Flexibility of Brain Networks May Curtail Cognitive Consequences of Poor Sleep

ABSTRACT

Previous research has shown that laboratory‐controlled sleep deprivation leads to cognitive impairments, including low vigilance and deficits in working memory. However, the robustness of sleep effects on behavior and brain dynamics in naturalistic settings remains underexplored. In this study, we investigated the impact of naturalistic, unfettered variations in sleep on behavioral performance and brain network dynamics in 39 healthy adults. Using a dynamic networks approach combined with ordinal regression, we show a significant increase in flexibility, a measure of rapid reconfigurations within the brain modules, with decreasing sleep time, particularly in the fronto‐parietal control network, during a psychomotor vigilance (PVT) and visual working memory (VWM) task. This change in network flexibility was not observed during the resting state. Critically, performance itself did not change as a function of sleep, providing preliminary evidence that brain networks may compensate for having a poor night's sleep by recruiting the necessary resources to complete the task. Additional analysis assessing the regularity of sleep indicates a wider change in flexibility during PVT for irregular sleepers in networks including the limbic system, ventral attention network, and somatomotor system. These results provide new insights into the neural and behavioral correlates of naturalistic sleep modulations.

Studying the effect of naturalistic sleep fluctuation, we found that low sleep related to increased brain network flexibility during tasks, especially in the “control” network, whereas task performance remained unchanged. Flexibility did not change during rest, highlighting task‐specific brain adaptations that compensate for low sleep to maintain behavior.

1. Introduction

Sleep and cognitive performance are intricately linked (Alhola and Polo‐Kantola 2007). Adequate sleep plays a crucial role in maintaining optimal cognitive function, whereas sleep deprivation can significantly impair our mental abilities including memory formation (Steptoe et al. 2006; Walker and Stickgold 2004, 2006), vigilance (Van Dongen et al. 2003), and attention (Lim and Dinges 2008) and can lead to poor decision making (Harrison and Horne 2000) and lower creativity (Wimmer et al. 1992). A large body of sleep literature relies on unnaturally modifying sleep, sometimes with extreme deprivation, to understand the cognitive consequences of sleep and their neural underpinnings. These rigorous, controlled laboratory studies have provided foundational knowledge and have led to well‐established theories of sleep effects on behavior and the brain, but researchers are now beginning to test the robustness of sleep effects on behavior and extend experimentation to more naturalistic contexts in order to provide ecologically valid results (Matusz et al. 2019). Recent studies hint that not only extreme deprivation but short sleep (DiFrancesco et al. 2019) and even sleep fluctuations (Knufinke et al. 2018) can impact behavior and lead to measurable alterations in the large‐scale brain dynamics.

To expand our understanding of the effect of real‐world sleep variations, we investigate the impact of unconstrained, naturalistic fluctuations in sleep on task performance as well as on large‐scale brain dynamics. Across a cohort of individuals, we measured performance on a variety of tasks including the psychomotor vigilance task (PVT), visual working memory task (VWM), and a modular math task (MOD), whereas their brain activity was recorded using functional magnetic resonance imaging (fMRI); and monitored unconstrained sleep using an actigraphy watch for up to 2 weeks prior to the experiment. We captured both instantaneous as well as longer term readouts of an individual's sleep by leveraging two different metrics: total sleep time (TST) the night before the experiment and sleep regularity index (SRI) for the 2 weeks leading to the day of the experiment (Phillips et al. 2017).

Coordinated interplay between different regions and networks within the human brain is foundational to brain function and critical to cognitive processes (Bressler and Menon 2010), as it efficiently integrates and segregates dynamic modules (Sporns and Betzel 2016). To investigate the dynamic reconfiguration of modules supporting effective interactions between distributed regions, researchers have borrowed methods from network science. One such method, so‐called dynamic community detection, has been successfully deployed to investigate temporal changes in networks (Bassett et al. 2013, 2015; Braun et al. 2015; Garcia et al. 2018; Lima Dias Pinto et al. 2022), characterizing the emergence and reconfiguration of network modules with time. Across several studies, network flexibility, the tendency of the brain regions to flexibly change module allegiance, has been associated with task demands during a plethora of cognitive processes including memory performance (Braun et al. 2015; Stehr et al. 2021), learning (Reddy et al. 2018), motor tasks (Monteiro et al. 2019), opinion change (Lima Dias Pinto et al. 2022), and other complex functions (Alavash et al. 2015; Shafiei et al. 2020).

Here, we employ dynamic community detection on fMRI‐derived blood–oxygen level dependent (BOLD) activity, recorded during task performance, to investigate how reconfigurations in the brain dynamics follow the association between sleep and task performance. Specifically, we investigate how brain flexibility, both brain‐wide and within cognitively defined functional networks or communities (Schaefer et al. 2018; Thomas Yeo et al. 2011), changes as a function of sleep time (TST) during different tasks and how the relationship between TST and flexibility is impacted by the regularity in sleep (SRI).

Our results indicate significant robust change in network flexibility as a function of sleep time during task performance such that flexibility of a distributed set of networks including control, limbic, default mode, and subcortical increases in a task‐specific manner with poor sleep. However, we observed a lack of clear impact of sleep time directly on task performance itself. To explain the maintained performance accompanied by an increase in flexibility when sleep is low, we extend the idea of a compensatory mechanism (Cabeza et al. 2018), largely discussed within the literature on aging, as an increase in neural activity to compensate for potential performance decline. This compensation process includes substantial reorganization in neural resource deployment often measured as an increase in neural activity (e.g., Grady et al. 1998; Reuter‐Lorenz et al. 2000; Cabeza et al. 2002) or changes in functional connectivity (e.g., Stanford et al. 2022). We extend this concept beyond the aging brain and postulate that even healthy brains can compensate for fluctuations in a variety of states including potentially transient and nonpathological fatigue states. Our findings suggest that the brain may dynamically reconfigure when facing computational hurdles due to external pressures (i.e., sleep loss), and perhaps even point to the brain's unique resilience to challenges within our everyday lives.

2. Materials and Methods

2.1. Participants

Fifty‐six participants between the ages of 18 and 35 years old (M = 22.21; SD = 2.99; 58% female) were recruited from the greater Santa Barbara area as part of the Cognitive Resilience and Sleep History (CRASH) research study. During an experimental session, participants completed a set of cognitive tasks including a PVT, a VWM task, and a MOD, along with a 5 min resting state, whereas their BOLD responses were recorded using fMRI and behavior was recorded during task performance across trials. Additionally, daily sleep measurements of actigraphy and sleep history were collected starting 2 weeks prior to the session (Figure 1A). During experimental data collection, many participants completed bi‐weekly experimental sessions over the course of 16 weeks, whereas others only participated in one session. In the current analysis, a total of 40 participants (mean age: 22.3 years, SD: 3 years, 24 females) were included who had at least one session's data across all the modalities (i.e., fMRI, behavioral, and actigraphy), and in the current study, we used the first session's data including sleep measurements. One participant was excluded due to being an outlier, having slept for an extreme duration of 15.67 h, resulting in a total of 39 participants. All methods and procedures in the present study were approved by both of the Institutional Review Boards at the University of California, Santa Barbara, and the U.S. Combat Capabilities Development Command Army Research Laboratory and carried out in accordance with this approved research protocol. All participants in this study provided written informed consent to participate in this approved research protocol.

FIGURE 1.

(A) Experiment schematic. (B) Histogram displaying the distribution of total sleep time the night before the session, measured by actigraphy with a cutoff at 6, 7, and 8 h leading to four categories of sleep: (i) < 6 h (with minimum 4.1 h), (ii) 6–7 h, (iii) 7–8 h, (iv) > 8 h (with maximum 11.45 h). The study utilized this literature driven categorization to examine the effects of sleep time. (C) Distributions of behavioral performances across sleep categories for psychomotor vigilance task (PVT) reaction time, visual working memory (VWM) accuracy, and modular math task (MOD) accuracy. We did not observe a direct effect of naturalistic sleep variations or low sleep on task performance. The red point depicts the median value.

2.2. Behavioral Tasks

2.2.1. Psychomotor Vigilance Task

PVT is a classic task to assess vigilance where participants were asked to respond to the appearance of a stimulus as fast as they can (Loh et al. 2004). On average, the participants completed 80 trials, and the performance was assessed by calculating the average response time across trials.

2.2.2. Visual Working Memory Task

In order to examine VWM performance, participants completed a standard VWM task (Luck and Vogel 1997). Participants were presented with squares of different colors for 150 milliseconds. After a delay period of 1180 milliseconds, participants were asked to recall whether the new presentation of colored squares was the same or different relative to the previously presented stimuli. Participants completed a total of 144 trials over the course of three blocks of 48 trials. The task took approximately 5 min to complete. VWM performance was assessed by calculating the accuracy of total correct responses divided by the total number of trials.

2.2.3. Modular Math Task

For MOD, participants were presented with a modular arithmetic expression with a remainder after dividing two integers and were asked if the expression was correct (Mattarella‐Micke et al. 2011). Each participant completed a total of 40 trials where difficulty levels were randomized. The task took approximately 2 min to complete. Performance was assessed by calculating the accuracy of total correct responses divided by the total number of trials.

2.3. Naturalistic Sleep (Wrist Actigraphy)

To measure fluctuations in naturalistic sleep, participants wore a Readiband Actigraph SBV2 watch (Fatigue Science, Vancouver BC) on the wrist during the study period, with the exception of removing the device during laboratory visits (approximately 3 h). The actigraph device measured movement using a 3D accelerometer with a sampling rate of 16 Hz and has been validated with respect to polysomnography and for internal consistency (Driller et al. 2016; Sadeh 2011). Since actigraphy data were stored locally on the device, data were downloaded during the experiment sessions. The device was also charged and inspected to ensure optimal functionality. Actigraph data were preprocessed by Fatigue Science software to estimate two discrete variables at each minute for a given 24‐h period: (i) the individual was “in bed” or “out of bed”; and (ii) the individual was “asleep” or “awake.” In line with previous studies (Berger et al. 2008; Thurman et al. 2018), sleep onset was defined as the first recorded instance of sleep occurring at or after 9:00 p.m., and sleep offset was defined as the last instance of transitioning from sleep to wake before 11:00 a.m. the following day. Prior work from our group on this same dataset shows strong agreement between actigraph estimates for sleep onset and sleep offset times and subjective sleep logs (Thurman et al. 2018). Due to this high correspondence, we only used actigraph sleep data in this study.

2.3.1. TST and Sleep Regularity Index (SRI)

For the purpose of the present study, using the measurements of actigraphy, we extracted two metrics of naturalistic sleep variation that allowed us to study both the immediate effect of sleep as well as relatively longer term sleep consistency while focusing on naturalistic fluctuations. TST was calculated for the day prior to the session as the length of the sleep period minus the amount of the time awake during the sleep period (TST = sleep offset—sleep onset—wake after sleep onset). We also examined the SRI, a measure to assess the regularity in individual sleep patterns both in terms of duration and time of the day. The Sleep Regularity Index (SRI, (Fischer et al. 2021; Phillips et al. 2017)) was calculated for the 2‐week period preceding the session. SRI measures the average proportion of minutes across consecutive days during which individuals were consistently asleep. An SRI value of 1 indicates perfect consistency, where every minute of actigraphy‐determined sleep and wakefulness occurred at the same times on consecutive days. Conversely, an SRI value of 0 signifies no overlap in sleep or wakefulness times between consecutive days.

In addition to using TST as a continuous variable, we also split the TST into four categories in order to reflect the meaningful categories of sleep from past sleep literature (Arora et al. 2011; Choi et al. 2017; Patel et al. 2004; Steptoe et al. 2006), which include less than 6 h of sleep (N = 8), between 6 and 7 h of sleep (N = 10), between 7 and 8 h of sleep (N = 5), and more than 8 h of sleep (N = 16). In general, a sleep duration of 7–8 h is typically considered within the normal range, whereas a sleep duration of less than 6 h is generally considered low. Additionally, subjects below ~50% of all subjects' SRI (0.84) were considered irregular sleepers (N = 18), and those above ~50% were considered regular sleepers (N = 21).

2.4. fMRI Data Acquisition

We collected fMRI data using a 3 T Siemens Prisma MRI. Functional image data were acquired using echo‐planar imaging that involved the collection of 64 coronal slices with a 3 mm slice thickness, a field of view of 192 × 192 mm, a flip angle of 52°, and a repetition time of 910 ms. The echo time was 32 ms, and the resulting voxel size was 3 × 3 × 3 mm. High‐resolution structural images were also collected for coregistration and normalization of functional brain images using a magnetization‐prepared rapid acquisition gradient echo (MPRAGE) sequence with a spatial resolution of 0.9 × 0.9 × 0.9 mm and a field of view of 241 × 241 mm. These structural images were collected with a repetition time of 2500 ms and an echo time of 2.22 ms.

2.4.1. fMRI Preprocessing

Neuroimaging data were preprocessed using ANTs (Avants et al. 2014; Tustison et al. 2021; Avants et al. 2009; Avants et al. 2014). The functional data underwent minimal preprocessing to correct for physiological artifacts and head motion. Physiological artifacts including respiration and cardiac cycle effects were corrected using RETROICOR (Glover et al. 2000) implemented in MEAP v1.5 (Cieslak et al. 2018). Head motion was estimated using antsMotionCorr. An unbiased BOLD template was created within each session using the means of the motion‐corrected BOLD time series from each run. The BOLD templates were coregistered to the corresponding T1‐weighted high‐resolution structural images. Each session was spatially normalized to a custom study‐specific multi‐modal template which included T1‐weighted, T2‐weighted, and GFA images from 24 randomly selected participants stratified to match the study population on gender. The template was affine transformed to share the coordinate space of the MNI152 Asymmetric template. The final BOLD time series images were created using the composed transforms from head motion correction, BOLD template coregistration, BOLD‐to‐T1w coregistration, and spatial normalization into 3 mm MNI space using a single Hamming weighted sinc interpolation. All co‐registration and normalization steps were computed using ANTs.

2.4.2. Atlas Parcellation

In order to examine the dynamic state in the brain in meaningful organization, cortical regions of interest were identified using the Schaefer template (Schaefer et al. 2018, resolution: 200 parcels), which were defined by the authors of the atlas using machine learning on resting state and functional data. Further, every parcel also belonged to one of eight functional networks. Subcortical regions (14 in total) were derived using the Harvard–Oxford atlas (Makris et al. 2006) and were grouped into one functional network named subcortical.

2.5. Functional Connectivity Analysis

To assess functional connectivity among ROIs, mean regional time courses were extracted and standardized using the nilearn package (Abraham et al. 2014) in Python 2.7, and confound regression was then conducted. In particular, the time series for each region was detrended by regressing the time series on the mean as well as both linear and quadratic trends. There were a total of 16 confound regressors, which included: head motion, global signal, white matter, cerebrospinal fluid and derivatives, quadratics, and squared derivatives. This functional connectivity preprocessing pipeline was selected based on conclusions from prior work that examined performance across multiple commonly used preprocessing pipelines for mitigating motion artifacts in functional BOLD connectivity analyses (Ciric et al. 2017; Lydon‐Staley et al. 2019). Following preprocessing, wavelet coherence was estimated for each pair of regions and was averaged across frequency bands between 0.06 and 0.12 Hz, a task‐relevant frequency range of coherence (Sun et al. 2004) within 22 s windows, yielding a 214 × 214 matrix of coherence values for each pair of regions for each time window (total 39 windows).

2.6. Dynamic Community Detection

Although human brain mapping efforts have demonstrated a relationship between spatial specificity and cognitive functions, techniques rooted in network science provide a useful framework for characterizing and understanding the spatiotemporal dynamics of the functional systems subserving cognition (Bassett and Sporns 2017). One of the core concepts at the basis of network science is network modularity, which is the idea that neural units are structurally or functionally connected, forming modules or clusters (Garcia et al. 2018). This organization allows for the system to perform both local‐level exchanges of information while maintaining system‐level performance.

Here, we examined how interactions of network communities during task performance were modulated by naturalistic fluctuations in sleep within each functional network. To measure such changes in network communities during the behavioral tasks, a dynamic community detection analysis was employed across 39 temporal windows (layers) of functional brain connectivity (Bassett et al. 2011; Mucha et al. 2010). The dynamic community detection algorithm optimizes a modularity quality function, Q, to distill complex connectivity matrices into a series of coarse clusters or communities of networks across time. It was implemented using a Louvain‐like greedy algorithm (Blondel et al. 2008) and standard optimization procedures (Lima Dias Pinto et al. 2022) to assign brain regions to communities.

$$Q_G=\frac{1}{2\mu }\sum _{\mathit{\text{ijlr}}}\left[\left(A_{\mathit{ijl}}-\gamma \frac{k_{\mathit{il}}{k}_{\mathit{jl}}}{2m_l}\right){\delta }_{\mathit{lr}}+{\delta }_{\mathit{ij}}\omega \right]\delta \left(C_{\mathit{il}},C_{\mathit{jr}}\right)$$ (1)

where $Q_G$ is the generalized multilayer modularity function, indices $l$ and $r$ denote consecutive time layers; $A_{\mathit{ijl}}$ is the weighted edge between nodes $i$ and $j$ reflecting their functional connectivity in layer $l$; $k_{\mathit{il}}$ is the degree of node $i$ in layer $l$; $m_l$ is the sum of the edge weights of layer $l$; $\mu$ is the sum of the edge weights of all time layers; $C_{\mathit{il}}$ is the community affiliation of node $i$ in layer $l$, and $\delta$ is the Kronecker delta, which equals 1 if $i=j$, and 0 otherwise. The community assignments are dependent on two parameters: (1) a structural resolution γ parameter and (2) a temporal resolution ω parameter. In this analysis, we used γ = 0.94 and ω = 0.23, which were optimized through a search over a predefined range of values of γ between [0.5, 1.2], and ω between [0.2, 1.5], as detailed in (Lima Dias Pinto et al. 2024). This, on an average, yielded 14 communities across tasks. Due to the heuristic nature of the generalized Louvain algorithm, we ran 100 iterations of community detection for each participant and task. This yielded 100 sets of community labels (affiliations) for the 214 nodes for each participant and task.

2.6.1. Flexibility

Flexibility of a network node broadly refers to the likelihood that a node changes its community affiliation across time. Increased flexibility has been broadly linked to a variety of cognitive phenomena, including memory (Braun et al. 2015), learning (Bassett et al. 2011), influence (Cooper et al. 2018), opinion change (Lima Dias Pinto et al. 2022), and has even been proposed to be a potential marker of cognitive flexibility (Xiao et al. 2022).

Using a standard definition, we extracted the flexibility of a node by calculating the number of times a node changes its community affiliation normalized by the total number of affiliation changes possible (Bassett et al. 2011). If g _i is the total number of times node i changes its affiliation and L is the total number of time layers, the flexibility of node i is.

$${\xi }_i=\frac{g_i}{L-1}$$ (2)

Node flexibility value was averaged across all the 100 iterations of community detection.

2.7. Data Analysis

In this study, we investigated how TST impacts task performance and brain flexibility during task performance to assess how naturalistic sleep variations relate to dynamic reconfigurations within the brain. Apart from using TST as a continuous measure, in our primary analysis, we split the continuous variable of TST into distinct sleep categories identified from sleep literature (< 6 h, 6–7 h, 7–8 h, > 8 h). This categorization allows us to capture meaningful effects of sleep quantity while reducing the effect of actigraphy‐related potential measurement errors in capturing exact sleep duration.

There are two common approaches for analyzing categorical data in this context. One way is to select a baseline category and compare the effects of other categories through pairwise comparisons. This method focuses on a single “typical” sleep duration, which would require establishing a baseline and comparing how increases or decreases in sleep relate to brain flexibility. However, this approach risks oversimplifying the analysis by reducing variability to just one baseline comparison and can make it harder to capture the broader patterns in the data. It would also necessitate multiple pairwise comparisons, potentially overlooking more subtle relationships.

Instead, we chose to use ordinal regression, which allows us to examine transitions across multiple ordered sleep categories within a single model (Agresti 2002). This approach enables us to explore the graded relationships between sleep duration and brain function without having to select a common baseline, providing a more comprehensive understanding of how variations in sleep impact brain dynamics during different tasks. Specifically, the ordinal regression considered the cumulative probabilities of the dependent variable falling within or below a particular given category. Mathematically, it estimates the probability $P\left(Y\le j\right)$ for each threshold $j$. We chose sleep categories as the dependent variable to model both behavioral and flexibility measures in relation to sleep to simplify the analysis and facilitate comparisons. In our data, $Y$ is the ordinal response indicating a sleep category, and $j$ represents each successive sleep category (less than 6, 6–7, 7–8, and more than 8 h of sleep). The model expresses the log odds of these cumulative probabilities, which is variable of interest, as a linear function of the independent variables. It uses the formula:

$$\log \left(\frac{P\left(Y\le j\right)}{1-P\left(Y\le j\right)}\right)={\alpha }_j-\left(\beta ·X\right)$$ (3)

Here, ${\alpha }_j$ represents the threshold parameters for each j, $\beta$ is the vector of coefficients associated with the independent variables, and $X$ is the vector of brain flexibility (or behavioral performance measures). Therefore, the model allows us to estimate how changes in brain flexibility and behavioral performances related to the odds of a subject's total sleep duration falling into a longer or shorter category.

2.7.1. Significance Testing

Using the “polr” function from the MASS package in R (Ripley and Venables 2009; Venables and Ripley 2002), which fits an ordinal regression model, we treated the ordered sleep categories as the response variable. We then evaluated the predictor variable—brain flexibility of each atlas parcel—against this ordinal scale. To assess the relationship between sleep duration and brain flexibility, we used the normalized ordinal regression coefficient (coefficient divided by its standard error), and to approximate significance, we compared it to a z‐test, assuming that the normalized coefficients followed a standard normal distribution, which is reasonable under the big data assumptions due to the large sample size and asymptotic properties of the estimators. The corresponding p value evaluates whether each coefficient is significantly different from 0, thereby indicating if the predictor has a measurable effect on the cumulative odds of the outcome variable. This offered a statistical measure of how variations in sleep levels are associated with the flexibility of each functional brain network. A negative coefficient with a small, significant p value suggests that a one‐unit increase in brain flexibility significantly raises the likelihood of subjects being in a shorter sleep category, and vice versa. To assess the robustness of these effects, we conducted a permutation‐based analysis in which sleep categories were randomly shuffled across participants within each task. This procedure preserved the task‐specific structure of the data while disrupting the relationship between individual sleep patterns and network flexibility.

3. Results

3.1. Sleep and Task Performance

Figure 1B shows TST from the night prior to the experimental session estimated via an actigraphy watch. Despite no manipulation of sleep schedules, substantial variability in TST was observed with an average TST of 452.05 min (~7.5 h, SD = 107.9 min or 1.8 h). First, we assessed if sleep variation, specifically, lack of sleep was associated with behavioral performance. VWM and MOD performance were determined by the percentage of correct responses. Response time in milliseconds was used to measure the PVT's performance. We did not observe a direct relationship between TST and task performance using continuous sleep measures in a linear regression analysis (Figure S1).

To test the relationship between task performance and categorized sleep, we used ordinal regression to look at whether having less or more sleep the night before the session is related to the behavioral performances during the scanning session. We found that none of the behavioral performances are related to having more or less sleep (using ordinal regression, p = 0.98, p = 0.62, and p = 0.42 for VWM, PVT, and MOD, respectively). As shown in Figure 1C, which depicts the distribution plots of behavioral performances for sleep categories for all three VWM, PVT, and MOD tasks, participants' behavioral performances overall were relatively close in each sleep category (also p > 0.05 using the analysis of variances). Despite the substantial variation in TST, the lack of effect of sleep on performance across tasks could be due to neurophysiological mechanisms combating external or environmental factors to maintain behavior (Drummond et al. 2004; Sullan et al. 2021).

3.2. Sleep and Brain Flexibility

We hypothesize that the absence of the effect of sleep is due to the compensatory mechanisms (Cabeza et al. 2018) designed to overcome environmental factors and provide adaptability, which at a macroscopic scale may be captured by network reconfigurations (Garcia et al. 2018). In the current study, we used dynamic community detection to distill the reconfiguration of network modules and further used network flexibility as an objective measure to quantify the extent of network reconfigurations. Modulations in flexibility are crucial for efficient execution of cognitive function (Bassett et al. 2011; Braun et al. 2015), and it has been associated with cognitive flexibility (Xiao et al. 2022) and resilience to opinion change (Lima Dias Pinto et al. 2022). Therefore, in our next analysis, we assessed whether the brain dynamics or network reconfigurations, measured by flexibility, are impacted by the total amount of sleep during task performance.

Flexibility was calculated for each brain region (node), identified as one of 200 parcels in (Schaefer et al. 2018) and also subcortical regions defined by the Harvard–Oxford atlas, using the dynamic community detection methodological approach on time‐evolving patterns of functional connectivity. Overall, the distributions of flexibility during PVT, VWM, and MOD tasks across all nodes ranged between 0.2 and 1 (on a scale of 0–1, Figure 2A) within the brain, showing a substantial change in community affiliations during task performance. Figure 2B shows the distribution of flexibility across sleep categories for each behavioral task. We conducted ordinal regression to examine the relationship between the probability of each subject having shorter sleep time the night before the session and the flexibility during the session and found that overall, the whole brain flexibility isn't different across the sleep categories (p = 0.06, p = 0.63, and p = 0.77 for PVT, VWM, and MOD, respectively). We also did not find a direct relationship between TST and whole brain flexibility in a regression analysis (Figure S2).

FIGURE 2.

(A) Three histograms depict the distribution of node flexibility during PVT, VWM, and MOD tasks. (B) Three panels depict the distribution of whole brain flexibility during PVT, VWM, and MOD tasks for sleep categories. The red point indicates the median value. We did not observe a direct effect of total sleep on whole brain flexibility using ordinal regression (p = 0.06, p = 0.63, and p = 0.77 for PVT, VWM, and MOD, respectively).

Although global brain flexibility provides a glimpse into the average tendency for the nodes to change community affiliation or reconfigure, it might not capture more localized, network‐based changes that are more task‐specific. Regional specificity has been observed in previous studies relating flexibility and behavior in several task performance experiments including memory performance (Bassett et al. 2011), learning (Reddy et al. 2018), motor tasks (Monteiro et al. 2019), opinion change (Lima Dias Pinto et al. 2022), and other complex scenarios such as multitasking (Alavash et al. 2015) and robot‐assisted surgery (Shafiei et al. 2020). Therefore, building on previous studies that utilized the Schaefer atlas to systematically investigate brain mechanisms by breaking the whole brain into smaller, functionally distinct regions (Dai et al. 2020; Stanford et al. 2022; Wang et al. 2024; Zhou et al. 2023), we narrowed down our analysis in the next step by examining the relationship between flexibility and TST at a functionally relevant scale, focusing on functional networks (or systems).

3.3. Sleep Impacts Functional Network Flexibility During Tasks

We examined how the TST relates to the flexibility in each of the functional networks for all three tasks. We used ordinal regression and described our findings in Figure 3 and Table 1. We looked at the normalized regression coefficient of how flexibility in each functional network relates to the probability of subjects sleeping less. The threshold of ±1.96 shows that there is a significant relationship between flexibility and sleep categories (Figure 3). In Table 1, we elaborate on the observed effects and direction of the effect. Notably, a negative coefficient value indicates that higher flexibility was associated with a decreased amount of sleep time, and a positive coefficient indicates that higher flexibility was associated with an increasing amount of sleep time.

FIGURE 3.

Flexibility relation with sleep categories. (A) Polar plot showing the normalized coefficients of flexibility obtained from ordinal regression fitted on sleep categories. The yellow, green, and red dots represent PVT, VWM, and MOD tasks, respectively, and each dot corresponds to a functional network. Significance level (1.96) corresponds to p value < 0.05. (B) Distribution of flexibility within the control network for the PVT task across sleep categories clearly depicts the change in flexibility as a function of sleep quantity. (C) Polar plot showing the normalized coefficient of ordinal regression during resting state.

TABLE 1.

The normalized coefficients of flexibility obtained from ordinal regression fitted on sleep categories.

Network	PVT		VWM		MOD		Rest
	Normalized coefficient	p	Normalized coefficient	p	Normalized coefficient	p	Normalized coefficient	p
Cont	−4.004	0.000	−2.697	0.007	0.388	0.698	0.572	0.567
Default	−2.809	0.005	−0.287	0.774	0.111	0.911	−1.054	0.292
DorsAttn	0.589	0.556	0.215	0.830	−0.106	0.915	0.003	0.997
HarvOx (subcortical)	−1.582	0.114	−2.370	0.018	0.841	0.400	−1.350	0.177
Limbic	−3.764	0.000	1.291	0.197	0.096	0.924	0.908	0.364
SalVentAttn	−3.521	0.000	−1.770	0.077	−0.898	0.369	0.067	0.946
SomMot	−1.671	0.095	−1.339	0.180	0.803	0.422	0.106	0.916
TempPar	−1.867	0.062	0.583	0.560	0.785	0.433	−0.339	0.735
Visual	0.687	0.492	1.394	0.163	0.011	0.992	−0.844	0.399

Note: A negative z‐score value indicates that higher flexibility is associated with decreased amounts of sleep time. Significant relationships (|normalized coefficient| > 1.96, p < 0.05) are highlighted in bold text.

Figure 3A and Table 1 show that certain networks showed a significant increase in flexibility (i) with decreasing amounts of sleep time, (ii) in a task‐specific manner. Specifically, for PVT, we observed that flexibility increased as sleep time decreased in the control network (coefficient = −4.004, p value < 0.001), default mode network (coefficient = −2.809, p value = 0.005), limbic network (coefficient = −3.764, p value < 0.001), and salient ventral attention network (coefficient = −3.521, p value < 0.001). Similarly, for VWM, we observed that flexibility increased as sleep time decreased in the control network (coefficient = −2.677, p value = 0.007) and subcortical network (coefficient = −2.37, p value = 0.018). In Figure 3B, we demonstrate the observed relationship for the control network during PVT, showing that flexibility increases as TST decreases.

Interestingly, none of the networks showed any significant relationship for the MOD task. We also observed the absence of any significant relationship between flexibility and TST during the resting state (Figure 3C), highlighting that the observed relationships emerge as a function of task‐specific processes. These findings remained unchanged after including age and sex as model covariates (Table S2). To test the robustness of these findings, nonparametric permutation testing was conducted by shuffling sleep category labels within each task while preserving the structure of the flexibility data. For all significant effects reported above, the observed coefficients exceeded the 95% range of the permuted null distribution (see Table S1 and Figure S3 for the distribution of permutation results).

3.4. Sleep Regularity Further Differentiates Flexibility Variation During Tasks

TST the night before the scanning session and experiments provides an instantaneous readout of an individual's sleep; to further understand the relationship between sleep quantity and flexibility during task performance, we used the SRI, a longer term metric of sleep consistency measured over the period of 2 weeks before the scanning session. Figure 4A shows a distribution of observed SRI values across individuals based on which we constructed two groups of regular and irregular sleepers such that individuals in the lower 50% of SRI values were identified as irregular. Similar to Figure 1, we did not find any relationship between behavior and sleep quantity when tested separately within these groups (Figure S4).

FIGURE 4.

(A) Distribution of SRI. Top and bottom 50% (cutoff SRI = 0.84) individuals are identified as regular and irregular sleepers, respectively. (B, C) Radar plots showing the normalized coefficients of flexibility obtained from ordinal regression fitted on sleep categories for irregular and regular groups. The yellow, green, and red dots represent PVT, VWM, and MOD tasks, respectively, and each dot corresponds to a functional network. Significance level (1.96) corresponds to p value < 0.05.

We observed a substantial impact of SRI on the relationship of sleep time and flexibility during task performance, as described in Figure 4B and Table 2. Importantly, the direction of most of the observed significant effects remained consistent, such that flexibility increased with decreasing amounts of sleep during task performance. Following the findings presented in Figure 3 and Table 1, during PVT, both groups showed a significant relationship for the control network (irregular: coefficient = −2.903, p value = 0.004; regular: coefficient = −3.224, p value = 0.001) as well as the limbic network (irregular: coefficient = −3.359, p value = 0.001; regular: coefficient = −2.119, p value = 0.034). The relationship for the default mode network was observed only in the regular group (coefficient = −3.622, p value < 0.001), whereas for the salient ventral attention, it was for the irregular group (coefficient = −3.496, p value < 0.001), with the regular group showing only a marginally significant relationship (coefficient = −1.732, p value = 0.083). In addition to these effects, we also observed a significant relationship for the somatomotor network for the irregular group (coefficient = −2.040, p value = 0.041) and the temporal parietal network for the regular group (coefficient = −1.992, p value = 0.046).

TABLE 2.

The normalized coefficients of flexibility obtained from ordinal regression fitted on sleep categories within each group of regular and irregular sleepers identified using the SRI.

Network	PVT				VWM				MOD
	Normalized coefficient		p		Normalized coefficient		p		Normalized coefficient		p
	Irregular	Regular	Irregular	Regular	Irregular	Regular	Irregular	Regular	Irregular	Regular	Irregular	Regular
Cont	−2.903	−3.224	0.004	0.001	−2.111	−1.428	0.035	0.153	−0.922	1.584	0.357	0.113
Default	−1.354	−3.622	0.176	0.000	0.782	−1.275	0.434	0.202	−0.828	0.952	0.407	0.341
DorsAttn	0.778	−0.169	0.436	0.865	0.114	0.362	0.909	0.717	−0.806	0.736	0.420	0.462
HarvOx	−1.404	−1.006	0.160	0.314	−1.807	−1.666	0.071	0.096	0.455	0.809	0.649	0.418
Limbic	−3.359	−2.119	0.001	0.034	1.791	−0.118	0.073	0.906	−0.595	0.801	0.552	0.423
SalVentAttn	−3.496	−1.732	0.000	0.083	−1.083	−1.382	0.279	0.167	−1.474	0.105	0.140	0.916
SomMot	−2.040	0.111	0.041	0.911	−0.493	−1.277	0.622	0.202	−0.781	1.924	0.435	0.054
TempPar	−1.047	−1.992	0.295	0.046	0.253	0.589	0.800	0.556	0.157	1.049	0.875	0.294
Visual	0.034	0.520	0.973	0.603	2.622	−1.127	0.009	0.260	−1.268	1.507	0.205	0.132

Note: A negative z‐score value indicates that higher flexibility was associated with a decreased amount of sleep time. Significant relationships (|normalized coefficient| > 1.96, p < 0.05) are highlighted in bold text.

During VWM, following the findings from Figure 3 and Table 1, we observed a significant relationship for the control network for the irregular group (coefficient = −2.111, p value = 0.035). Relationships for the subcortical networks were only marginally significant (irregular: coefficient = −1.807, p value = 0.071; regular: coefficient = −1.666, p value = 0.096). Additionally, we observed an inverse relationship for the visual network for the irregular group such that flexibility decreases with decreasing sleep time (coefficient = 2.622, p value = 0.009). For MOD, none of the networks showed any significant relationship, consistent with the findings discussed in the previous section. The resting state, however, showed specific and drastically different group‐dependent relationships between functional network flexibility and sleep (Figure S5 and Table S3).

These findings highlight differences in how low sleep impacts brain flexibility during task performance based on the longer term sleep consistency of an individual. Specifically, somatomotor, salient ventral attention (for irregular) and temporal parietal networks (for regular) showed only group‐specific relationships during PVT, and the default mode showed a stronger relationship (for regular) when the groups are separated. Another key difference emerges during VWM, with the visual network (for irregular) showing an inverse relationship between flexibility and sleep time as compared to all the other significant relationships observed.

4. Discussion

The current study examines the relationship between naturalistic fluctuations in sleep and flexibility within the brain during multiple tasks. Using a network measurement of flexibility, derived from a well‐established dynamic community detection method (Bansal et al. 2024; Lima Dias Pinto et al. 2022, 2024), we estimated, on average, how much each brain region within a functional network may flexibly combine and recombine with the dynamic modules in the brain to perform different tasks that vary in complexity and underlying cognitive demands. This measure relies on the complex, dynamic network representation of the brain regions and provides an aggregate, objective measure to capture dynamic reconfigurations within this network as different modules (or communities) emerge and evolve in response to neurophysiological processes and cognitive demands (Garcia et al. 2018; Mucha et al. 2010). Although there is a long history of inspecting how the interactions between regions, subregions, or systems of the brain give rise to cognition (Tognoli and Kelso 2014)—that is, so‐called functional (or effective) connectivity (Friston 2011)—it was only in the past few years that researchers have emphasized that the changes within the patterns of connectivity across a variety of time scales (milliseconds (Garcia et al. 2020); seconds (Shine et al. 2016); minutes (Betzel et al. 2017)) are behaviorally relevant and can also characterize certain patient populations (Braun et al. 2016; Cooper et al. 2018).

4.1. Naturalistic Variations in Sleep Influence Brain Reconfigurations Measured Through Flexibility

A unique feature of our study is that despite no manipulation of sleep, we find robust associations linking fluctuations in sleep duration with network flexibility in the brain during both PVT and VWM tasks. Previous research has directly linked task performance to sleep quantity (Basner and Dinges 2011; Chee and Chuah 2008; Drummond et al. 2004; Knufinke et al. 2018), but our work extends this literature to natural fluctuations in the previous night's sleep to neural processes subserving key executive functions. Specifically, the absence of a relationship (i) between the amount of sleep and flexibility during the resting state (Figure 3C) and (ii) between the amount of sleep and global brain flexibility (Figure 2) highlights that the observed relationships between sleep and flexibility emerge with the task demands within specific networks of the brain (Table 1). Interestingly, we did not observe any relationship between functional network flexibility during the MOD task and sleep quantity. This could be due to factors including more nuanced reconfigurations during MOD tasks occurring at a much finer scale than the functional network segregation considered here (9 networks across 214 regions) or the task designs themselves, where most of the participants performed well (Figure 1C).

Although our study used classic task paradigms for sleep research, our protocol, unlike more traditional sleep studies, measured sleep without depriving participants of sleep. It is also worth noting that no participants in our sample met the criteria for habitual short sleep, as all individuals averaged more than 6 h of sleep per night over the 14 days preceding the scan. This supports the interpretation that any effects of reduced sleep reflect acute, rather than chronic, sleep restriction. Therefore, the current research speaks to the growing need for studies to leave the bench and replicate the effects “in the wild” (Niell and Stryker 2010; Zaki and Ochsner 2009). This may lead to at least partially replicating previous results in more realistic contexts that may better represent our everyday experiences (Matusz et al. 2019).

4.2. Flexibility of Brain Networks May Compensate for Behavioral Consequences of Poor Sleep

Our results extend prior research suggesting that flexibility in the brain may also offset behavioral decrements associated with a potentially detrimental state (i.e., reduced sleep duration). Critically, however, we observed that across tasks, performance did not change as a function of sleep. Observed dynamic cortical changes without associated behavioral change lead us to posit that the increases in flexibility may indicate a compensatory mechanism for low levels of sleep.

Compensatory mechanisms, or the additional recruitment of neural resources to accomplish the same task, have been observed in the literature on the aging brain (Grady 2008), and in those with psychopathologies (Cabeza et al. 2002; Cirstea and Levin 2000; Feigin et al. 2006; Grady et al. 2003). For example, in age‐related studies, increases in neural activation patterns have been seen during several tasks including episodic memory tasks (Cabeza et al. 2002), working memory tasks (Grady et al. 1998; Reuter‐Lorenz et al. 2000), perceptual tasks (Grady et al. 1994), and walking (Nóbrega‐Sousa et al. 2020). Researchers have interpreted this hyperactivation in tasks as a compensation for age‐related cognitive decline. For increased brain activity or the recruitment of additional brain regions in older adults to be considered compensatory, it has been recently suggested that these changes must be present within the context of task demand and should be behaviorally beneficial for behavioral performance (Cabeza et al. 2018). We extended this idea in this work.

In the sleep deprivation literature, links have been made between the increase/alterations in brain activity during specific tasks, for example, an increase in frontal lobe activity during working memory tasks in aging adults compared to sleep‐deprived young adults (Chee and Chuah 2008; Harrison and Horne 2000); topological configurations during attention tasks in young adults (Pesoli et al. 2022); increased connectivity between the anterior cingulate cortex and insula during vigilance tasks (Fu et al. 2022); as well as increased stimulus‐related responses in a distributed set of networks supporting attention, executive control, motor, visual, and default‐mode in adolescents (DiFrancesco et al. 2019) during PVT. This could suggest a type of compensatory neural activity even in the young population. These activity patterns, however, do not shed light upon the nature of this compensation. Could the increased activity in the elderly be a consequence of different neural populations “working harder” to perform the same task, or is another network intervening to compensate for the network changes? Although our results cannot speak to age‐related changes, our results provide preliminary support for the idea that day‐to‐day compensations in sleep fluctuations might be due to a combination of resource deployment and a recruitment of other areas to maintain adequate performance during tasks.

Our results suggest that the term and action of compensation within the brain may be extended, and we propose that successful compensation, previously defined as the recruitment of additional neural resources to increase performance, could be extended to healthy brains and naturalistic fluctuations in a variety of states including potentially transient and nonpathological fatigue states. Here, we have used dynamic community detection to identify rapid reconfigurations in brain networks that provide the way this deployment of resources may occur, and we specifically interpret resources to be a change in functional connectivity. Future research may extend this further, perhaps exploring other environmental factors that may decrease performance, which could be mitigated by dynamic and rapid reconfigurations within the brain.

4.3. Network Dynamics, Resting State, and Task Performance

Our results suggest that flexibility may be a compensatory mechanism under the pressure of external events (i.e., sleep loss) based on an extension of a previously applied definition of successful compensation within the aging literature. We postulate that after a regular life event that may induce a poor night's sleep (e.g., irregular environmental events at night, everyday life stressors), our brain may deploy additional resources, reconfiguring a processing network in order to maintain performance. We did not observe such a reconfiguration during the resting state. Interestingly, other work has found that flexibility may even increase during the resting state to a higher degree than during a task state (Varangis et al. 2022). At first glance, this may appear inconsistent with our results; however, it seems likely that an even more unfettered thought process—like during the resting state—where individuals freely (and often) think about previous and future events (Binder 2012) may lead to reconfigurations distinct from a highly structured task, which contrived to test one's working memory or vigilance employed in this study. In fact, when the regularity of sleep is considered, we find differences in the relationships between functional network flexibility and sleep quantity between regular and irregular groups during the resting state, as discussed in the Figure S4, indicating a higher population variability and lack of robust overall effect during rest. Future studies may disentangle these findings and explore regularity‐based individual differences further.

4.4. Strong Role of the Control Network for the Compensation During PVT and VWM Task Performance

Our results demonstrate that the control (fronto‐parietal control) network consistently exhibits a compensatory effect during PVT and VWM tasks across all participants. For both PVT and VWM, the flexibility of the control network showed a robust relationship with sleep quantity (Tables 1 and 2). Sleep deprivation is known to impair attention and executive control functions, primarily mediated by the fronto‐parietal network (Seeley et al. 2007; Vincent et al. 2008). This network is critical for coordinating behavior in a rapid, accurate, and flexible goal‐driven manner and has been shown to act as a flexible hub interacting with and altering other functional brain networks during cognitive control (Marek and Dosenbach 2018). Prior research suggests that reduced integration within this network under sleep deprivation can lead to decreased PVT performance, emphasizing the importance of maintaining network connectivity for optimal cognitive function (Kurtin et al. 2023; Yao et al. 2023). Higher flexibility of control regions associated with shorter sleep further highlights the ability of this network to alter interactions with other modules or networks within the brain to maintain performance. Notably, the limbic and default mode networks also showed a robust relationship for their flexibility with sleep quantity, particularly for PVT. The limbic system showed an inverse effect across the combined groups, which aligns with findings that robust connectivity within limbic and default mode networks is associated with cognitive resilience (Stanford et al. 2022).

4.5. Sleep Regularity Modulates the Relationship Between Flexibility and Sleep Quantity

We also observed that the sleep regularity, a relatively longer term measure assessing an individual's sleep consistency (Fischer et al. 2021), moderates the relationship between network‐specific flexibility and total sleep the night before. This indicates the importance of individual (sleep) traits in determining the nature of reconfigurations during task performance that follow total sleep the night before, a relatively instantaneous measure.

When examining the participants by sleep regularity, distinct patterns emerged. Among regular sleepers, significant effects were observed within the default mode and temporal–parietal networks, consistent with the idea that stable sleep patterns preserve the functional integrity of these networks. Conversely, irregular sleepers exhibited unique effects within the ventral attention network, indicating a compensatory mechanism, likely driven by acute partial sleep loss. As sleep patterns became increasingly disturbed over the week, the brain recruited additional resources to maintain task performance, particularly when comparing across subjects. These findings are supported by evidence linking regular sleep to the stability and functionality of the default mode network (Lunsford‐Avery et al. 2020). They further indicate that different pathways of compensatory mechanisms come into play across different groups with different sleep trends. Although any adversarial or beneficial effect of these specific configurations is unknown, these observations align with the literature emphasizing the role of sleep for the mental (cognitive) and physical well‐being of an individual (Palmer and Alfano 2017).

4.6. Potential Implications of State Changes in Neural Plasticity

It is well established that short‐term sleep loss has robust effects on cognition (Harrison and Horne 2000); it is also well known that long‐term sleep disturbances are associated with a variety of health conditions, mental disorders, and diseases (Krystal et al. 2008). This study takes a multidisciplinary approach by combining network neuroscience methods, statistical modeling, and cognitive experiments to capture naturalistic fluctuations in sleep, uncovering neural changes due to a mere 25% reduction in sleep (8 vs. 6 h). Although previous research has shown a clear but relatively modest effect of sleep deprivation on task performance (Basner and Dinges 2011; Chee and Chuah 2008; Drummond et al. 2004; Knufinke et al. 2018), our results show that even without severe restriction in sleep, there are still measurable changes in brain dynamics that affect how people perform the following day. This research not only expands our understanding of the functional consequences of slight reductions in sleep, but it may also provide an analytical framework to probe neuroplastic changes associated with state changes. In particular, our results highlight the effects of transient sleep loss on cognitively relevant brain dynamics within relatively short time periods. This methodological scheme has the potential to provide insights into other potentially potent state changes in people's everyday dynamic environments.

4.7. Methodological Constraints

A majority of previous sleep studies have used controlled environments and designed sleep deprivation to assess the impact of sleep on behavior and neurological processes. Here, we employed a more realistic, ecologically relevant approach by considering naturalistic fluctuations in sleep. Although it might look somewhat unstructured at first glance, previous work has shown measurable effects of naturalistic sleep variations (Knufinke et al. 2018). Moreover, the variation we observed within our data and the categorization we used, especially less than 6 h, has been characterized as “short sleep” causing sufficient impairment in performance and measurable changes in the large‐scale brain dynamics (DiFrancesco et al. 2019). Furthermore, our study included multiple tasks and different sleep metrics, providing not just an instantaneous TST but also a long‐term readout of an individual's sleep through SRI.

However, we believe that these results should be considered only preliminary evidence of potentially widespread and regular compensatory action within the brain, accommodating for moment‐to‐moment hardships from the environment (sleep variation) and triggering reconfigurations in brain networks to maintain adequate performance. To ensure these effects were not simply driven by demographic differences, we repeated the analyses while statistically adjusting for age and sex (modeling sleep categories as a function of flexibility, age, and sex), and the results remained unchanged. Future studies may incorporate a larger population with a wider age group and a larger variety of tasks to test the robustness of these findings. Especially, the lack of effect during MOD and the varied group‐dependent effect during rest will need further exploration to be fully understood.

In this study, we used sleep categorization and ordinal regression to be able to extract functionally relevant, but coarse dependencies of brain reconfigurations on sleep quantity and regularity, which do not rely on accurate measurement of sleep time to the minutes level. With recent improvements in actigraphy devices and sleep tracking, further robust details of an individual's sleep can be obtained, providing a deeper understanding of sleep quality. These factors, such as sleep stages and cortical arousals, can be incorporated into the modeling approaches to further fine‐tune these findings. Finally, it is well noted that individual differences exist with regard to the amount of sleep a person may need or the quality of their sleep (Hennevin et al. 1995; Khalsa et al. 2017; Otte et al. 2011). The relatively small sample size may limit the generalizability of our findings and leave open the possibility that unmeasured nuisance factors contributed to the observed effects on brain flexibility. Although demographic and scanner‐related variability were minimized through a homogeneous sample and consistent acquisition protocol, a larger sample would allow for more comprehensive control of inter‐individual variability. As an initial investigation, this study offers a proposed mechanism linking recent sleep patterns to task‐related brain dynamics. Although beyond the scope of the present study, a potential direction for future research is to examine whether these sleep duration effects on brain dynamics depend on either an individual's sleep need or quality of sleep to understand subsequent behavioral performance. Additionally, future work should examine the robustness of these findings across different brain parcellations. Although we used the Schaefer functional atlas to best capture task‐related dynamics, testing other atlas frameworks would strengthen generalizability.

5. Conclusions

In conclusion, we examined how naturalistic variations in sleep influence brain flexibility during varying cognitive tasks including PVT, VWM task, and MOD. Our findings demonstrate that lack of sleep, even in naturalistic settings, leads to increased brain network flexibility, which is a measure of the dynamic reconfiguration of functional brain modules, particularly within the fronto‐parietal control network during PVT and VWM tasks. Critically, the task performance itself did not change with sleep time. This means that the brain's compensatory flexibility, defined as the over‐recruitment or hyperactivation of brain regions to meet task demands, was sufficient to maintain performance in the face of reduced sleep duration. Such over‐recruitment likely reflects a greater demand on neural resources or less efficient usage. No significant sleep‐related changes in flexibility were observed during the resting state, highlighting the task‐specific nature of these compensatory dynamics. Notably, sleep regularity may enhance these compensatory mechanisms, with regular sleepers showing more efficient network engagement in specific tasks. Together, these findings suggest that compensatory mechanisms during naturalistic variations in sleep involve over‐recruitment of brain regions as a targeted response to task demands, requiring greater neural resources due to inefficiencies in network usage.

Author Contributions

X.Z. conducted the analysis. X.Z., J.O.G., K.B. developed the framework and modeling approach. K.B., J.O.G., E.F.‐E. provided input on data analysis. X.Z., J.O.G., K.B. wrote the manuscript. B.G., S.G., J.C.E., J.M.V. planned the experimentation and oversaw the data collection. S.M.T. helped with data curation. V.J.L. provided input on statistical modeling. J.O.G., K.B. supervised the research. X.Z., K.B. prepared the figures and tables. All authors reviewed and provided feedback on the manuscript.

Supporting information

Figure S1: We did not find any directly linear relationship between behavior and total sleep time (TST) the night before for any of the tasks. Here we show the scatter plots between behavior and TST for PVT, VMW, and MOD. Blue lines represent the linear fit, none of the tasks showed a significant fit.

Figure S2: We did not find any directly linear relationship between brain‐wide flexibility and total sleep time (TST) the night before for any of the tasks. Here we show the scatter plots between brain‐wide flexibility and TST for PVT, VMW, and MOD. Blue lines represent the linear fit, none of the tasks showed a significant fit.

Table S1: To test the robustness of the relationships between sleep and network flexibility as reported in Table 1, we computed normalized coefficients from 1000 permutations in which sleep category labels were shuffled. Two‐tailed significance was determined using the 2.5th and 97.5th percentiles of the permuted coefficient distribution (α = 0.05). Observed (unshuffled) normalized coefficients were compared against this null distribution. Values highlighted in bold indicate coefficients that exceeded the permutation‐derived thresholds, reflecting statistically significant associations unlikely to arise by chance. Shuffled p value indicates the fraction of shuffled coefficients that were higher than observed (unshuffled) coefficient.

Figure S3: Permutation‐based null distributions of flexibility effects on sleep category across tasks. For visualization purposes, each histogram shows the distribution of normalized coefficients from 1000 permutations of the flexibility–sleep category relationship within the task (PVT, VWM, MOD) for the control network (corresponding to row one in Table S1 and Table 1). The x‐axis represents the normalized coefficients from the ordinal regression model predicting sleep categories from brain network flexibility in the control network. Red dashed lines indicate the true (unshuffled) normalized coefficient. Black dotted lines mark the 2.5th and 97.5th percentile cutoffs of the null distribution (two‐tailed α = 0.05). True normalized coefficients in the PVT and VWM tasks exceed this threshold, indicating statistically significant effects.

Table S2: The normalized coefficients of flexibility obtained from ordinal regression fitted on sleep categories, whereas age and sex are included as covariates. A negative z‐score value indicates that higher flexibility was associated with decreased amount of sleep time while keeping the age and sex constant. Significant relationships (|normalized coefficient| > 1.96, p < 0.05) are highlighted in bold text.

Figure S4: Consistent with Figure 1, we did not find any significant relationship between sleep categories and behavior using ordinal regression even after separating regular (p = 0.98, p = 0.62, p = 0.42 for PVT, VWM and MOD, respectively) and irregular groups (p = 0.26, p = 0.11, p = 0.99 for PVT, VWM, and MOD, respectively).

Figure S5: Polar plots showing the coefficients of flexibility during resting state obtained from ordinal regression fitted on sleep categories (Wald test z‐score) for irregular and regular groups. Significance level (1.96) corresponds to p value < 0.05. Further details are discussed in the Table S3.

Table S3: The normalized coefficients of flexibility obtained from ordinal regression fitted on sleep categories during resting state and the p value. A negative z‐score value indicates that higher flexibility was associated with decreased amount of sleep time. Significant relationships (|normalized coefficient| > 1.96, p < 0.05) are highlighted in bold text.

Zhou, X. , Lauharatanahirun N., Thurman S. M., et al. 2025. “Flexibility of Brain Networks May Curtail Cognitive Consequences of Poor Sleep.” Human Brain Mapping 46, no. 14: e70362. 10.1002/hbm.70362.

Data Availability Statement

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