Abstinence Alters Triple Network Dynamics in Moderate‐to‐Heavy Drinkers

ABSTRACT

Alcohol misuse is a significant public health concern, yet little is known about the neural dynamics associated with habitual heavy drinking, particularly during abstinence. The Triple Network Model, comprising the salience network (SN), default mode network (DMN), and central executive network (CEN), provides a framework for understanding large‐scale brain network dysfunction associated with heavy alcohol use. Using resting‐state fMRI and a Hidden Semi‐Markov Model (HSMM), we examined dynamic brain state changes in moderate‐to‐heavy drinkers (n = 38) across two conditions: typical drinking and alcohol abstinence. Our findings revealed six distinct brain states, with significant differences in state occupancy, transitions, and duration between drinking conditions. Abstinence was associated with decreased time spent in a DMN‐dominant state, a lower probability of transitioning to a state with high SN activation, and more frequent but shorter durations in a state without a distinct dominant network. These results suggest alcohol abstinence alters the temporal dynamics of these brain networks, potentially disrupting attention shifting and cognitive control mechanisms that may contribute to relapse risk. Understanding these neural adaptations will provide critical insight into the neurobiology of habitual heavy drinking and inform potential targets for future interventions.

A hidden semi‐Markov model was used to investigate differences in the dynamics of brain states among moderate to heavy drinkers under two conditions: typical drinking and alcohol abstinence. The results demonstrated altered dynamics, including a reduced probability of transitioning to a state characterized by high salience network activation in the abstinence condition.

1. Introduction

Alcohol misuse is a major economic and public health concern. In the United States, alcohol misuse costs approximately $250 billion annually (Sacks et al. 2015) and is associated with many adverse health outcomes including cancer (Boffetta and Hashibe 2006), fatal injury (Alpert et al. 2022), and cognitive decline (White et al. 2023). In 2023, the National Survey on Drug Use and Health reported more than 85% of adults aged 18 or older reported lifetime alcohol consumption and 51% of adults reported alcohol consumption within the past month, with more than 6% of US adults reporting heavy alcohol use within the past month (NIAAA 2025a; SAMHSA 2023). Alcohol use that exceeds recommendations for safe consumption (NIAAA 2025b) accounts for a tremendous amount of the comorbid disease burden, and even early mortality (Chikritzhs et al. 2015; Griswold et al. 2018; Rehm 2011), making heavy drinking a crucial area of research in the hopes of mitigating the negative effects of alcohol on society.

Substantial effort has gone toward studying the neural basis of alcohol misuse using resting‐state functional magnetic resonance imaging (fMRI) (Rodríguez et al. 2025), an effort that has expanded our understanding of the pathophysiology of alcohol use disorders (AUD). A broader understanding of what changes in the brain accompany heavy alcohol use remains unclear, in part due to the focus on AUD in alcohol‐related research. A large proportion of alcohol‐related research focuses on periods of alcohol abstinence, as it is known to be a state associated with a variety of changes in subjective experience and neurobiology. In contrast to our growing understanding of AUD, relatively little is known about the brain function of moderate‐to‐heavy drinkers, and in particular the resting brain signatures present during abstinence in this population.

Over the past two decades, there has been a growing interest in studying functional brain networks using fMRI. The key insight from these studies is that complex cognitive and perceptual processes do not emerge from isolated brain regions, but from a network of interacting regions. A widely studied network model of cognitive function is the Triple Network Model (Menon 2018), which includes the salience network (SN), the default mode network (DMN), and the central executive network (CEN). The SN includes brain regions that are heavily implicated in the neurocircuitry of addiction, particularly the preoccupation associated with abstinence (Koob and Volkow 2016), including the insula and anterior cingulate. The DMN is a well‐documented functional brain network comprised of the medial prefrontal cortex, precuneus, posterior cingulate, inferior parietal lobules, and lateral temporal cortex (Raichle 2015; Raichle et al. 2001). The DMN is best known for its association with self‐referential thought and internal processing (Davey et al. 2016; Gusnard et al. 2001; Knyazev 2013; Knyazev et al. 2012; Vatansever et al. 2018), but it has also been increasingly implicated in withdrawal affects and craving in individuals with substance use disorders (Zhang and Volkow 2019). Effective connectivity of the DMN has been used to differentiate AUD patients from healthy controls (Khan et al. 2021), and even among alcohol consumers who do not meet diagnostic criteria for an AUD, acute consumption has been documented to decrease functional connectivity within the DMN (Fang et al. 2021; Weber et al. 2014). In contrast to the DMN, the CEN is thought to support external, goal‐oriented processing. In a sample of alcohol consumers, those with a positive familial history of AUD have shown weaker connectivity between the SN and CEN than those without a family history (Parrish 2014). Modafinil, administered to improve cognitive control in alcohol‐dependent patients, has been associated with negative coupling between the DMN and CEN (Schmaal et al. 2013). The Triple Network Model proposes that the SN acts as a “dynamic switch” that directs attention toward either the internal processing of the DMN or the external processing of the CEN. However, aberrant functioning in any of the three subnetworks would have wide‐ranging effects on cognition and behavior.

While the Triple Network Model was built on findings from traditional static functional brain networks, the model inherently contends that the interactions within and between the SN, DMN, and CEN are dynamic. In recent years, attention has been drawn toward dynamic functional connectivity due to mounting evidence in support of the view that brain networks are constantly changing over time (Allen et al. 2014; Chang and Glover 2010). Among numerous strategies for studying dynamic connectivity, state‐based methods have been particularly useful (Allen et al. 2014; Shappell et al. 2019). These methods identify a set of recurring brain states, where each state is characterized by a particular pattern of connectivity between brain regions. Leveraging these methods, researchers have investigated variations in state dynamics across different groups, experimental conditions, and other relevant covariates associated with alcohol use (Maleki et al. 2022; McIntyre et al. 2025; O'Donnell et al. 2025; Zhang et al. 2022). For example, Zhang et al. used a sliding window analysis to show brain state changes associated with acute alcohol intake in healthy individuals. They identified a state characterized by weaker connections in the sensorimotor, basal ganglia, and visual networks as well as the DMN and CEN, linking the state to alcohol's effect on the brain (Zhang et al. 2022). AUD has also been explored in the context of dynamic connectivity. Maleki et al. reported dynamic differences in a cluster encompassing reward, sensorimotor, and frontal functional networks in individuals with AUD (Maleki et al. 2022). However, to our knowledge, no study has examined dynamic functional connectivity in moderate‐to‐heavy drinkers, particularly in the context of alcohol abstinence.

In this study, we aimed to examine differences in dynamic brain states in moderate‐to‐heavy drinkers during a week of typical drinking and a week of alcohol abstinence. Using a Hidden Semi‐Markov Model (HSMM) to infer brain network dynamics (Khodaei et al. 2024; Shappell et al. 2019), we identified brain states and contrasted their temporal dynamics in typical drinking versus abstinence. We focused our analysis on the Triple Network due to extensive prior work implicating each of its subnetworks in different aspects of alcohol use and abstinence.

2. Methods

2.1. Participants

A total of n = 39 (22 female) participants underwent MRI scanning in this study. While significant efforts were taken to promote a diverse sample of participants and a fairly even split of sex was achieved across the sample, participants primarily reported identifying as white (90%), had completed some higher education (average 16.3 years of education), and reported a median annual personal income of $40,000. Study inclusion criteria included self‐reported consumption of at least 7 drinks per week for females and 14 drinks per week for males, ensuring NIAAA criteria for heavy drinking were met by all participants (NIAAA 2011, 2025b). In this study, average consumption was measured as the number of standard drinks consumed, each containing approximately 14 g of alcohol, for example, 12 oz of beer, 5 oz of wine, or 1.5 oz of distilled spirits (NIAAA 2011, 2024). Participants were eligible only if they reported consuming alcohol on at least 50% of days during the 90 days prior to study enrollment, as assessed via the Timeline Followback (Sobell et al. 1996), and reported maintaining a similar drinking pattern for at least 3 years. In an effort to study the effect of alcohol abstinence in as much isolation as possible in a real‐world sample and setting, exclusion criteria included current or historical AUD diagnosis, binge drinking episodes (defined as four or more drinks for females and five or more drinks for males consumed in about 2 h or less (NIAAA 2025b; NIAAA 2011)) occurring more frequently than once monthly on average, regular morning drinking or “eye openers,” or use of other substances (including consumption of more than 500 mg of caffeine daily on average, smoking more than 1.5 packs of cigarettes or equivalent nicotine use daily on average, and use of methamphetamine, cocaine, marijuana, amphetamines, opiates, and benzodiazepines, assessed using saliva drug screenings). Following the imposed period of abstinence (see Section 2.2), potential withdrawal symptoms were assessed via the Clinical Institute Withdrawal Assessment of Alcohol revised (CIWA‐Ar) (Sullivan et al. 1989). No participants reported experiencing more than mild withdrawal symptoms in this sample. Further exclusion criteria ensured participant safety during MR scanning: participants with a history of head injury resulting in loss of consciousness were excluded due to potential history of brain injury, and participants with any ferromagnetic implants, such as cardiac pacemakers, cerebral aneurysm clips, metal shards or shrapnel, or permanent make‐up or tattoos on the face were deemed ineligible for participation.

This sample of habitual heavy drinkers allowed for the examination of fMRI differences across drinking states within and between individuals, namely allowing for the precise measure of abstinence effects while accounting for effects such as brain development (minimum age requirement of 24 years) or brain aging (upper cut‐off of 60 years of age), concurrent substance use, or potential problematic behavioral patterns associated with their alcohol use (as could be indicated by behaviors such as regular morning drinking). The lower age cutoff ensured all participants met the legal drinking age in the United States (21 years of age) and allowed for 3 years of participants' maintained drinking patterns, making 24 years of age the youngest eligible age for participation in this study. The upper age limit was selected to avoid aging effects that occur both in the brain and with alcohol consumption while recognizing that an ageing‐specific study is appropriate and likely needed in this area. Rather than studying individuals with AUD, this study focused on habitual heavy drinkers due to the known association between heavy alcohol use and comorbid disease and mortality, including increased risk of ischemic cardiovascular disease, cancers, diabetes, and traumatic injury (Chikritzhs et al. 2015; Griswold et al. 2018; Rehm 2011; Rumgay et al. 2021). However, it should be noted that participants in this study were not assessed by a clinician or licensed addiction specialist. The study team attempted to ensure the recruitment of non‐AUD alcohol consumers by excluding individuals with a known AUD diagnosis and potential participants who indicated experiencing two or more potential AUD symptoms, as assessed by DSM‐5 screening criteria (APA 2013; Grant et al. 2015; WHO 2001). Participants in this study did report moderate Alcohol Use Disorder Identification Test (AUDIT) scores, but no measures of alcohol consumption quantity or frequency, which are assessed by the AUDIT, are specified in order to meet diagnostic criteria (Veach and Moro 2018). AUDIT scores were not used to determine eligibility in this sample but did confirm the recruitment of moderate‐to‐heavy drinkers.

2.2. Study Design

All study procedures were scheduled to avoid any individual major life stressors or periods of atypical routine. The full study protocol consisted of four in‐person study visits: a screening visit to ensure participants met eligibility requirements, a behavioral baseline visit to collect surveys and cognitive testing, and two MRI sessions scheduled in a sex‐stratified random order. Only data collected during the MRI sessions are presented here. During the screening visit, all participants provided written informed consent, including consent for MRI scanning, as approved by the Wake Forest Health Sciences IRB.

Prior to each MRI scan visit, participants completed two 3‐day experimental periods, one during which they were instructed to drink according to their typical routine, and one during which they were instructed to abstain from all alcohol. The order of these periods was randomized between participants. Both periods were scheduled on days when participants typically consumed alcohol. Compliance with the abstinence period was assessed with a BACtrack C6 mini breathalyzer, with which participants were required to self‐administer breath tests as a measure of blood alcohol content (BAC) at random intervals throughout the scheduled abstinence days. Lack of compliance with this scheduling led to study exclusion. This occurred for one participant. Adherence to typical drinking patterns was assessed via self‐report prior to MRI scanning following the typical drinking period. Participants underwent MRI scanning at the end of each three‐day experimental period. MRI scanning was scheduled on the fourth day of each experimental period and scheduled for each participant during the time of day when they would typically consume alcohol. The brief 3‐day period of alcohol abstinence was anticipated to induce abstinence‐driven changes in brain function, as periods of acute withdrawal following abstinence can last for periods ranging from hours to days, with behavioral effects such as changes in reward thresholds remaining elevated for up to a week (Koob 2013; Koob and Volkow 2010), and ASAM Clinical Practice Guidelines recommend assessing and monitoring patients experiencing alcohol withdrawal for 5 days following the cessation or reduction of alcohol use (Alvanzo et al. 2020).

During each scan session, structural image acquisition was followed by functional image acquisition during the resting state (eyes open, viewing a fixation cross). Participants also completed neutral and alcohol imagery tasks, which are not reported on here. Given that the majority of brain connectivity literature associated with alcohol use uses resting state data, this analysis focused on resting state data for comparative purposes. Scanning was scheduled during the time of day when participants would typically be drinking (usually 5–8 p.m.) during two consecutive weeks on the same day of the week, when possible. A minimum washout period of 6 days was required between scanning sessions to ensure no carryover effects between the typical drinking and abstinence states. At the start of each MRI study visit, participants were screened for MR compatibility and safety (e.g., ensuring they had no ferrous material embedded in their body) and given thorough instructions regarding their experience in the MRI scanner.

2.3. MRI Data Acquisition

MRI data were collected on a 3 T Siemens Skyra scanner, using a 32‐channel head coil and a rear‐projection screen for task image presentation. The imaging protocol included a T1‐weighted structural scan acquired in the sagittal plane using a single‐shot 3D MPRAGE GRAPPA2 sequence (TR = 2.3 s, TE = 2.99 ms, 192 slices) and blood‐oxygen level dependent (BOLD)‐weighted functional scans acquired in the transverse plane using an echo‐planar sequence (4 × 4 × 5 mm, 7 min acquisition, TR = 2.0 s, TE = 25 ms, flip angle 180°, 35 slices per volume, 177 volumes). Participants viewed a fixation cross during resting state scanning (eyes open).

2.4. Preprocessing

Structural MR images were initially processed by implementing tissue segmentation using a unified segmentation algorithm in SPM12 software (http://www.fil.ion.ucl.uk/spm) (Ashburner and Friston 2005). The gray and white matter segmented images were summed, and voxels with ≥ 0.5 probability value were retained as a mask of brain tissue of interest, with voxels of non‐brain tissue and cerebrospinal fluid (CSF) excluded. Any misclassified voxels in the gray and white matter mask were manually corrected. The corrected mask was applied to the structural image, which was then warped to the Montreal Neurological Institute (MNI) template using Advanced Normalization Tools (ANTS), and deformation maps were derived (Avants et al. 2011).

Functional image processing began by discarding the first 10 volumes (20 s) to allow the MR signal to achieve equilibrium. Functional images were then slice‐time corrected and realigned to the series' first volume. One participant was removed from data analysis due to having head motion that exceeded the voxel dimensions. We further verified that the mean framewise displacement (FD) for all remaining participants did not exceed 0.6. We also confirmed that mean framewise displacement did not differ significantly between the two conditions (Paired t‐test: t = 0.272, p = 0.787; see Table S1). A band‐pass filter (0.009–0.08 Hz) was applied to remove the scanner drift and high‐frequency physiological noise such as cardiac and respiratory fluctuations (Powers et al. 2012). Regression was used to remove the mean signals for gray matter, white matter, and CSF and signal associated with 6 motion parameters. We removed an additional five time points from the beginning and end of the signal to correct for bandpass filtering artifacts. When testing the removal of only three time points, although we observed condition differences in state transitions and occupancy time, more participants remaining in a single state for the entire scan suggest a weaker fit of the HSMM model. Thus, we chose to remove five time points for improved model accuracy.

To extract fMRI signal of each brain region, we utilized the Schaefer‐2018‐17‐atlas, comprised of 100 functionally derived cortical regions (Schaefer et al. 2018). Functional images were first co‐registered to each participant's structural T1 image, which was then normalized to the MNI standard space using deformation maps calculated with ANTs (Avants et al. 2011). The atlas was resliced to match the resolution of the normalized functional data in MNI space, and the mean fMRI signal was extracted from each atlas region for every participant.

Reliably fitting the HSMM on a large number of brain regions requires large amounts of data (i.e., many participants and/or long scans). Since our sample size was small, we could not fit the HSMM on all 100 Schaefer atlas regions. Therefore, we focused on 51 regions of the DMN, SN, and CEN that together define the triple network (Menon 2011). The list of these regions and their network classification is provided in Table S2.

2.5. Hidden Semi‐Markov Modeling

Compared with the sliding‐window method, one of the most widely used approaches for identifying brain states, HSMM provides several key advantages. The sliding‐window method requires setting multiple parameters, such as the window function, length, and overlap; however, selecting appropriate values is challenging because there is no ground truth in fMRI data. Sliding‐window methods also struggle to detect sudden changes in connectivity. In comparison, HSMM models the states directly from the time series and thereby avoids the challenges of windowing. This also enables HSMM to assign a brain state to each time point and detect sudden changes in functional connectivity.

The goal of the HSMM is to identify the hidden states ($S$) in the fMRI data ($Y_t$) associated with each timepoint (t) across all the brain regions (p). In other words, for all timepoints of $Y_1,Y_2,\dots ,Y_n$, it estimates their corresponding states $S_1,S_2,\dots ,S_n$ (Figure 1). Each state is defined by a connectivity matrix and a mean vector associated with the mean fMRI signal at all brain regions. The key assumption of this approach is that each $Y_t$ follows a multivariate Gaussian distribution ($Y_t~N\left({\mu }_{s=k},{\sum }_{s=k}\right)$), where $s=k$ represents the current state of $Y_t$. In this formulation, ${\mu }_s$ represents states' mean signal/mean activation, and ${\sum }_{s=k}$ represents the states' covariance/correlation matrix of the p regions. The key advantage of HSMM over the hidden Markov model (HMM) is its ability to model the sojourn time distribution (dashed line in Figure 1) explicitly. Sojourn time refers to the number of consecutive timepoints a state is present before transitioning to another state. In HMMs, the sojourn time is implicitly assumed to follow a geometric distribution, which biases toward shorter intervals of timepoints. HSMMs address this limitation by allowing the hidden state process to follow a semi‐Markov chain process, which leads to greater flexibility in modeling the duration of brain states. In summary, the HSMM uses fMRI time series as input and estimates states' covariance matrices (${\mathrm{\Sigma }}_{1:\mathrm{k}}$), mean vectors (${\mu }_{1:k}$), transition probabilities (P), and sojourn distributions ($d_{1:k}\left(u\right)$) using maximum log likelihood, which for one participant can be formulated as:

$$l^{*}\left({\mu }_{1:k},{\mathrm{\Sigma }}_{1:k},P,\mathrm{\pi },d_{1:k}\left(u\right)\right)=\mathit{\text{logf}}\left(\tilde{s},\tilde{y}\right)=\mathit{\text{logf}}\left(\tilde{y}|\tilde{s}\right)+\log \left(\tilde{s}\right),$$

where $f\left(\tilde{y}|\tilde{s}\right)$ and $f\left(\tilde{s}\right)$ denote the probability of having observation $\tilde{y}$ given $\tilde{s}$ and the probability of the state sequence. We used the MIND‐Map toolbox (Khodaei et al. 2024) to perform the HSMM analysis using an equal partitioning approach for initialization of states' covariance matrices and mean vectors and K‐Smooth Non‐Parametric sojourn distribution. The equal partitioning method divides the time series into k equal segments (where k is the number of states) and then estimates the covariance matrix and mean vector for each segment. These are used as the initial state covariance matrices and mean vectors in the HSMM model.

FIGURE 1.

The adapted HSMM model for fMRI data. The participant number is denoted by i. The observations are the fMRI signal for all brain regions (top) at each timepoint. The latent states are covariance matrices associated with the connectivity of the regions at each state. Sojourn times are illustrated with dashed lines, and red arrows show the transition from one state to another.

2.6. Maximum Likelihood Estimation

The MIND‐Map toolbox utilizes the mhsmm R package (O'Connell and Højsgaard 2011) to estimate the HSMM parameters, including states' mean vectors and covariance matrices, sojourn time distributions, and state transition probabilities. A key advantage of this package is its ability to perform inference for multiple participants simultaneously. To achieve this, the toolbox concatenates the time series data from all participants, creating a single time series where the total number of rows equals the sum of all scans time points, and the number of columns corresponds to the number of regions (51 in our dataset). The mhsmm package then applies HSMM to the entire dataset, estimating the parameters for the full sample using the Expectation–Maximization algorithm (Dempster et al. 1977). Finally, the toolbox uses all estimated parameters and estimates each participant's most likely state sequence using the Viterbi Algorithm (Forney 1973). This algorithm utilizes the fMRI time series along with the model estimates, including the states' mean vectors, covariance matrices, transition probabilities, sojourn densities, and initial state probabilities, to estimate the most probable state sequence for each participant (Shappell et al. 2019).

2.7. Number of States

HSMM requires user specification of the number of states. This is addressed in the MIND‐map toolbox by running the model with different numbers of states and then selecting the number of states that has the highest minimum distance between its state pairs. The toolbox uses Euclidean distance to estimate the distance between the state's functional connectivity matrices. The underlying idea is that genuine brain states should be maximally distinct from one another, whereas spurious states tend to closely resemble (and have low distance from) genuine states. We applied HSMM to the entire sample using a range of 3–7 states and selected the model with the largest minimum Euclidean distance between its states. Fitting the HSMM on more states increases the number of parameters that must be estimated, meaning that more data is needed for reliable results. Given our relatively small sample size, we limited our analysis to a maximum of 7 states. Although this approach has shown its capability for identifying the optimum number of states on simulated data, it requires a successful HSMM model fit. Therefore, we also investigated the individual state sequences to ensure they did not follow the initial sequence patterns or show clustering of participants, which would indicate a poor fit due to insufficient data.

2.8. Drinking Condition Brain Dynamics Comparison

To compare the dynamics between the two drinking conditions (typical drinking and abstinence), the HSMM was first fit on all 76 scans (2 conditions for n = 38 participants) to identify a common set of states between the conditions. These states were then used as fixed parameters in two distinct HSMMs (one per condition) which revealed characteristics of dynamics for the two conditions separately. This approach ensures that the model does not estimate new states between conditions but instead focuses on extracting the dynamics of states within each condition, which enables a direct comparison of their state‐related dynamics.

Using permutation testing for each state, we compared participant occupancy time, percentage of the number of transitions, and empirical sojourn distributions between the two conditions. Occupancy time represents the total number of time points a participant spends in each state across the full duration of the scan. The number of transitions percentage reflects the probability of a participant switching between states. While sojourn time refers to the consecutive time points a participant remains in a state before transitioning to another, we also identified full sojourn distributions that represent the distribution of sojourn times for each state. Finally, we compared the transition matrices, which represent the probabilities of switching between each pair of states given that a state switch is occurring. To examine differences in transition probability matrices, one transition matrix was subtracted from the other and visualized using the Networkx Python package (Hagberg et al. 2008).

Statistical permutation testing was also performed using the MIND‐Map toolbox. We employed paired permutation testing with 10,000 permuted samples to compare each participant's average time in and transit into the states across the two conditions. The permutation groups were created by flipping the sign of the difference between the two conditions. For the empirical sojourn distributions comparison, we began by counting the number of consecutive timepoints each participant spent in each state. These counts were then aggregated for each state to calculate a distribution for each condition. The toolbox used a permutation test with 500 permuted samples to compare the Kullback–Leibler divergence (Kullback and Leibler 1951) between the sojourn distributions of each state across the two conditions.

2.9. States Visualization and Characterization

To investigate the characteristics of brain states, we generated two brain maps for each state. First, we produced brain mappings of the modular organization of each state, representing the distinct network structures associated with different brain states. This was done by performing modularity analysis using the Brain Connectivity Toolbox (BCT) (Rubinov and Sporns 2010). The modularity analysis identifies brain network communities that are highly intraconnected but weakly connected to nodes in other modules (Newman and Girvan 2004). We applied Newman spectral community detection with a lambda value of 1.2 to identify the modules (Newman 2013) of the positive correlation matrices associated with the states. Our analysis with a lambda value of 1 produced only a small number of modules for each state. To obtain more distinct community structures, we increased the lambda parameter to 1.25. This allowed the states to split into additional modules and provided a clearer view of the organization of the three subnetworks within the triple network. Because the same lambda was applied to all states, their community structures remained comparable. Nodes corresponding to each module were assigned the same color and mapped onto the brain for visualization. Second, the HSMM‐estimated mean activation vector of each state was mapped to brain space to illustrate the average normalized BOLD signal during each state when compared to other states. These mean activation vectors were also used to calculate the average normalized BOLD signal (normalized activation) within each of the three subnetworks. In addition, we extracted the average normalized BOLD signal of the anterior medial and posterior medial DMN, as studies have shown they can play unique roles in addiction (Zhang and Volkow 2019).

In addition to brain maps, we estimated a set of topological network measures for each state's correlation matrix after applying a threshold of zero, retaining only positive edges, using the BCT toolbox (Rubinov and Sporns 2010), including global efficiency (Latora and Marchiori 2001), shortest path (Bassett and Bullmore 2006), clustering coefficient (Watts 1999), local efficiency (Latora and Marchiori 2001), assortativity (Newman and Newman 2002), modularity (Girvan and Newman 2002; Newman 2006, 2013), strength, and network density. These network measures provide information about local and global information flow, preferential attachments, community structures, and network over all density of connections.

3. Results

3.1. Participants

Participant demographic information is shown in Table 1. Participants averaged 41 (±11) years of age and reported regularly consuming alcohol for an average of 25 (±12) years. Overall, participants consumed an average of 17 (±5) standard alcoholic drinks per week, or 2.9 (±0.9) per day, with male participants consuming significantly more [23 (±6) drinks per week or 3.7 (±1.2) per day] than female participants [13 (±5) drinks per week or 2.4 (±0.7) per day; t = 5.22, p < 0.001], as designed by the study procedures. Participant identification of race, age, and total number of years drinking were not significantly different between male and female participants. Participants reported an average AUDIT score of 18.4 (±2.9), which is indicative of hazardous drinking, and the World Health Organization indicates “brief counseling and continued monitoring” should be implemented (p. 21) (WHO 2001).

TABLE 1.

Participant demographics.

Demographic	Full sample	Males	Females
N	39	17 (44%)	22 (56%)
Age	41.2 (±11.7)	38.8 (±9.8)	43 (±13)
White	35 (90%)	15 (88%)	20 (91%)
African American or Black	4 (10%)	2 (12%)	2 (9%)
Total Years of Education	16.3 (±1.5)	16.2 (±1.5)	16.2 (±1.5)
Annual Personal Income	$40 k median (±$35 k)	$50 k median (±$30 k)	$40 k median (±$35 k)
Total Years Drinking	24.7 (±11.8)	21.9 (±9.2)	26.9 (±13.2)
Average Consumption Per Week	17 (±5)	23 (±6)*	13 (±5)*

^*Tested via t‐test, p < 0.001.

3.2. Condition Comparison

We fit the HSMM to the data using six states, as determined by the state's minimum distance plot from the MIND‐Map toolbox (Figure S1). The correlation matrices for the six brain states and the participants' state sequence are presented in Figures S2 and S3, respectively. For each state, we compared the occupancy time, the number of transitions, and the sojourn time between the typical drinking and abstinence conditions. Table 2 reports the p values for these comparisons. The occupancy time for State 4 was significantly different between the two conditions, with less time spent in this state during abstinence compared to typical drinking (Figure 2A). During the abstinence condition, participants spent on average more time in States 1, 2, and 5. When the percentage of the number of transitions to each state was compared, State 3 differed significantly between the two conditions, with a decrease in transitions to this state in abstinence (Figure 2B). The increased transition to State 3 during the typical drinking condition is also evident in the circular graph of transition probabilities between abstinence and typical drinking conditions (Figure 3). During the typical drinking condition, transitions to State 3 were more likely from States 2, 4, and 6. During abstinence, there were more transitions to other states (1, 2, 5, and 6), particularly to States 1 and 5. The drinking conditions also had different sojourn distributions for State 1, with participants having a lower probability of a longer sojourn time during abstinence compared to typical drinking. The sojourn distributions for each state are shown in Figure 4.

TABLE 2.

Results from permutation testing comparing the dynamics between the typical drinking and abstinence conditions.

States number	Occupancy time p	Number of transitions p	Sojourn distribution p
State 1	0.472	0.099	0.002*
State 2	0.093	0.720	0.252
State 3	0.464	0.005*	0.8
State 4	0.023*	0.306	0.440
State 5	0.148	0.165	0.9
State 6	0.790	0.848	0.418

Note: Asterisks indicate statistically significant condition differences (p < 0.05).

FIGURE 2.

(A) Occupancy time for each state during the two conditions. (B) Percentage of the number of transitions to each state during the two conditions. The error illustrates the standard error of the means in both figures. Asterisks indicate states with significant condition differences.

FIGURE 3.

Differences in transition probabilities between states under the two conditions. Solid purple indicates the top five strongest transitions during the typical drinking condition, whereas solid green represents the top five strongest transitions during abstinence. The thickness of the edges reflects the scaled strength of the transition probability. The size of the blue circles shows the scaled occupancy time during the two conditions for each state. During the typical drinking condition, transitions were higher from State 1 to States 3 and 5, and from States 2, 4, and 6 to State 3. Conversely, during abstinence, transitions were higher from State 1 to State 2, from States 2, 3, and 4 to State 1, and from State 4 to State 5. The difference in transition probability matrices of the two conditions is presented in Table S3.

FIGURE 4.

Sojourn distribution of each state during each condition.

3.3. State Characteristics

To investigate the unique characteristics of the states, we analyzed their modular structure, average normalized BOLD signal, and topological measures, with a particular focus on States 1, 3, and 4, which exhibited significant dynamic differences between the conditions. We also investigated other states, including 2 and 5. Although they did not show significant condition differences, States 2 and 5 were among the main states participants transitioned to and spent more time in during abstinence. However, the difference in dynamics between the two conditions was not significant, possibly due to the effect being distributed across States 1, 2, 5, and 6. Each state's extracted topological network measures are presented in Table S4.

3.3.1. States' Modularity Maps

The modularity map for each state is presented in Figure 5. State 1 was divided into four modules, with the precuneus and part of the middle posterior cingulate and retrosplenial divided into a single module, while the remaining parts of the DMN, CEN, and SN were clearly pronounced and followed their expected structure. State 4 showed similar isolation of the precuneus into a single module; however, unlike State 1, it demonstrated stronger connections between the CEN and SN. State 3 also exhibited a modular structure that fused and divided the expected subnetworks. The anterior and posterior DMN (aDMN and pDMN, respectively) were divided into separate modules, while regions of the SN largely merged with regions of the CEN into a single module.

FIGURE 5.

Modularity maps associated with each state. The states marked with asterisks indicate significant condition differences in state dynamics.

Although no significant differences were observed between drinking conditions in States 2, 5, and 6, participants did occupy these states during both their typical drinking and abstinence periods. States 2 and 5 also illustrated unique modular structures. State 2 was divided into three modules, with the DMN, CEN, and SN each forming a distinct module, with regions following their expected network classifications. State 5, on the other hand, was divided into five modules, making it the state with the highest number of modules and maintaining the least of the expected subnetwork structure. Four modules were identified in State 6, with the precuneus and posterior cingulate organized in one module, other regions of the DMN largely remaining interconnected in another module, and regions of the CEN and SN dividing into two modules.

3.3.2. States' Average of the Normalized BOLD Signal (Normalized Activation)

We evaluated the average of the normalized BOLD signal within each subnetwork for each state. Figure 6A illustrates the average of the BOLD signal of each subnetwork across each brain state relative to other states. Although less pronounced than in State 6, State 4 showed DMN dominance coupled with reduced SN activity. In contrast, State 3 exhibited the highest SN activation, along with notably low DMN activity. State 1 similarly presented low DMN activation but is not dominated by either the CEN or SN. Additionally, States 2 and 5 revealed distinct patterns where State 2 had activation levels near the overall mean, whereas State 5 was dominated by CEN activity. Lastly, State 6 was characterized by the highest DMN activation and the lowest CEN activity compared to the other states. Figure S4 indicates the average BOLD signal in all triple network regions for each state. Additionally, based on our modularity analysis and reported alterations in the anterior medial DMN (amDMN) and posterior medial DMN (pmDMN) in the literature associated with alcohol use (Zhang and Volkow 2019), we estimated the average BOLD signal in the amDMN and pmDMN. As shown in Figure 6B, amDMN exhibited lower than average activation in States 1, 2, 3, and 5, while it demonstrated higher than average activation in States 4 and 6. On the other hand, pmDMN had lower than average activation in States 1, 3, and 4 but showed higher‐than‐average activation in States 5 and 6.

FIGURE 6.

(A) Mean of the normalized bold signal in the triple networks for each state. (B) Comparison of the average of the normalized BOLD signal in the anterior medial DMN (medial prefrontal cortex) and the posterior medial DMN (posterior cingulate and precuneus) across the six states.

4. Discussion

We utilized a Hidden Semi‐Markov Model (HSMM) to extract brain states from the resting‐state fMRI time series of moderate‐to‐heavy drinkers collected during a period of typical drinking and a period of alcohol abstinence. We then compared the dynamics of the states between the two conditions, including comparisons of the state occupancy time, percentage of transitions to states, and state sojourn time distributions, focusing on the brain networks of the Triple Network Model. Our study revealed six distinct brain states, with significant differences in measures of brain network dynamics between the two drinking conditions in three of the six states, indicating that alcohol abstinence influences the temporal characteristics of these specific brain states. More specifically, our analysis revealed differences in the occupancy time of State 4, the number of transitions to State 3, and the sojourn time distribution for State 1. To better understand the unique characteristics of these states, we further investigated their mean normalized activity and modularity maps.

4.1. The Triple Networks

The Triple Network Model posits that aberrant functional organization of the SN, CEN, and DMN may underlie a variety of psychopathologies (Menon 2018), and previous work has suggested that alcohol use and abstinence are linked to dysfunction within and between these systems, but little work has examined their temporal dynamics. By modeling how individuals transition among network configurations, our results extend this literature from static to dynamic analyses, revealing how abstinence alters the flexibility and stability of Triple Network interactions.

Consistent with prior findings of disrupted DMN connectivity in heavy drinkers and abstinence individuals (Fang et al. 2021; Muller‐Oehring et al. 2015; Shokri‐Kojori et al. 2017; Vergara et al. 2017; Weber et al. 2014; Zhu et al. 2017), we observed altered engagement in a DMN‐dominant state (State 4) across conditions. The shorter occupancy of this internally focused state during abstinence parallels evidence that withdrawal is associated with a decrease in DMN connectivity in heavy drinkers during resting state, with more substantial effects associated with higher alcohol craving (Fang et al. 2021), and weaker within‐network connections of the DMN in abstinent alcohol drinkers, with weaker connections associated with poor cognitive performance and increased mood disturbances (Muller‐Oehring et al. 2015). This is also consistent with our other dynamic connectivity study, which found a negative association between the occupancy time of a high‐DMN state and future drinking (McIntyre et al. 2025). Although the current DMN‐dominant state did not show concurrent salience‐network activation as that state did, both findings highlight the importance of DMN‐driven states in shaping vulnerability to alcohol use. Similarly, differences in transitions involving the SN‐dominant state (State 3) align with previous reports of changes in the structure and function of the SN reported in association with addiction (Cushnie et al. 2023) and cognitive impairment in AUD associations with the gray matter structure of regions of the SN (Galandra et al. 2018). This SN‐dominant state also exhibited strong SN–CEN coupling. This finding aligns with the meta‐analytic review by Wilcox et al., which identified reduced SN–CEN connectivity across substance‐use disorders and proposed this network pattern as a potential treatment target for future interventions (Wilcox et al. 2019). Together, these results support the idea that alcohol consumption may transiently stabilize network dynamics that are otherwise dysregulated during abstinence among habitual moderate‐to‐heavy drinkers.

Our findings also contribute to the growing literature on functional differences in the anterior DMN (aDMN) and posterior DMN (pDMN) associated with differing cognitive functions and behaviors (Andrews‐Hanna et al. 2014; Xu et al. 2016). The aDMN has also been found to be more activated while participants reflect on their present mental state, while the pDMN is more activated when they imagine scenarios for their future selves (Xu et al. 2016). Rumination and anxiety, both components of withdrawal states and subsequent preoccupation with substance use (Koob and Volkow 2010), are related to different patterns of EEG activity in the aDMN and pDMN (Ho et al. 2024). In a review study, Zhang and Volkow have reported different behaviors in the functional connectivity of the anterior medial DMN (amDMN) and posterior medial DMN (pmDMN) in association with substance use disorders, where participants with addiction show decreased functional connectivity in the amDMN, while the pmDMN tends to exhibit increased connectivity in individuals with addiction (Zhang and Volkow 2019). The observed dissociation of these subnetworks across states, and in particular the high activation of the aDMN in State 4 and the low activity of the pDMN in States 1 and 3, suggests that abstinence selectively disrupts the balance between internally and externally oriented thought. These dynamic results provide temporal evidence consistent with previous static connectivity findings linking altered DMN‐CEN‐SN coordination to craving, rumination, and cognitive control deficits during abstinence (Koob and Volkow 2010; Marsden 2022; Mayhugh et al. 2016).

In support of the disruption of the three networks that comprise the Triple Network Model in alcohol consumers, Muller and Meyerhoff have reported on a reconfiguration of functional network communities in successfully abstinent AUD patients, which they associated with purposeful, top‐down mind wandering, thought to serve as a deliberate alternative to rumination on cravings (Muller and Meyerhoff 2021). The reorganization included reallocation of the dorsomedial prefrontal cortex from the DMN to a community comprised of regions of the CEN and SN, and this reconfiguration was not observed in AUD patients who released (Muller and Meyerhoff 2021). This prior work resonates with our observation that abstinence was associated with more frequent but less stable transitions into a segregated, transitory state (State 1). Such segregation is broadly consistent with the Triple Network Model, in which the DMN, CEN, and SN form distinct but interacting modules during rest. However, the brief dwell time within this state suggests reduced temporal stability of network organization. This suggests that even in moderate‐to‐heavy drinkers, temporary abstinence may induce instability in cross‐network coordination similar to patterns seen in clinical AUD.

The behavioral associations of the Triple Network Model and the underlying subnetworks allow for neurobiologically and behaviorally relevant explanations of observed brain connectivity. However, behavioral correlates were not assessed in this study, and therefore, the relationship between the connectivity of these networks and potential behavioral results cannot be proven from this analysis. Nevertheless, in an effort to explore potential effects resulting from the patterns of brain connectivity identified in this analysis, potential behavioral associations are described with each discussion of major results. The states with no significant condition differences are described in detail in Supporting Information.

4.2. State 1—An Interruptive, Transitive State

State 1 showed the lowest activity in the DMN of any of the observed states, while the CEN and SN exhibited higher‐than‐average activity. Correspondingly, activity in both the aDMN and pDMN was lower than average in State 1. The modularity analysis revealed that the precuneus and the middle posterior cingulate parts of the CEN, and retrosplenial as well as parahippocampal regions of the DMN were highly interconnected in this state, while the remainder of the regions belonging to the DMN and CEN, as well as regions of the SN, showed modular structures that highly resembled the expected network organization of each subnetwork, suggesting a high degree of interconnection within each subnetwork, with fewer connections between them. While participants did not differ in their occupancy time in State 1 when comparing typical drinking to abstinence, there was a significant difference in the sojourn distributions, such that participants had a higher probability of spending a lower number of continuous timepoints in State 1 during abstinence. There was a corresponding trend toward a greater number of transitions into State 1 during abstinence compared to typical drinking, suggesting that participants were potentially more likely to transition into State 1 during abstinence, but then occupied the state for a shorter time during abstinence than during typical drinking. During abstinence, participants showed a high probability of transition from States 2, 3, and 4 to State 1.

The dynamics of State 1 may relate to an interruption of processes occurring in other states, an effect that was more prevalent in the abstinence condition than following typical drinking. Participants transitioned from States 3 and 4, which also showed significant temporal dynamics between drinking conditions, into State 1 during the abstinence condition, moving from states of integrated or reconfigured network organization, with high SN and DMN activation, respectively, into a State 1 with functional segregation, as may be expected from the Triple Network Model, between the subnetworks. During abstinence, participants spent fewer consecutive timepoints in State 1, and on average transitioned into State 1 more frequently, suggesting this interruption of high SN or DMN activity in States 3 or 4 by occupancy of State 1 occurs more frequently and more rapidly following abstinence, perhaps capturing abstinence‐specific alterations in the dynamic coordination of the triple networks and the transition between SN and DMN dominant states. The shorter sojourn times observed for State 1 during abstinence may potentially indicate reduced stability in this particular network configuration compared to periods of typical alcohol consumption. While this analytic method cannot allow us to ascertain the underlying causes of differences in temporal dynamics, a potential explanation for this reduced stability may be the neurotransmitter imbalances that occur following the cessation of alcohol use, leading to neuronal hyperactivity (Mattle et al. 2022). This neurotransmitter imbalance and hyperactivity could result in less stable state occupancy.

4.3. State 3—An Externally‐Oriented State

State 3 also showed low activity in the DMN, while also being the state with the highest mean activation in the SN. The modular structure of State 3 showed only three communities, but rather than aligning with the three expected subnetworks, the SN and CEN regions were merged into one module, while the aDMN and pDMN were split into two modules. This suggests that while the activity in regions of the SN was substantially greater than the activity in regions of the CEN, these regions were interconnected, indicating segregated information sharing. Additionally, activity in the pmDMN may be predominantly driving the overall low activation of the DMN, as activity in the pmDMN was relatively much lower than in the amDMN. During the typical drinking condition, participants transitioned into State 3 significantly more frequently than they did following abstinence, including more transitions from States 2, 4, and 6 into State 3. Participants did not differ in their occupancy time or sojourn time distributions for State 3.

The high SN activation in State 3 suggests this may potentially be a state of selecting salient stimuli and directing corresponding behavior, and in particular, there seems to be a preference for externally focused or goal‐directed thought, as regions of the SN were highly interconnected with regions of the CEN. Correspondingly, DMN activation was much lower than average in State 3, perhaps indicating a lack of cognitive resources being directed toward internal or self‐referential thought. Interestingly, the low DMN activation may be driven by the pmDMN, which segregated from the amDMN, to the extent that the two portions became separate modules. Participants occupying State 3 may be unlikely to consider their future actions, or retrieve memories regarding present stimuli, instead directing their attention to more salient, external stimuli. Participants transitioned into this SN and CEN integrated state more frequently during typical drinking than during abstinence, suggesting perhaps this state could play a role in how individuals interact with external cues, such as alcohol‐related cues, while they have been drinking, compared to when they have been abstaining. Perhaps following abstinence, moderate‐to‐heavy drinkers are more likely to remember salient cues and ruminate on their desire to drink, while during the typical drinking condition, higher transition probability into State 3 allows them to remain attentive to external stimuli. Participants showed a high probability of transitioning from State 4, the state of internally directed focus, as well as States 2 and 6 into State 3 during the typical drinking condition, whereas they are not highly likely to transition into State 3 during the abstinence condition.

4.4. State 4— an Internally Focused State

State 4 showed the second highest mean activation of the DMN (following State 6), with relatively high activation in the amDMN and near average activation in the pmDMN. Despite this difference in activation, most regions of the DMN were classified into a single module, denoting a high degree of interconnectedness, excluding the retrosplenial and parahippocampal regions, which isolated into a unique module, as was seen in State 1. The modular organization of the regions of the CEN and SN was more complicated, forming two modules, with regions from each subnetwork belonging to each module. Therefore, some regions from each subnetwork were more connected to regions from the other subnetwork than they were to other regions in their own subnetwork, while other regions were most highly connected to other regions of the same subnetwork. This suggests a more complex integration of communication between these subnetworks than was observed in other states. Although State 4 showed the highest occupancy time across both drinking conditions, participants exhibited significantly longer occupancy time in State 4 following typical drinking. Sojourn time distributions did not differ between conditions for State 4, but participants did appear to show a greater number of transitions into State 4 on average in the typical drinking condition, although the difference between conditions was not significant. During the typical drinking condition, participants were more likely to transition from State 1 to State 4.

Following both typical drinking and abstinence, participants spent the largest proportion of their scan time in State 4, a state with higher‐than‐average DMN activation, particularly in the amDMN, perhaps indicating State 4 as a state of internal awareness and current self‐reflection. On average, participants transitioned into this state more frequently during typical drinking and spent significantly longer occupancy time, on average, during the typical drinking condition. Following typical drinking, participants were also likely to transition from State 1, a potentially mediating or interrupting state, into State 4, where they were likely to remain for long periods of time. During alcohol withdrawal and following alcohol cue exposure, the resting state functional connectivity of the DMN decreases (Zhang and Volkow 2019). Subsequently, it may be expected that moderate‐to‐heavy drinkers may show less occupancy of a DMN‐related brain state following alcohol abstinence, as abstinence‐related effects may decrease the overall connectivity of the DMN. Furthermore, nicotine‐dependent individuals have shown increased duration of DMN activation following typical nicotine use (Wang et al. 2020), similar to the increased occupancy time in this DMN‐related state observed in this sample following typical alcohol consumption.

5. Limitations

This study, while employing emerging dynamic state‐based analyses of fMRI data in an understudied drinking population, was not without limitations. While we employed the preprocessing order used in earlier studies, applying band‐pass filtering prior to regression, more recent work has suggested the reverse order. This work indicates that performing regression after band‐pass filtering can introduce high‐frequency noise. We investigated this issue in our dataset and found that although such high‐frequency components did appear after regression, their amplitude was small, minimizing their potential impact. Nonetheless, we acknowledge that they may still have some effect on our results. The limited sample size restricted the number of regions that could be included in our modeling, as more regions would require estimating additional parameters. Our model did successfully converge with the regions of the triple network included, but a larger sample size (from either a greater number of participants or longer scan durations) would allow for the inclusion of more brain regions, with the eventual goal of including a full whole‐brain network. The utilized HSMM fit the same six states across data collected during both drinking conditions. This technique was used because it makes it possible to directly compare dynamics between the typical drinking and abstinence conditions, but it also makes it more difficult to detect states that may exist only in one condition or the other. For example, it is plausible that the subjective experience of craving that arises during abstinence may be due to the introduction of a unique brain state that is not typically occupied by participants while they have been consuming alcohol. Additionally, while this analysis has identified significant differences in the temporal dynamics of functional connectivity in this sample, direct associations with behaviors cannot be drawn from this analysis. Future work expanding on this foundation would be strengthened by the inclusion of additional behavioral metrics.

While we have completed extensive analyses in previous work (Khodaei et al. 2024) to validate our approach for detecting the number of states present in the time series for a sample, the true number of brain states that appear in the sample is not known. It is plausible that while six states seem to provide the best fit to the data, the actual number of states that participants entered may be different. As brain network dynamics have not been investigated in this study's population previously, we intended to lay the groundwork for further investigation using resting state data. However, the brain states which participants occupy during tasks may meaningfully differ from states seen at rest. Future work should investigate brain network dynamics in this population during tasks such as delayed discounting and alcohol‐cue presentation. Additionally, expanding the current work into a broader range of alcohol consumers would allow for further exploration of the observed effects, which would allow us to investigate whether they are common to the alcohol drinking experience or unique to moderate‐to‐heavy drinkers. Furthermore, the participants in the study did not meet DSM criteria for an AUD but did report heavy, risky drinking based on AUDIT scores. Of note, these participants were not assessed by a clinician or licensed addiction specialist, and therefore we cannot report with complete certainty that this sample of participants would not be diagnosed if assessed in a clinical setting. This study aimed to examine the drinking and abstinence experience of moderate to heavy drinkers, rather than AUD patients, but this may limit the generalizability of these results to broader categorizations of alcohol consumers. Finally, while sex differences are increasingly of interest in the alcohol literature, we did not analyze any sex differences (beyond drinking quantities) because our sample was not powered to detect sex difference, though they may exist. It will be important to investigate potential sex effects in future work, as well as other potential group differences, by studying a larger, more diverse sample.

6. Conclusions

We investigated changes in the subnetworks of the Triple Network Model in moderate‐to‐heavy drinkers during typical drinking and abstinence using HSMM, aiming to elucidate alterations in the resting‐state connectivity underlying alcohol use and cessation in this population. Our findings indicate that during the typical drinking condition, participants spend on average more time in a state of high DMN activation. More importantly, during the typical drinking condition, participants had a higher probability of transitioning to a state with high activation of the SN, which was strongly connected to the CEN. This connection could facilitate the selection of important external stimuli and the transfer of information to the CEN for higher cognitive processing. In addition, during the abstinence condition, participants transition more frequently into the transitory state but remain for shorter consecutive stretches of time. Overall, this pattern of results suggests more regulated, systematic dynamic transitions between the subnetworks of the Triple Network Model, suggesting a disruption of triple network functioning during alcohol abstinence. Such alteration may provide the foundation for impaired cognitive control that may facilitate relapse risk. Understanding these dynamic network changes forms a critical foundation for developing neuromodulation interventions for heavy alcohol use and may play a role in more thoroughly understanding alcohol misuse and developing clinical interventions.

Funding

This work was supported by the National Institute on Alcohol Abuse and Alcoholism (P50AA026117) National Institute of Biomedical Imaging and Bioengineering (K25EB032903).

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Data S1: hbm70432‐sup‐0001‐Supinfo.docx.

Khodaei, M. , Peterson‐Sockwell H., McIntyre C. C., et al. 2025. “Abstinence Alters Triple Network Dynamics in Moderate‐to‐Heavy Drinkers.” Human Brain Mapping 46, no. 18: e70432. 10.1002/hbm.70432.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.