larger, more balanced cohorts and expanded modeling frameworks to further explore the neural signatures of meditation and their implications for mental well-being and cognitive health.

Author contributions

N.N, N.S., J.K., B.N. and A.L.C. conceived and designed the study. J.S., J.T. and B.N. acquired the data and conducted the experiments. B.N. and A.L.C handled data curation. G.R. carried out formal analysis. G.R., F.B., A.G., and A.L.C. analyzed the data and worked on results interpretation and visualization. G.R., F.B, A.G., B.N. and A.L.C. prepared the original draft. All authors edited and revised the manuscript. A.G., E.P.S., N.V., B.N and A.L.C supervised the study.

Funding

The research leading to these results received partial funding from the Italian Ministry of Education and Research (MIUR) in the framework of the ForeLab Project (Departments of Excellence). This research is partly funded by the European Union – Next Generation EU, in the context of The National Recovery and Resilience Plan, Investment 1.5 Ecosystems of Innovation, Project Tuscany Health Ecosystem (THE), Spoke 3 ”Advanced technologies, methods, materials and health analytics” CUP: I53C22000780001.

Data availability

The preprocessed 5-min long resting-state datasets presented in this article are available as supplementary material. The code used for carrying out the connectivity analysis presented in this paper is available on GitHub at https://github.com/rhogianluca94/Dynamic-Causal-Modeling-of-Low-Density-Resting-State-EEG-in-Long-Term-Meditation-Practitioners.

Declarations

Competing interests

The authors declare no competing interests.