Chronic insomnia (CI) is a prevalent sleep disorder characterized by persistent difficulties in initiating or maintaining sleep, accompanied by daytime impairments. Epidemiological studies indicate that CI affects approximately 10–15% of the global adult population, representing a significant public health burden (1, 2). The condition of CI patients can significantly affect their cognitive function, emotional regulation, and overall quality of life. Additionally, the severity of insomnia correlates with an increased risk of developing comorbid psychiatric disorders and cardiovascular diseases (3, 4). In clinical practice, symptoms of CI are commonly alleviated by pharmacological agents such as benzodiazepines and Z-drugs (5), or managed through cognitive behavioral therapy for insomnia (CBT-I) (6). However, the complex and heterogeneous etiology of CI, combined with the potential for side effects, dependency, or limited accessibility of these standard treatments, often undermines long-term therapeutic efficacy and patient adherence (7, 8). Therefore, it is imperative to explore alternative and complementary interventions to more effectively manage this disorder.

Various viable alternatives have been investigated for the treatment of CI, among which electroacupuncture (EA) has been increasingly adopted for its holistic regulatory approach and favorable safety profile (9–11). In China, EA is recommended as a complementary therapy for sleep disorders, including insomnia (12). Neuroimaging and electrophysiological studies have demonstrated that EA can modulate dysregulated neural circuits and oscillatory activity associated with sleep–wake cycles, thereby facilitating improvements in sleep architecture and subjective sleep quality (13, 14).

To elucidate the electrophysiological basis of this modulation, techniques such as EEG have become important research tools. Recent research has employed periodic and aperiodic EEG features to assess the modulatory effects of acupuncture on brain dynamics. For instance, parameterizing EEG power spectra into periodic components (e.g., adjusted alpha power) and an aperiodic component (i.e., the aperiodic exponent) allows for a finer-grained analysis of acupuncture-induced neural changes. This approach provides a more comprehensive evaluation beyond traditional band-power analysis by separately quantifying oscillatory activity and background 1/f dynamics (15). Based on functional network analysis utilizing phase synchronization metrics such as the phase lag index, acupuncture has been shown to significantly enhance whole-brain synchronization within specific frequency bands—particularly in the δ and α bands—and to optimize the topological architecture of functional networks. This is specifically manifested in the strengthening of their small-world network properties, thereby improving the efficiency of information integration and transmission (16). Furthermore, advanced methods like low-dimensional neural manifold decoding have been used to characterize brain state transitions under acupuncture stimulation, exploring this within the framework of an “acupuncture-brain interface.” (17) For example, some studies have attempted to use deep learning models (e.g., Transformer networks) to directly identify specific acupuncture manipulation states from scalp EEG (18). A functional magnetic resonance imaging meta-analysis focusing on mild cognitive impairment has also provided spatial evidence supporting acupuncture’s modulatory effects on key brain regions and networks (19). Collectively, this multi-faceted evidence suggests that acupuncture possesses the potential to reshape brain network dynamics.

It is also noteworthy that individual differences in resting-state brain network dynamics are associated with clinical symptoms and treatment responses in sleep disorders, as shown by previous neurophysiological examinations (20, 21). Recent research indicates that EEG microstates—brief, recurrent patterns of global neural synchronization—can reflect these large-scale network dynamics and may distinguish between different clinical profiles (22, 23). EEG microstates are characterized as semi-stable, topographical maps of scalp voltage that persist for tens of milliseconds before rapidly transitioning, representing the coordinated activity of distributed neural assemblies (24–27). On the relationship between EEG microstates and the treatment of neuropsychiatric conditions, studies have shown that microstate parameters can change following therapeutic interventions (28–30). Additionally, other research has indicated that baseline microstate features may predict response to treatments such as antidepressants (31, 32). This implies that EEG microstates hold promise as potential biomarkers for assessing and predicting the effectiveness of therapeutic interventions.

Existing studies have preliminarily identified abnormalities in EEG microstates among patients with CI. Compared to healthy controls (HC), individuals with insomnia exhibit shorter durations of microstate B, lower occurrence and coverage of microstates B and C, as well as higher occurrence and coverage of microstates D and E (33). However, another study reported reduced occurrence and contribution of microstate D (34). These findings remain largely descriptive and inconsistently replicated across studies, with the underlying relationship between such microstate alterations and response to clinical interventions yet to be elucidated. Furthermore, although multiple randomized controlled trials have demonstrated the efficacy of acupuncture in improving insomnia, its specific mechanisms of modulating brain network dynamics remain poorly understood. Crucially, there is a lack of objective neurophysiological markers capable of predicting individual treatment responses prior to intervention.

In this context, this study proposes an integrative research framework combining EEG microstate analysis with machine learning (ML) modeling to preliminarily explore the association between EA and changes in brain microstate dynamics in patients with CI and to attempt to construct an exploratory predictive model based on baseline EEG features. We hypothesize that CI patients exhibit characteristic microstate abnormalities, that EA can modulate these aberrant dynamics, and that baseline microstate parameters can effectively predict clinical response. By comparing resting-state microstate features between patients and HC, analyzing changes in microstate parameters before and after treatment, and employing multi-stage feature selection and ML algorithms for exploratory modeling, this study aims to provide preliminary data for understanding the relationship between EA intervention and brain dynamics, and to offer a reference for future exploration of neural predictors for individualized therapy.

By leveraging a combination of EEG microstate analysis and ML, the present study aimed to address two core objectives: to investigate the neurophysiological changes associated with EA treatment for CI before and after intervention, and to preliminarily develop a predictive model for identifying individualized treatment responses. Our key findings revealed three pivotal observations: First, patients with CI exhibited markedly abnormal baseline temporal dynamics in microstates, characterized by significantly higher coverage and occurrence of microstate A, as well as a higher occurrence of microstate C, along with significantly lower coverage and duration of microstate B. Additionally, dynamic abnormalities in microstate transitions were observed in these patients, which may be associated with dysregulated resting-state brain networks in this population. Second, during the four-week EA intervention, the coverage and duration of microstate B showed a trend toward normalization to the levels observed in HC, while corresponding adjustments were also noted in the temporal characteristics of microstates C and D. Third, baseline microstate features, particularly microstate durations and specific inter-microstate transition probabilities, demonstrated predictive potential for treatment response as measured by PSQI score improvement in this study. Collectively, these findings offer new observational perspectives on the neurophysiological changes accompanying EA treatment for CI, while also suggesting that microstate parameters may have potential biomarker value for predicting individualized treatment responses.

This study employed the classic A–D four-class microstate model for analysis. This choice was based on dual considerations of methodological maturity and alignment with the research objectives. First, this model represents the most classic and widely adopted analytical framework in resting-state EEG microstate research. According to existing literature, the four canonical microstate classes are thought to potentially map onto the activity of large-scale brain networks; for instance, microstates A, B, C, and D are often associated with the auditory network, visual network, SN, and attention network, respectively (29, 52, 53). Research indicates (34, 54, 55) that the four-class model can effectively capture the core dynamic patterns of large-scale brain networks, providing a foundation for its applicability in this study (56). Second, to develop a ML model that is robust, interpretable, and holds potential for clinical prediction, utilizing the well-validated four-class features as inputs helps ensure the comparability of analytical results and provides a clear neuroscientific basis for interpreting the model. In the present study, we observed multi-dimensional abnormalities in the temporal parameters of microstates in patients with CI. These aberrant patterns may collectively constitute the characteristic disturbances in resting-state brain network dynamics associated with CI. Specifically, CI patients exhibited increased coverage and occurrence of microstate A. Given that microstate A is frequently associated with the ventral attention network (VAN) in the literature (57)—a network whose functions include external environmental perception and attentional orienting—we speculate that this abnormality may reflect a “hyperarousal” neural signature specific to CI. This may suggest that, even in the eyes-closed resting state, the brains of CI patients may remain in a state of heightened vigilance, unconsciously enhancing the processing of internal and external sensory signals (58, 59).

Meanwhile, compared with healthy participants, CI patients showed reduced coverage and duration of microstate B, alongside hyperactivation of microstate C as evidenced by its increased occurrence. This pattern aligns with previous research (33, 60): following sleep deprivation, microstate B was suppressed and microstate C exhibited hyperactivation, with the occurrence rate of microstate B showing a significant negative correlation with Subjective sleepiness (e.g., Karolinska Sleepiness Scale scores). Furthermore, the increased occurrence rate of microstate C in patients with CI may be related to its overactivation, which is implicated in processing personally salient information and self-referential thinking (26). This elevation in microstate C occurrence is also associated with functional abnormalities in the DMN and the executive control network (ECN)—a finding supported by the work of Khoo S. Y. et al. (61). It is noteworthy that reduced duration of microstate B has been linked to prolonged reaction times in psychomotor vigilance tasks, increased subjective/objective sleepiness, and decreased positive affect, suggesting that it may serve as a potential neurodynamic indicator linking nocturnal sleep disturbances to impaired daytime cognitive functions, such as attention. Prior studies have indicated that microstate C exhibits a competitive relationship with other microstate classes, suggesting it plays a unique role in the dynamics of canonical microstate transitions (33, 62). Additionally, we found that patients with CI showed a significant increase in transition probability from microstate A to C and a significant decrease in transition probability from microstate B to D, which may indicate a dysregulation in the dynamic coordination of brain networks in this population (34, 63).

Enhanced A → C transitions may reflect abnormal coupling between the SN and DMN, potentially leading to a tendency toward abnormal switching in CI patients between internal introspective states (e.g., DMN dominance) and states of readiness to respond to external stimuli (e.g., involving the SN and central executive network) (64). In contrast, reduced B → D transitions may suggest decreased efficiency in the brain’s dynamic shift from sensory processing to cognitive control. Collectively, these microstate anomalies may represent one potential neurophysiological substrate underlying clinical manifestations of CI, such as impaired cognitive flexibility and attentional regulation deficits (65).

Our study found that after 4 weeks of EA, scores on the PSQI, ISI, FSS, HAMA, HAMD, and HAS were significantly improved—this is consistent with our previously reported significant efficacy of EA in treating CI (9, 66, 67). However, the mechanism underlying EA’s therapeutic effect in CI requires further investigation. Our team previously reported that after EA treatment, CI patients exhibited reduced high-frequency EEG activity, increased low-frequency EEG activity, and decreased entropy values, suggesting that modulating EEG signals in CI patients may be one of the mechanisms by which EA improves sleep (9). Nevertheless, whether and how EA modulates EEG microstates—a key indicator reflecting the dynamic coordination of large-scale brain networks—remains unclear. The present study provides new electrophysiological observations for this: after 4 weeks of EA treatment, the originally abnormal parameters of microstate B (coverage and duration) in CI patients showed a trend toward normalization, approaching levels observed in HC. Microstate B is primarily associated with the DMN and visual network, and typically dominates during rest, introspection, and relaxation. The occurrence rate of microstate C, which is often associated with the SN and attentional network activity, was altered following EA treatment. Building on this association, the changes in microstate B parameters suggest that EA intervention may facilitate a shift in brain activity toward a state more inclined to internal focus and relaxation, thereby potentially creating conditions conducive to sleep initiation. The acupuncture point prescription employed in this study (Baihui GV20, Yintang EX-HN3, bilateral Shenmen HT7, and Sanyinjiao SP6) was selected based on traditional Chinese medicine (TCM) theory and modern neuroanatomical evidence. This multi-site combination has been widely used in previous studies for insomnia treatment, and its effects may be associated with the modulation of multiple neurophysiological processes involved in sleep–wake regulation (68). Based on TCMtheory, Baihui (GV20) and Yintang (EX-HN3) are located on the Governor Vessel and extraordinary points, respectively, and are commonly used for calming the spirit, stabilizing the mind, and regulating mental activity. Neuroimaging studies suggest that stimulation of these frontal scalp points can modulate activity in the prefrontal cortex and limbic system (69). Shenmen (HT7), located on the Heart Meridian, is a key point for nourishing the heart and calming the spirit. Studies have shown that stimulating this point can influence autonomic balance and emotional processing (70). Sanyinjiao (SP6) is the confluence point of the Liver, Spleen, and Kidney Meridians. It is used to nourish Yin and subdue Yang, aiming to correct the underlying imbalance considered in TCM theory to be responsible for insomnia (71). This combination of acupoints is designed to synergistically regulate the spirit, mental activity, and systemic balance, which corresponds to the multi-network modulation observed in the brain. Animal studies have demonstrated that EA at GV20, HT7, and SP6 may modulate the activity of the hypothalamic–pituitary–adrenal axis, affect various monoamine neurotransmitters (such as 5-HT), and influence neurotrophin pathways. Furthermore, a recent randomized controlled neuroimaging study also found that acupuncture can modulate abnormal functional connectivity between the hypothalamus—a core structure of the sleep–wake cycle—and the orbitofrontal cortex in patients with CI, and this change in connectivity is correlated with clinical improvement (72). One study found that acupuncture may improve clinical symptoms in patients with chronic insomnia by modulating the expression of specific genes related to neuroinflammation, thereby remodeling the functional connectivity of key brain regions, such as the striatum and prefrontal cortex (73). The treatment protocol utilized the even reinforcing-reducing needle technique combined with 2 Hz continuous-wave EA stimulation, which has frequently been reported in previous studies to be associated with sedative effects and the modulation of low-frequency neural oscillations (74–76). Thus, this multi-target stimulation protocol with specific parameters may be more inclined to exert multi-level modulation on the distributed brain network systems that maintain the balance between wakefulness and sleep (77, 78). This could thereby be associated with the clinical symptom improvements and microstate dynamic changes observed in the present study. Of course, this network modulation interpretation based on microstate changes remains a correlative finding and hypothesis; its definitive causal relationships and precise neural pathways require further in-depth validation in future studies incorporating control groups and more advanced multimodal techniques.

Although symptom alleviation was observed based on PSQI assessment following EA treatment, there remained significant inter-individual variability in therapeutic efficacy. Therefore, distinguishing potential responders before treatment is valuable for exploring individualized strategies. In recent years, studies have attempted to predict treatment responses in insomnia patients using baseline EEG features (51). For instance, Zhu Lin et al. (79) found that, among 36 patients with primary insomnia, the ML model based on baseline theta connectivity achieved the highest accuracy in predicting treatment response to repetitive transcranial magnetic stimulation (rTMS; AUC = 0.843). Another study by Yongjian Guo et al. (34) found that transition probabilities between microstates were useful for identifying patients with potential high responsiveness to repetitive rTMS treatment, with the final model achieving an average prediction accuracy of 80.13%. Following this line of reasoning, we exploratorily developed for the first time a predictive model for EA efficacy based on baseline microstate features. Among the models tested, the RF algorithm exhibited the optimal classification performance in this dataset (AUC = 0.849). Key predictive features included Duration_A, OrgTM_D → B, and OrgTM_C → B. These preliminary findings suggest that pre-treatment resting-state brain activity may contain information relevant to subsequent subjective treatment outcomes, potentially enabling the identification of patient subgroups most likely to benefit from EA. The ML methodology employed herein demonstrates distinct advantages over traditional statistical approaches in modeling high-dimensional, nonlinear relationships among microstate parameters, thereby enhancing predictive precision. The Precision-Recall curve is highly sensitive to class imbalance. If the performance improvement were merely due to overfitting, the PR-AUC should not have shown such a substantial leap. This result confirms that oversampling enabled the model to genuinely learn the characteristic patterns that distinguish responders from non-responders. SHAP analysis further improved model interpretability by quantifying feature contributions: Duration_A, OrgTM_D → B, and OrgTM_C → B emerged as critical positive predictors, whereas Occurrence_C and Duration_B served as negative predictors.

The preliminary results of this predictive framework may provide a preliminary theoretical reference for future exploration into pre-treatment prognostic assessment using neurophysiological indicators, yet its clinical translational potential is currently severely limited by the small sample size, the potential overestimation of model performance due to bootstrap oversampling, as well as the lack of external multi-center validation and the exclusive use of subjective PSQI scores (without objective polysomnography (PSG) sleep measurements) to define treatment response. For patients predicted to potentially respond poorly, future studies could explore the necessity of adjusting treatment plans—for instance, by adopting combined strategies such as integrating EA with pharmacotherapy or CBT-I, or implementing sequential multimodal interventions—in well-characterized cohorts with large sample sizes and objective sleep assessments.

While this study yielded several meaningful findings, it is not without limitations. First, this study adopted an open-label, single-arm, before-and-after comparison design without including a sham acupuncture or waitlist control group. Therefore, although significant improvements temporally associated with the EA intervention were observed, the potential influence of placebo effects, regression to the mean, or non-specific treatment factors cannot be fully excluded. Future studies require randomized, double-blind, placebo-controlled trial designs to confirm the specific efficacy of EA. Second, the machine learning predictive analysis in this study is strictly exploratory and preliminary in nature. The constructed prediction model is based on single-center data with a limited sample size, which directly leads to a significant risk of overfitting and severely restricts the model’s stability and generalization ability. Although bootstrap oversampling was employed to address class imbalance, this strategy may have potentially overestimated the model’s performance evaluation metrics (such as AUC and PR-AUC). Moreover, the model has not undergone any external multi-center validation, making its direct application to clinical practice impossible at this stage. Additionally, the definition of treatment responders relies entirely on subjective PSQI score reduction (≥50%) without incorporating objective polysomnographic sleep measures (such as sleep efficiency, total sleep time, and number of awakenings). This may introduce subjective bias into the classification outcomes and limit the reliability of the predictive model. This point is indirectly reflected in the performance variability of the model, as indicated by the large standard deviation of the reported AUC values (e.g., 0.849 ± 0.133 for the random forest model). Therefore, the clinical applicability of the current prediction model remains unclear. Future studies must utilize larger, multi-center datasets to validate the robustness of the model and consider incorporating objective PSG-derived metrics into the definition of treatment response to enhance its objectivity and reliability. Third, this study focused solely on resting-state EEG microstate features and did not incorporate EEG activity during sleep or acupuncture treatment sessions. Furthermore, the analysis primarily addressed linear microstate characteristics (e.g., duration, occurrence, and transition probabilities), overlooking their non-linear dynamic attributes. Subsequent research could integrate PSG recordings and EEG data collected during acupuncture sessions to comprehensively assess microstate dynamics across different sleep stages. Additionally, incorporating non-linear metrics of microstates (e.g., entropy rate, Lempel–Ziv complexity) could provide complementary insights. More importantly, the brain’s response to acupuncture is multifaceted. An analytical framework that extends beyond a single modality may reveal more complete mechanisms. Future studies could significantly enhance predictive accuracy and mechanistic understanding by constructing a multimodal EEG biomarker framework. For example, integrating the microstate temporal dynamics examined here with complementary spatio-temporal features—such as the periodic-aperiodic parameters of neural oscillations, measures of functional and effective brain network connectivity, and functional magnetic resonance imaging (fMRI) data—could provide a more comprehensive picture of brain activity, ranging from local rhythms and network synchronization to global dynamics. Integrating these multidimensional biomarkers within advanced ML pipelines holds great potential for developing robust, individualized neuromodulation strategies for CI. Finally, the definition of a “responder” as a participant with a ≥ 50% reduction in PSQI score, while clinically conventional, remains somewhat subjective. Future work may improve objectivity by incorporating PSG-derived measures—such as sleep efficiency and the number of awakenings—into the criteria for defining treatment response.

This study revealed EEG microstate dynamics associated with CI, characterized by significantly higher coverage and occurrence of microstate A, increased occurrence of microstate C, and notably lower coverage and duration of microstate B, along with dynamic abnormalities in microstate transitions. EA treatment was associated with changes in these abnormal patterns, with the parameters of microstate B showing a trend toward normalization. Based on these findings, an exploratory predictive model was constructed to identify responses to EA treatment for CI. This work provides preliminary clues for understanding brain dynamic responses under EA intervention and exploratorily explores the potential feasibility of using neurophysiological indicators for efficacy prediction, with no clinical application value at present. Future research is needed to validate these findings with more rigorous designs, larger multi-center sample sizes and objective sleep measurements, and to further investigate the underlying neural mechanisms.