Exploratory GABAa-informed control network modulates hyperarousal brain dynamics in chronic insomnia

Abstract

Chronic insomnia disorder is characterized by hyperarousal, a heightened cortical activation pattern that disrupts normal sleep. While hyperarousal has been linked to altered brain state dynamics, the underlying neurobiological mechanisms remain poorly understood, particularly regarding the influence of inhibitory neurotransmitter signaling through GABAa receptors. This study demonstrates that hyperarousal in chronic insomnia is characterized by more frequent and unpredictable transitions between brain states compared to healthy controls, as revealed by hidden Markov modeling of resting-state functional MRI data. By conducting an exploratory integration of DTI-based structural connectivity and regional GABAa receptor distribution within a network control theory framework, we find that chronic insomnia is associated with a flattened energy landscape reflecting hyperarousal, indicating that less energy is required for the brain to transition between states. Notably, accounting for GABAa receptor distribution increases the control energy required for state transitions and is associated with greater stability in brain state dynamics. These findings provide neurobiological insights into hyperarousal in chronic insomnia, with an exploratory analysis suggesting a modulatory role of GABAergic signaling in shaping the brain’s dynamic dysfunction.

Chronic insomnia disorder is characterized by a flattened control energy landscape underlying hyperarousal, and GABAa receptor-mediated inhibitory signaling may play a key role in regulating these altered brain state dynamics.

Introduction

Chronic insomnia disorder (CID) is characterized by persistent difficulty initiating or maintaining sleep, resulting in significant impairment in daily functioning^1–3. A core feature of CID is hyperarousal, defined as a state of heightened cortical and physiological activation that interferes with the natural reduction in arousal necessary for sleep initiation and maintenance^4–6. Advances in functional magnetic resonance imaging (fMRI), especially static resting-state functional connectivity (RSFC), have offered important insights into the neural mechanisms of hyperarousal in CID, identifying specific brain regions and networks linked to heightened arousal states⁷. For example, the salience network (SN), comprising brain regions such as the anterior insula and dorsal anterior cingulate cortex, has been consistently implicated in the hyperarousal state of insomnia due to its role in detecting and processing emotionally and physiologically salient stimuli^8,9. However, static RSFC analyses, typically performed during daytime resting states^10,11, provide only a limited snapshot, overlooking transient fluctuations in brain activity that might critically reflect the dynamic nature of hyperarousal in CID.

Recently, advancements in dynamic functional imaging techniques, particularly the hidden Markov model (HMM), have provided novel insights into transient brain dynamics^12,13. The HMM is a promising method that can identify recurrent latent brain states from resting-state fMRI data, offering a deeper understanding of spatial-temporal fluctuations in brain activity¹². In healthy individuals, HMM analyses have revealed a significant decrease in temporal complexity during consolidated sleep compared to wakefulness¹⁴, indicating that sleep is characterized by less diverse brain activity. Additionally, the hyperarousal condition caused by sleep deficits was associated with increasing transition probabilities between different dynamic brain states¹⁵. Consequently, we hypothesized that hyperarousal in CID patients would manifest as increased transitions between distinct brain networks and greater diversity in brain activities compared to healthy controls (HCs). Despite the potential of HMM, its application in CID research remains scarce, and few studies have explicitly examined how dynamic temporal features relate to the hyperarousal state in insomnia^16–18.

Network control theory (NCT) further advances our understanding by providing a quantitative framework to examine how structural connectivity constrains functional brain dynamics and transitions between brain states¹⁹. NCT allows quantification of the control energy required for brain state transitions, effectively linking structural connectivity to functional brain dynamics^19,20. Considering the hyperarousal state in a core feature of CID, we hypothesized that patients with CID would exhibit a flattening of the control energy landscape to control network transitions and persistence compared to HCs. Previous applications of NCT have demonstrated that neurotransmitter receptor distributions significantly shape brain dynamics^15,20. For example, excitatory psychedelics like psilocybin modulate brain control energy through serotonin 2a receptors, facilitating transitions across diverse brain states²⁰. Serotonin 2a is an excitatory receptor that mediates a flattening of the control energy landscape, manifested by accelerated transitions between brain functional networks and increased diversity (entropy) of brain activity²⁰. This flattening energy landscape is consistent with the hyperarousal brain state observed in CID patients. However, no studies to date have employed inhibitory neurotransmitters, such as gamma-aminobutyric acid type A (GABAa), to examine whether hyperarousal can be suppressed—an approach that could provide theoretical support for CID treatment.

GABAa receptor, the primary inhibitory neurotransmitter in the human brain, is essential for regulating cortical excitability and arousal^21,22. Dysregulation of GABAergic signaling, particularly through GABAa receptors, is implicated in various neuropsychiatric disorders characterized by hyperarousal, including insomnia^23–25. Previous studies have demonstrated that, concomitant with the alleviation of insomnia symptoms, the levels of GABAa receptors in patients gradually increase^26,27. This finding suggests that the flattening of the control energy landscape, serving as a hyperarousal biomarker for CID patients, may be modulated by the upregulation of GABAa receptors. Extending the NCT framework to incorporate GABAa receptor distribution could clarify how disruptions in inhibitory control mechanisms influence the energy landscape of brain dynamics in CID, potentially explaining the hyperarousal observed in these patients.

In the current study, we applied a comprehensive analytical approach combining HMM and receptor-informed NCT to investigate the neurobiological underpinnings of CID-related hyperarousal. First, we applied HMM to a large cohort of 377 individuals to characterize whole-brain dynamic patterns and to determine how these patterns relate to insomnia (Fig. 1a). Next, by integrating structural connectivity from diffusion MRI²⁸ and receptor-density mapping from positron emission tomography (PET)²⁹ in third-party databases, we exploratorily tested whether regional variations in GABAa receptor availability modulate the control energy required for transitions between distinct brain states (Fig. 1b). Drawing on prior literature^{15,19–21,27}, we hypothesized that patients with CID would display altered dynamic states—potentially reflecting more frequency of transition between brain states and more diverse activity patterns—along with a flattening of the energy landscape to control transitions compared to HCs. Additionally, we further hypothesized that disruptions in GABAa receptor-mediated inhibitory control contribute to the flattening energy landscape observed in CID, providing new insights into the neural mechanisms underlying chronic insomnia and potentially informing future therapeutic strategies.

Fig. 1. Study workflow overview.

a Hidden Markov Model (HMM) analysis was applied to concatenated time-series data from all participants to determine the spatial-temporal dynamics of latent states. b Using network control theory and a diffusion MRI-derived structural connectome, we calculated the minimum energy required to transition between states (or maintain the same state) using each individual’s brain states derived from the CID and HC groups separately. Our calculations reveal a flatter energy landscape in CID patients compared to HCs. Weighting the energy calculations of the CID brain states with inputs from PET-derived receptor density maps of the gamma-aminobutyric acid type A (GABAa) receptor resulted in an increased energy landscape, providing a mechanistic explanation for how the flattened brain state dynamics in CID patients can be controlled by the GABA expression.

Results

HMM-driven clustering of brain activity patterns

The HMM identified six distinct spatial states based on dynamic functional activity patterns (Fig. 2) that were different under the CID and HCs groups (Supplementary Fig. 3 and Supplementary Fig. 4). State #1 exhibited lower averaged functional activity (AFA) in the DAN. State #2 showed elevated AFA in the LN, DMN, and subcortical regions, but reduced AFA in the SN. State #3 had higher AFA in SN and lower AFA in LN relative to state #2. State #4 was characterized by increased AFA in the SMN and decreased AFA in the VN. Compared to state #1, state #5 displayed higher AFA in DAN and lower AFA in subcortical regions. State #6 resembled state #3 with higher AFA in SN and lower AFA in LN, and additionally showed higher AFA in the FPN compared to state #3.

Fig. 2. fMRI signal profiles for each brain state during rest across groups.

Left column: Six distinct, recurring brain states identified by HMM, representing variations in average functional activity across the whole brain. Right column: Relative contribution of each brain state to seven canonical networks: dorsal attention network (DAN), default mode network (DMN), frontoparietal network (FPN), limbic network (LN), somato-motor network (SMN), ventral attention network (VAN), and visual network (VN). Consistent color coding for each state is used throughout the manuscript. The blue-red color bar indicates relative loadings for average activity in the HMM model.

CID effects on the brain state dynamics

CID is significantly associated with altered brain state dynamics and transition probabilities, as evidenced by comparisons between CID patients and HCs. In terms of ALT, CID patients exhibited distinct patterns after adjusting for age and sex with FDR correction. They demonstrated significantly higher ALT in states #5 (t = 2.27, p_FDR = 0.038) and #6 (t = 3.18, p_FDR = 0.010), while showing lower ALT in state #3 (t = 2.31, p_FDR = 0.038) compared to HCs. No significant differences emerged in states #1 and #4, indicating that CID disrupts temporal stability in a state-specific manner. Similarly, FO analysis revealed a shift in the proportion of time spent in certain states. CID patients displayed higher FO in states #2 (t = 2.85, p_FDR = 0.009) and #6 (t = 3.67, p_FDR < 0.001), and CID had lower FO in states #1 (t = 2.20, p_FDR = 0.043) and #3 (t = 3.87, p_FDR < 0.001), with no differences in states #4 and #5. These findings, illustrated in Fig. 3a, b underscore a reorganization of dynamic brain activity in CID.

Fig. 3. Group differences in dynamic brain state characteristics and their correlations.

Violin plots comparing healthy controls and all CID patients for a averaged lifetime and b fractional occupancy (FO) of each state. Asterisks denote statistical significance after false discovery rate (FDR) correction (p_FDR < 0.05). Correlation analyses: c FO of state #3 negatively correlates with Pittsburgh Sleep Quality Index (PSQI) scores (r = −0.19, p_FDR = 0.019). d FO of state #6 positively correlates with PSQI scores (r = 0.17, p_FDR = 0.043).

Previous studies indicate that insomnia is a disorder marked by internal heterogeneity, which can be expressed through varying symptoms and their corresponding brain functional activities^30–32. We examined this heterogeneity of CID by investigating changes in dynamic functional activity (ALT and FO) across insomnia severity groups (mild, moderate, and severe). We found that the heterogeneity within CID can, to some extent, be reflected by the severity of insomnia symptoms. For example, mild insomnia patients had significantly lower ALT in state #4 compared to both moderate (t = 2.96, p = 0.001) and severe (t = 2.55, p = 0.005) groups (Fig. 4e). Changes in FO across HCs and CID patients with varying insomnia severity were also observed (Supplementary Fig. 5). See Supplementary materials for details. These results demonstrate that CID is associated with marked alterations in brain state dynamics, which may help elucidate the neurobiological underpinnings of insomnia and its severity-dependent manifestations.

Fig. 4. Differences in brain state transition probabilities between groups.

Group-averaged transition probability matrices for a HCs and CID patients (b), with diagonal elements omitted for clarity. Middle column: Top 20% most frequent transitions for c HCs and d CID patients. Arrow thickness represents the transition probability. e Transitions with significantly higher probabilities in HCs compared to CID patients, and f transitions significantly higher in CID patients than HCs, identified using Network-Based Statistics (p_FWE < 0.05). Arrow diameters are scaled based on fractional occupancy (refer to Fig. 3b). Each arrow denotes a specific state transition.

The relationship between these dynamic features and insomnia symptoms further highlights their relevance. PSQI scores correlated with FO, but not ALT, suggesting that occupancy patterns may better reflect symptom severity. Specifically, state #3 FO negatively correlated with PSQI scores (r = −0.19, p_FDR = 0.019), while state #6 FO showed a positive correlation (r = 0.17, p_FDR = 0.043), as depicted in Fig. 3c, d. No significant correlations emerged for other states or ALT measures, indicating that FO could serve as a sensitive marker of insomnia-related brain dynamic changes. Transition probabilities further elucidated these dynamics, revealing increased overall state transitions in CID patients, primarily due to decreased occupancy in states #1 and #3. Detailed analysis (Fig. 4a-d) using Network-Based Statistics showed more transitions among states #1, #2, #3, and #6 in CID patients compared to HCs (p_FWE < 0.05), as visualized in Fig. 4e, f. Additionally, we asked whether a more complex (entropic) brain state exists in CID patients as an additional brain dynamic marker of hyperarousal. Supporting our hypothesis, we found patients with CID have increased complexity of brain state transition patterns than HCs (p < 0.001, Fig. 5a).

Fig. 5. Network control theory unveils a flattened brain control energy landscape in CID patients.

a Patients with CID have increased complexity of brain state transition patterns than HCs (p < 0.001). b Group comparison of the transition energies in CID and HC groups using uniformly weighted whole-brain inputs. c Weighting the control vector by the GABA receptor density map results in significantly higher energies compared to uniformly weighted control vector inputs in the CID group. d The true GABAa receptor map resulted in significantly higher energy required for transition compared to the control vector constructed using the spin-permuted receptor maps. e The GABAa-induced control energy changes were significantly negatively correlated with the empirically observed transition probability in the CID group (r = −0.769, p < 0.001). f The GABAa-induced control energy changes were negatively correlated with entropy in brain state timeseries in the CID group (r = −0.197, p = 0.007).

Network control theory unveils a flattened brain control energy landscape in CID patients

We next sought to provide a direct test of our hypothesis about decreased control energy requirements to transition between different states in CID patients. The NCT was used to calculate the control energy, which is the minimum energy needed for transition between brain states. The control energy required to maintain each state is referred to as persistence energy, and the control energy required to change states is referred to as transition energy. Using this framework, previous studies have shown that the brain favors transitions to brain states that require less control energy^15,20,33. According to the NCT, energy is injected into the system at control points to induce the desired transition³³. First, we allowed the control inputs to be uniformly weighted from all brain regions (model 1). Compared with HCs, the CID group was found to have decreased control energy between brain states and the persistence energy needed to maintain each state (Fig. 5b and Supplementary Fig. 6a). Next, we sought to determine if the specific regional distribution of GABAa receptors in the brain could correspond to especially suitable control points for inducing an increase in transition energy (model 2). In every possible transition, we observed that the GABAa-weighted inputs provided higher control energy than the uniform inputs (Fig. 5c and Supplementary Fig. 6b) in the CID group.

However, it could be argued that giving additional control to some regions will result in a higher control energy, regardless of their particular spatial arrangement. To demonstrate that our results are specific to GABAa receptors’ spatial distribution across brain regions, we compared the control energy obtained from the true GABAa distribution versus 10,000 spin permutations that preserve the set of weights and their spatial autocorrelations^34,35. The true distribution of GABAa still resulted in significantly higher energies (p_spin < 0.05, Fig. 5d), demonstrating the critical role of the specific regional distribution of GABAa receptors for increasing control energy to induce low-energy state transitions.

GABAa-induced control energy changes in CID are associated with transition probability and the complexity of brain dynamics

Crucially, the results demonstrate the specific role of GABAa receptors in steepening the control energy landscape in the CID group. Therefore, we sought to test how the GABAa-weighted control energy may affect corresponding brain dynamics. Specifically, we show that, across the 6 states transitions, the GABAa-induced control energy change was significantly correlated with the empirically observed transition probability in the CID group (r = −0.769, p < 0.001, Fig. 5e). This negative correlation indicated that an increased frequency of state transitions in CID patients can be controlled by a GABA-induced control energy landscape. Additionally, we found that the increased control energy in the CID group was correlated with less entropy in their brain state timeseries (r = −0.197, p = 0.007, Fig. 5f). This result directly and quantitatively links a heightened GABAa-weighted energy landscape that controls more entropic brain activity in CID patients.

Discussion

Hyperarousal is widely recognized as a core feature of insomnia^4–6, but the brain dynamic mechanisms underlying this phenomenon remain unclear. A critical, unanswered question is how inhibitory neurotransmitter signaling, particularly through GABAa receptors, influences the dynamic characteristics of hyperarousal in CID. In this study, we integrated fMRI data with separately obtained PET and diffusion MRI data using a framework combining HMM and NCT to systematically investigate these questions. Our results revealed that CID is characterized by increased transitions between dynamic brain states and elevated entropy, indicating disrupted stability and predictability of neural activity patterns. Furthermore, NCT analyses identified a flattened energy landscape in CID patients, reflected by a reduced minimum control energy required for transitioning between brain states compared to HCs. Importantly, incorporating regional GABAa receptor densities into the control model significantly modulated this flattened energy landscape, suggesting a pivotal role of GABAergic inhibitory signaling in regulating these altered dynamics. Additionally, GABA-induced control energy change was negatively correlated with empirically observed transition probabilities and entropy patterns, further underscoring the critical influence of GABAa receptor distribution on CID-associated hyperarousal. Together, these findings provide novel insights into how impairments in GABA-mediated inhibitory control might contribute to the neurobiological underpinnings of hyperarousal observed in CID.

Using an HMM, we identified six distinct dynamic brain states, each characterized by unique patterns of functional network activation. Specifically, we observed higher FO and ALT in state #6—marked by increased SN and FPN activity and reduced LN activity—within the CID group. The prolonged occupancy and stability of state #6 suggest a sustained hyperarousal condition in patients with insomnia. The SN is integral to detecting and prioritizing salient stimuli³⁶, while the FPN underlies executive control processes³⁷. Meanwhile, the diminished LN involvement could signal weakened affective regulation³⁸. Abnormalities in the SN, such as hyperactivity in the anterior insula and dorsal anterior cingulate cortex, can lead to an exaggerated detection of stimuli, causing patients to remain overly vigilant⁸. Simultaneously, persistent engagement of the FPN, particularly in regions like the dorsolateral prefrontal cortex, may foster excessive cognitive processing—manifesting as rumination or worry—which hinders the deactivation necessary for sleep onset^11,39. In addition, reduced LN activity may reflect a diminished capacity to integrate and regulate emotional information, thereby impairing stress and anxiety modulation⁴⁰. In contrast, HCs showed higher FO and ALT in state #3, associated with relatively greater SN activity and lower LN activity, indicative of a restful, emotionally regulated state. Reduced occupancy of this restful state in CID likely exacerbates difficulties transitioning to sleep, thereby perpetuating insomnia. Moreover, significant correlations between insomnia severity, measured by PSQI scores, and the temporal features of states #3 and #6 further underscore their clinical relevance. For instance, lower FO in state #3 negatively correlated with PSQI scores, suggesting that reduced engagement in restful states is linked to more severe insomnia symptoms. Conversely, higher FO in state #6 positively correlated with PSQI scores, indicating that increased activity in this state may be associated with elevated hyperarousal levels.

The human brain’s dynamic nature, transitioning smoothly through states that support cognitive function, is essential for understanding functional brain plasticity^41,42. Our analysis of transition probabilities between the six HMM-driven states revealed that CID patients exhibit increased transitions between states #1, #2, #3, and #6. This increased transition frequency underscores heightened neural instability, suggesting impaired inhibition over transitions between states associated with varying arousal levels⁴³. The elevated frequency of transitions thus aligns with the clinical observation that insomnia patients experience ongoing challenges in sustaining sleep and exhibit elevated sensitivity to internal and external stimuli⁴. Moreover, we found elevated entropy (complexity) in brain state transition patterns among CID patients compared to HCs. Higher entropy reflects greater unpredictability and less organized temporal patterns of neural activity. Increased brain state entropy in CID suggests a more entropic and disordered neural environment, complicating the brain’s capacity to reliably and predictably transition into and sustain restful, low-arousal states necessary for effective sleep. Collectively, these findings highlight that altered brain dynamics in CID—including prolonged occupancy of hyperarousal states, increased state transitions, and heightened entropy—constitute critical neural mechanisms underpinning the persistent hyperarousal and clinical symptoms observed in chronic insomnia.

Identifying hyperarousal neuroimaging markers of CID can significantly enhance the monitoring of remission, prediction of relapses, and the development of behavioral and pharmacological interventions⁷. Unlike previous case-control studies that focused on structural imaging, task-based fMRI, or static RSFC and reported minimal differences between CID patients and HCs^44–48, our research revealed significant distinctions in dynamic rs-fMRI features. These dynamic measures not only differentiated CID patients from controls but also correlated with the severity of insomnia. Additionally, prior fMRI studies incorporating both static and dynamic brain activity metrics have demonstrated the superior sensitivity of dynamic features in diagnosing psychiatric conditions^49,50. Future longitudinal research could validate the efficacy of the dynamic features found in our study in monitoring CID progression and treatment responses.

Moreover, our findings demonstrated a flattened brain control energy landscape in CID, characterized by reduced control energy required for transitioning between different dynamic brain states compared to HCs. According to NCT, the brain transitions between states by exerting control inputs at specific neural nodes, and the magnitude of these energy inputs determines the ease or difficulty of transitioning^19,20. A flattened energy landscape implies lower energetic barriers, facilitating frequent and spontaneous state transitions due to decreased constraints imposed by preceding neural states. Previous research supports this concept, indicating that the brain naturally favors state transitions that require minimal control energy, thereby sustaining neural efficiency and stability¹⁹. For instance, similar flattened energy landscapes have been observed in psychedelic-induced states, leading to enhanced brain state diversity and increased spontaneous state transitions, reflecting decreased energetic barriers among states²⁰.

In the context of CID, the flattened control energy landscape likely promotes frequent, spontaneous shifts between brain states with varying arousal profiles, thereby sustaining a heightened state of cortical and physiological activation, clinically recognized as hyperarousal. Specifically, the lowered energy barriers diminish the energetic cost to depart from stable, low-arousal states typically conducive to sleep, facilitating frequent and effortless transitions toward more aroused, vigilant states. Over time, this weakened control over transitions could reduce the brain’s ability to maintain restful states, resulting in the persistent hyperarousal and sleep disturbances characteristic of insomnia. Thus, the flattened energy landscape in CID patients provides a neurobiological basis for the increased instability and heightened diversity in brain activity, offering critical insight into why CID patients experience difficulties in initiating and sustaining sleep.

In this study, we used a shared population-averaged structural connectome for both CID patients and HCs in our NCT model—a common simplification in large-scale functional modeling^15,20,51. This assumes broadly similar macro-scale white matter architecture across groups. Supporting this assumption, a meta-analysis by Sanjari et al.⁵² reported no consistent white matter abnormalities in insomnia, and recent large-scale studies^53,54 found only modest, region-specific associations with white matter microstructure. Nonetheless, we acknowledge that individual-level or group-specific differences in white matter architecture could influence control energy estimates. Specifically, if CID patients exhibit even subtle alterations in structural connectivity, such as reduced network efficiency or regional disconnection, our current model may underestimate the true energetic costs required for state transitions in these individuals⁵¹. This could bias our estimates toward detecting a “flattened” energy landscape in CID patients. Importantly, since the structural connectome was held constant across groups, the observed alterations in control energy primarily reflect differences in brain state dynamics derived from fMRI. Therefore, we believe that the conclusion regarding a flattened energy landscape in CID patients remains valid and interpretable.

Furthermore, our results demonstrate that weighting control inputs by the regional distribution of GABAa receptors significantly modulates the flattened brain control energy landscape observed in CID. Specifically, we found that when control inputs were uniformly distributed across all brain regions, CID patients exhibited lower control energy requirements for transitioning between brain states compared to HCs, indicating a flattened brain control energy landscape. However, when control inputs were weighted by regional GABAa receptor distribution, there was a marked increase in the control energy required for these state transitions in CID patients. This finding highlights a crucial modulatory role of GABAergic signaling in restoring stability within the otherwise flattened energy landscape.

Mechanistically, GABA functions as the primary inhibitory neurotransmitter, crucially regulating cortical excitability and arousal through GABAa receptors, which mediate inhibitory neurotransmission across cortical and subcortical circuits^21,22. The regional specificity of these receptors allows them to act as key modulators of neural dynamics by influencing the energetic demands necessary for transitioning between different brain states. Our findings suggest that regions with higher GABAa receptor density require greater control energy to shift the brain away from stable, low-arousal configurations. Thus, in CID, diminished or altered GABAergic signaling potentially weakens inhibitory control, reducing the energetic barrier between states and promoting easier transitions into hyperaroused states.

This interpretation aligns well with previous research indicating that the inhibitory system, particularly GABAergic mechanisms, is central to regulating cortical excitability and ensuring effective transitions from wakefulness to sleep^40,43. Moreover, our empirical observations support this model, as GABAa-induced control energy changes correlated negatively with the frequency of transitions between brain states and entropy measures, underscoring the importance of intact GABAergic inhibitory function in maintaining neural stability. Thus, our findings underscore the mechanistic importance of GABAa receptor distribution in modulating the brain’s control energy landscape, suggesting a critical neurobiological pathway through which inhibitory neurotransmission may influence hyperarousal and associated clinical symptoms in CID^55,56. Targeting these receptor-mediated mechanisms could potentially offer novel therapeutic strategies aimed at restoring functional stability and mitigating hyperarousal in chronic insomnia patients²⁷.

Several limitations warrant consideration. First, insomnia is a heterogeneous condition with various subtypes, which may limit the generalizability of our findings^30,31. Future studies should recruit larger, more balanced CID cohorts and employ longitudinal designs to capture within-subject variations in dynamic features and remission trajectories. Second, the relatively short fMRI scan duration and long repetition time for each participant limited the amount of temporal information available. To address this, we concatenated time series across participants and applied the HMM to the aggregated data, thereby partially enhancing temporal resolution despite the brevity of individual scans. Nevertheless, future studies should employ longer scan durations and shorter repetition times to accurately capture the dynamic characteristics of brain activity. Finally, our NCT model assumed stable structural connectivity across both CID patients and HCs, although alterations in white matter connectivity may exist in CID. While subject-specific multimodal datasets are preferable, integrating structural and functional data from separate cohorts is a well-established and increasingly common practice in NCT research, particularly in large-scale studies where diffusion imaging is not available for all participants. Several recent NCT studies have employed similar strategies, utilizing population-averaged structural connectomes and neurotransmitter receptor maps derived from independent healthy cohorts to model functional brain dynamics in clinical populations^15,20,57. This approach enables researchers to leverage high-quality, publicly available structural data to inform functional modeling in extensive resting-state fMRI datasets. Nevertheless, future work should aim to validate these findings using multimodal neuroimaging data acquired from the same cohort, allowing for more precise characterization of individual-level structure-function relationships in CID.

Our study integrates HMM and GABAa receptor-informed NCT to elucidate the neural dynamics underlying hyperarousal in CID. We found that CID patients exhibit significantly altered dynamic brain activity, marked by more frequent transitions between distinct brain states and elevated entropy compared to HCs. Moreover, NCT analysis revealed a flattened brain control energy landscape in CID, characterized by significantly lower control energy barriers for state transitions. Notably, incorporating regional GABAa receptor distribution into our NCT analyses underscored its modulatory role on this energy landscape. These findings offer crucial mechanistic insights into CID pathophysiology and suggest that GABAa receptor-targeted interventions may effectively normalize hyperarousal brain dynamics.

Methods

Recruitment and assessment

We recruited 212 patients diagnosed with CID from the outpatient clinics of the Affiliated Hospital of Chengdu University of Traditional Chinese Medicine and Chengdu Second People’s Hospital. Additionally, 208 HCs were recruited through social media advertisements from surrounding communities. CID diagnoses were established through structured interviews conducted by five neurologists, each possessing 8–15 years of psychiatric experience. Diagnoses adhered to the criteria outlined in the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-V) and the International Classification of Sleep Disorders, Third Edition (ICSD-3)⁵⁸. Detailed inclusion and exclusion criteria are provided in the Supplementary Methods. All participants were fully informed about the study objectives and provided written informed consent. The study was approved by the Ethics Committee of the Affiliated Hospital of Chengdu University of Traditional Chinese Medicine. All ethical regulations relevant to human research participants were followed. Initially, 21 healthy subjects and 22 CID patients were excluded due to excessive head movement during fMRI preprocessing. This resulted in a final sample of 190 CID patients and 187 HCs for analysis. The clinical and demographic characteristics of the CID and HC groups are presented in Table 1.

Table 1.

Demographic and clinical characteristics of the study participants

	HCs (n = 187)	CID (n = 190)
Age
Mean (SD)	34.8 (12.5)	35.1 (11.5)
Median [Min, Max]	30.0 [19, 65]	31.0 [18, 65]
Sex
Female	127 (67.9%)	128 (67.4%)
Male	60 (32.1%)	62 (32.6%)
PSQI
Mean (SD)	3.7 (1.4)	12.9 (3.0)
Median [Min, Max]	4.0 [1.0, 6.0]	13.0 [8.0, 20.0]
Missing	85 (45.4%)	0
Severity
Severe (PSQI > 15)		41 (21.6%)
Moderate (10 < PSQI < = 15)		104 (54.7%)
Mild (7 < PSQI < = 10)		45 (23.7%)
Head movement
Mean FD (mm)	0.09 (0.03)	0.10 (0.04)
Mean DVARS	0.90 (0.26)	0.91(0.27)

DVARS, the temporal derivative of time courses of RMS variance over voxels. FD Framewise displacement, CID Chronic insomnia disorder, HCs, Healthy controls, PSQI Pittsburgh sleep quality index.

Neuroimaging data acquisition

T1-weighted (T1w) and rs-fMRI images were acquired using a GE 3.0 T MRI scanner (Discovery MR750; GE Healthcare, Milwaukee, WI, USA). The following settings were used to acquire sagittal 3D T1-weighted images: repetition time (TR) = 7.06 ms; echo time (TE) = 3.04 ms; flip angle (FA) = 12°; acquisition matrix = 256 × 256; slice thickness = 1 mm, no gap; and 188 sagittal slices. The rs-fMRI images were obtained using an interleaved 2D echo-planar imaging (EPI) sequence with the following parameters: TR = 2000 ms, TE = 30 ms, FA = 90°, acquisition matrix = 64 × 64, voxel size = 3.75 × 3.75 × 3.2 mm³, thickness = 3.5 mm, slice gap = 0.7 mm, number of slices = 33, number of volumes = 240, resulting in a total scan time of 8 min.

MRI data preprocessing

T1w and rs-fMRI data were preprocessed using fMRIPrep 24.0.1⁵⁹. For T1w images, the pipeline included intensity non-uniformity correction, skull-stripping, tissue segmentation, and nonlinear spatial normalization to standard space. In the rs-fMRI pipeline, a reference volume was generated, head-motion parameters were estimated before spatiotemporal filtering, slice-timing corrections were applied, the T1w reference was co-registered, confounding time-series were extracted, and the data were resampled into standard space. During these steps, framewise displacement (FD) and DVARS (the standard temporal derivative of time courses of RMS variance across voxels) were calculated for each subject as measures of head motion. In addition, fMRIPrep estimated 36 confounds from the preprocessed time points^60,61, which were then utilized by xcp_d 0.10.1^61,62 to mitigate motion-related artifacts and noise in the rsfMRI data. Next, the functional time series were extracted from brain regions defined by the Schaefer 100-region cortical atlas⁶³ and the Tian 32-region subcortical atlas (Scale II)⁶⁴. These parcellation schemes were selected due to their high reproducibility and comparability in clinical neuroimaging studies, offering a more consistent anatomical and functional framework compared to data-driven approaches such as ICA. This choice also aligns with recent studies that have applied HMM to clinical populations⁵⁰. To enable network-level interpretation, cortical parcels were further categorized into canonical large-scale functional systems based on the Yeo 7-network parcellation⁶⁵, comprising the dorsal attention network (DAN), default mode network (DMN), frontoparietal network (FPN), limbic network (LN), somatomotor network (SMN), SN, and visual network (VN). This network mapping is publicly available from GitHub (https://github.com/ThomasYeoLab/CBIG/tree/master/stable_projects/brain_parcellation/Schaefer2018_LocalGlobal). The first five time-points were discarded due to pronounced head motion, and subjects were excluded if over 20% of time points exceeded 0.5 mm FD or 1.5 SD in DVARS. Additional exclusions were based on visual inspection of MRI scans (e.g., incomplete cortical coverage or segmentation inaccuracies). Consequently, 21 healthy individuals and 22 patients with CID were excluded for excessive head motion (see Supplementary Methods for details). Ultimately, the dataset included 377 participants, each with 235 time points and 132 brain regions, resulting in a 2D matrix of 88,595 × 132. This matrix was then standardized (mean = 0, standard deviation = 1) before being implemented in the HMM model.

Decoding brain dynamics utilizing a hidden Markov model

We employed the HMM-MAR toolbox in MATLAB (https://github.com/OHBA-analysis/HMM-MAR) to decode transient brain states from resting-state fMRI data, using variational Bayes (VB) inversion of an HMM with up to 500 iterations¹⁷. We used the VB approach for HMM inference because it provides an efficient and scalable way to estimate the model’s posterior distribution, especially when analyzing large-scale, concatenated group data¹⁷. In this framework, the observed fMRI time series X_t at each time point t is assumed to be generated from a latent hidden state S_t∈{1,2,…,K}. Each state k is characterized by a multivariate Gaussian distribution, such that the emission model is given by:

$$X_t\mathrm{\mid }{S}_t=k~\mathcal{N}\left({\mu }_k,{\mathrm{\Omega }}_k\right)$$

Where μ_k is the state-specific mean activity vector and Ω_k is the precision (inverse covariance) matrix. Temporal dependencies between states are captured by a first-order Markov process, with a transition probability matrix Θ_l,k = P(S_t = k|S_t − 1 = l), which models the probability of transitioning from state l to state k. The VB inference algorithm seeks to approximate the posterior distributions of the hidden variables and model parameters by minimizing the variational free energy. The number of hidden states K was treated as a prior and varied from 2 to 20, following established recommendations^50,66,67. Although there were some approaches to guide the choice of the number of states, different numbers of states in practice offer only different levels of detail of brain dynamics⁶⁷. For each value of K, we performed multiple HMM runs using both random and Gaussian mixture initializations to avoid convergence to local minima. Based on the inspection of free energy values, state occupancy distributions, and model interpretability^14,17,18, a six-state solution was selected (Supplementary Fig. 1). To enhance robustness, the six-state HMM was repeated 15 times using different random seeds¹³, and the ninth inversion, showing the highest average spatial correlation across runs, was selected for subsequent analyses (Supplementary Fig. 2). The HMM operates under the assumption that fluctuations in time-series data across brain regions can be explained by a limited number of latent states, referred to as recurrent spatial states. Consequently, each time point is characterized as a transient state that may either persist to the next time point or transition to a different state (Fig. 1a). This framework enables the calculation of the probability of each state at every time point and the likelihood of transitions between states. Key temporal metrics derived from the HMM for subsequent analyses include fractional occupancy (FO), which represents the proportion of time a subject spends in a specific state, and averaged lifetime (ALT), which indicates the average number of time points a subject remains in a state before transitioning to another state, thereby reflecting the stability of that state. Additionally, transition probability (TP) is computed to quantify the likelihood of transitioning between all possible states, effectively mapping the transition network of brain dynamics for each participant. This allows us to identify which particular state transitions occur more frequently in patients with CID relative to HCs, thus offering a granular and state-specific view of dynamic instability.

Complexity/entropy index

We calculated complexity/entropy indices for each participant’s brain state time series. The complexity index was computed using the seqici() function from the TraMineR R (version 2.2-7) package. This function computes a normalized composite complexity index that combines two aspects of temporal dynamics: the number of transitions in the sequence q(s), which reflects temporal instability, and the longitudinal entropy ℎ(s), which captures the distributional unpredictability of states. The complexity index is defined as:

$$\mathrm{C}\left(\mathrm{s}\right)=\sqrt{\frac{\mathrm{q}\left(\mathrm{s}\right)}{{\mathrm{q}}_{\max }}\cdot \frac{\mathrm{h}\left(\mathrm{s}\right)}{{\mathrm{h}}_{\max }}}$$

Where q_max = l – 1 is the maximum number of transitions for a sequence of length l, and ℎ_max = log 2 | A | , with |A| being the number of distinct states (alphabet size). The complexity index ranges from 0 (minimum complexity), observed in sequences with only one state and no transitions, to 1 (maximum complexity), which occurs when all states are equally represented and transitions are maximal. This metric captures both state diversity and temporal irregularity, providing a compact measure of uncertainty in brain state dynamics over time^68,69. Each participant has a single entropy value, which does not specify which states are involved but rather assesses whether the transition pattern is more or less ordered. The more complex the brain state transition pattern is, the less predictable and more entropic it is.

Structural connectivity network construction

Since diffusion MRI was not acquired in the CID study, we calculated a structural connectome derived from a third-party database²⁸ for NCT analysis. Specifically, we employed diffusion-weighted imaging (DWI) and T1w data from 180 healthy Chinese subjects (77 males, 103 females), aged 22–79 years, which closely matches the age range of our study participants²⁸. While subject-specific diffusion MRI would be ideal, using a population-averaged connectome from a demographically aligned cohort is a pragmatic and widely accepted approach, particularly in large-sample clinical studies where multimodal imaging is not always feasible. The acquisition details for the DWI and T1w data are documented in the original study²⁸, and the data are publicly available at NITRC. Here, we outline the process used to construct a population-averaged structural connectome using QSIPrep (v0.24.1)⁷⁰ and QSIRecon (v1.0.1)⁷⁰. T1w images were aligned to the MNI152NLin2009cAsym template via a 6-DOF affine and SyN nonlinear registration, with brain extraction (SynthStrip) and segmentation (SynthSeg). Diffusion images underwent MP-PCA denoising, N4 bias correction, and FSL eddy current correction, then were resampled to AC–PC space (3 mm isotropic). Anatomical atlases (Schaefer 100 cortical and Tian 32 subcortical structures) were mapped into DWI space using T1w-based normalization. Then, data were reconstructed using the mrtrix_singleshell_ss3t_noACT workflow in QSIRecon⁷⁰, with all tractography parameters set according to default recommendations unless otherwise specified. Specifically, structural connectivity was estimated as the streamline count (fiber density) between each pair of regions using a deterministic tractography pipeline implemented in MRtrix3, following constrained spherical deconvolution (CSD) based on multi-tissue fiber orientation distributions (FODs). To enhance cross-subject comparability and minimize intensity-related biases, FODs were intensity-normalized using the mtnormalize function, as part of the MRtrix3Tissue workflow. Structural connectivity networks were then constructed by integrating the reconstructed diffusion data with the parcellated cortical and subcortical regions, yielding a 132 × 132 connectivity matrix. This matrix reflects normalized streamline counts and was averaged across 180 healthy Chinese subjects to generate a population-based structural connectome, which was used in all subsequent NCT analyses.

GABAa receptor mapping

The spatial distribution of GABAa receptor densities was delineated using PET data, employing techniques previously elaborated⁷¹ and concisely reviewed here. A cohort of sixteen healthy individuals (9 females, mean age 26.6 ± 8 years) underwent PET imaging on a High-Resolution Research Tomograph (HRRT) with [¹¹C]flumazenil to quantify benzodiazepine receptor availability. The radioligand was administered either through bolus injection or bolus infusion, with emission data captured over 90 min. Structural MRI scans, aligned to a standardized atlas, adhered to acquisition protocols described by Nørgaard et al.²⁹. Voxel-wise maps of mean receptor density were derived and segmented into regions of interest using the Schaefer 100 cortical and Tian 32 subcortical atlases, yielding a 132*1 brain receptor expression matrix. Full methodological details on the acquisition of GABAa receptor distribution can be found in the original publication by Hansen et al.⁷¹.

Energy calculations

NCT enhances our comprehension of how white matter (WM) architecture shapes brain dynamics, enabling the calculation of the minimum control energy needed for state transitions and maintenance based on WM connectivity³³. In this study, we leveraged NCT to explore two objectives: 1) to determine if CID exists reductions in the minimum control energy required for brain state transitions, and 2) to evaluate whether control energy landscapes correlate with the spatial distribution of GABAa (Fig. 1b). We adopted a linear time-invariant model expressed as:

$$\stackrel{°}{x}\left(t\right)=Ax\left(t\right)+Bu\left(t\right)$$

Here, x(t) represents regional brain activity at time t, A is an N × N structural connectome matrix with N denoting the number of brain regions, B specifies the control input weights indicating the energy injectable into each region, and u(t) denotes the temporal profile of control energy applied. We computed the control energy under two distinct scenarios by varying the B matrix. In Model 1, applied to both HC and CID groups, we assumed uniform control inputs across all 132 regions, setting diagonal elements of B to 1 (identity matrix). In Model 2, exclusive to the CID group, control inputs were weighted by GABAa distribution across the same 132 regions. Using Model 1, we assessed differences in minimum control energy requirements for state transitions between HC and CID groups. Subsequently, within the CID group, we compared Models 1 and 2 to examine how GABAa’s spatial pattern influences the control energy. Details on the computation of minimum control energy are provided in the Supplementary materials.

Statistics and reproducibility

To compare FO and ALT between patients with CID and HCs, we utilized linear regression models. These models included group (CID vs. HCs), age, sex, and the age × sex interaction as fixed effects, with FO and ALT for each state as dependent variables. Before analysis, dependent variables underwent inverse normal transformation⁷². We conducted exploratory correlation analyses to examine associations between self-rated Pittsburgh Sleep Quality Index (PSQI) scores and the FO and ALT of spatial states showing significant group differences. Pearson correlation was used when both variables were approximately normally distributed (assessed via the Shapiro–Wilk test), and Spearman’s rank correlation was applied when normality assumptions were not met. All p-values were corrected for multiple comparisons using false discovery rate (FDR) adjustment. State transition probabilities, averaged across participants, were compared between CID patients and HCs. A 20% threshold identified the most frequent transitions. The Network-Based Statistics toolbox (version 1.2)¹³ was employed to detect networks of state transitions significantly more prevalent in CID patients compared to HCs, and vice versa. Specifically, we used the NBSglm function to perform mass-univariate general linear model (GLM) analyses at each connection across the transition probability matrices. Each connection was tested using a two-sample t-test, implemented within the GLM framework. To assess statistical significance at the network level, we performed 5000 non-parametric permutations, randomly shuffling group labels to build an empirical null distribution of the maximal subnetwork size. Family-wise error rate (FWER)-corrected p-values (p_FWE) were then computed based on this null distribution to identify significant group differences in transition network topology¹³. For between-group differences in control energy metric and complex index, we utilized linear regression models, with age, sex, and the age × sex interaction as covariates. For within-group differences of the CID group in the control energy metric, a two-sided paired t-test was used. All p-values were corrected for multiple comparisons with FDR, where correction is indicated. The GABA-weighted inputs from the true receptor distribution were compared with the randomly shuffled (132 regions) via a spatial permutation test, and the energy matrix was recalculated 10,000 times. P values (p_spin) were calculated as the fraction of times that the randomized distribution resulted in a higher energy than the true distribution. Additionally, the changed control energy in the CID group was calculated as GABAa-weighted control energy minus unweighted control energy. We then conducted Spearman correlation analyses to examine associations between transition probabilities and changed control energy, as well as associations between changed control energy and complexity index in the CID group.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

Liyong Yu and Siyi Yu conceived and designed the study. Liyong Yu, Liang Gong, and Xiaoqin Chen wrote the manuscript. Yuqi He and Rong Li performed the study and collected materials. Siyi Yu, Qi Zhang, and Xiaojuan Hong helped coordinate the study and reviewed the manuscript.

Peer review

Peer review information

Communications Biology thanks the anonymous reviewers for their contribution to the peer review of this work. Primary Handling Editor: Jasmine Pan. A peer review file is available.

Data availability

The raw fMRI data used in this study are not publicly available due to privacy and ethical restrictions. The processed fMRI data supporting the findings of this study are available at Figshare (https://figshare.com/articles/dataset/SourceData_CID_States/29328695)⁷³. The diffusion MRl are publicly available at NITRC²⁸. The GABAa receptor data was publicly available at neuromaps (https://github.com/netneurolab/neuromaps)⁷¹.

Code availability

The scripts used for HMM analyses are available at: https://github.com/OHBA-analysis/HMM-MAR. The complexity index was calculated using the seqici() function from the TraMineR R package: https://traminer.unige.ch/. NCT analyses were performed following procedures described in a previous study²⁰, with code available at: https://zenodo.org/records/6968138.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s42003-025-08439-4.