Electroencephalography microstates as novel functional biomarkers for insomnia disorder

Participants

This study was conducted at the Second Hospital of Hebei Medical University in Shijiazhuang, China. Good sleep controls (GSC) and ID patients were recruited from outpatient clinics. Individuals aged between 18 and 65 years who were right-handed were screened based on the diagnostic criteria for ID in the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition. A total score of ≥5 on the Pittsburgh Sleep Quality Index (PSQI) and ≥8 on the Insomnia Severity Index (ISI) was required for enrolment. Those with severe neurological or medical illness, other sleep disorders, restless leg syndrome and pregnancy (for females) were excluded. Additionally, individuals who received psychiatric medication, hypnotics or treatment with cognitive behavioural therapy for insomnia within the past 2 weeks were also excluded. For GSC, age-matched and gender-matched individuals were included based on the following criteria: no symptoms or history of psychiatric or sleep disorders, a total PSQI score of <5 and a total ISI score of <8 at the time of screening and no lifetime use of any psychotropic or hypnotic medication. For both groups, exclusion criteria included comorbid other psychiatric disorders and Beck Depression Inventory (BDI) scores >20 or Beck Anxiety Inventory (BAI) scores >45. To screen for other sleep disorders, participants were questioned in person about conditions including obstructive sleep apnoea syndrome, restless leg syndrome, etc. Also, the PSG data of the participants were collected and analysed. Individuals with other sleep disorders were excluded from the study.10

Two datasets were included in this study. The primary dataset included 40 GSC and 60 patients with ID. Patients with ID were randomly divided into active rTMS treatment (n=30) and sham rTMS treatment (n=30) groups. This dataset was used to explore abnormal brain microstates in patients with ID and to assess the effect of rTMS intervention on abnormal microstates in patients with ID. The second dataset included 90 patients who received active rTMS treatment; it was used to validate whether baseline microstates in patients with ID could predict the outcomes of 20-day active rTMS treatment. The inclusion and exclusion criteria for ID patients recruitment in the second dataset were consistent with those in the primary dataset. A flowchart of participant registration and follow-up is shown in figure 1. The characteristics of the participants are summarised in online supplemental table S1 and online supplemental table S2.

Request permission

Repetitive transcranial magnetic stimulation treatment

The rTMS therapy was carried out by using a pulsed magnetic stimulation device (M-100 Ultimate; Shenzhen Yingchi Technology, Shenzhen, China) with a figure-eight stimulation coil at the left DLPFC, as described in our previous study.10 The specific parameters were as follows: stimulus frequency at 1 Hz, stimulus intensity at 80% of the motion threshold, stimulation number at 10 pulses per string and 150 strings, string interval at 2 s, and total stimulation pulses at 1500. For each rTMS session, the total stimulation time is 30 min, one session per day. Sham rTMS was performed by orienting the coil away from the skull at 90°. Twenty active or sham sessions of left DLPFC rTMS treatment were delivered over four consecutive weeks, with a frequency of five times per week.

Clinical assessment

All patients with ID received pre-treatment and post-treatment clinical assessments, including full-night sleep PSG recordings using a Grael 4 K system (Compumedics, Victoria, Australia) and sleep-related scales. The Yet Another Spindles Algorithm (YASA) was used for sleep stage classification, and PSG metrics were calculated, including sleep onset latency (SOL), sleep efficiency (SE) and non-rapid eye movement sleep stage 3 (NREM 3) duration. The scales included the PSQI, ISI, Epworth Sleepiness Scale (ESS), BDI, BAI, Mini-Mental State Examination (MMSE) and the Montreal Cognitive Assessment (MoCA). GSC participants were subjected to the same assessment but only measured at baseline (ie, pre-treatment).

Resting-state electroencephalography data acquisition and preprocessing

The 10 min resting-state EEG recording was carried out in a shielded, sound-attenuated room. Participants were instructed to remain awake with their eyes closed, and relax without any active thinking. Resting-state EEG data were acquired at a sampling rate of 512 Hz (analogue bandpass filtering: 0.1 and 100 Hz), with a 19-channel Nicolet system (Cephalon, Denmark). The ground electrode was attached to the forehead, and reference electrodes were located on A1 and A2. For each electrode, the impedance was maintained below 10 kΩ. EEGLAB toolbox and custom MATLAB (MathWorks, Massachusetts, USA) scripts were employed for EEG data offline preprocessing.21 By using a notch filter, we removed the 50 Hz AC line noise artefact. After filtering with a 2–20 Hz bandpass, the filtered EEG data were re-referenced to the common average. The bad channels and bad epochs were manually rejected and then interpolated from the EEG signals of adjacent channels. Finally, the eye movement and blink artefacts were removed using independent component analysis.

Microstates analysis

The microstate analysis was conducted using the functions from the EEGLAB plugin for microstates (http://www.thomaskoenig.ch/index.php/software/Microstates-in-eeglab) in MATLAB R2021b. The process of EEG microstates analysis is described in detail below (online supplemental figure S1).

First, the global field power (GFP) was calculated:

Where k is the number of electrodes in the EEG data (19 in this case);    is the potential of the ith electrode at a certain time point and    is the average value of the instantaneous potential across electrodes.

Based on the above equation, a GFP curve that reflects the degree of change in the EEG potential between all electrodes at a given time can be obtained. GFP curve local maxima represented instants of the highest field strength.22 EEG topographies tend to remain stable during high GFP periods and change rapidly around the local minima of the GFP. Therefore, topographies at GFP peaks are representative of topographies at surrounding time points, and restricting the microstate analysis to these GFP peaks provided optimal topographic signal-to-noise ratios.23

Second, for the resting-state recording, all topographies at GFP peaks were selected and submitted to modified k-means clustering (despite spatial polarity) with 100 repetitions. The number of clusters was set to 4, and microstate maps (ie, cluster centres) were estimated. Third, a second cluster was performed among the subjects within each group. All microstate maps were submitted to modified k-means clustering (despite spatial polarity) with 100 repetitions. Finally, the group microstate maps were then fit back to the original data at GFP peaks, assigning each GFP peak to one microstate class based on the maximal spatial correlation between topographies. Microstate labels for data points between GFP peaks were interpolated with microstates starting and ending halfway between two GFP peaks. Potentially truncated microstates at the beginning and end of each epoch were excluded from the analysis.

After obtaining the microstates, the corresponding EEG microstates temporal parameters were extracted: (a) Duration_X: the mean duration of microstates of class X, in seconds; (b) Mean Duration: the mean microstate duration across all classes in seconds; (c) Occurrence_X: the mean frequency of observation of microstates of class X, per second; (d) Mean Occurrence: the mean frequency of observation of microstates across class X, per second; (e) Contribution_X: the proportion of the total time spent in microstates of class X; (f) Mean GFP_X: the mean GFP of microstates of class X in microvolts; (g) OrgTM_X->Y: the proportion of all observed microstate transitions from X to Y; (h) ExpTM_X->Y: the proportion of expected microstate transitions from X to Y, given only the observed occurrences; (i) DeltaTM_X->Y: the difference between the observed and the expected transition probabilities. Details are available in the online supplemental material.

Statistical analyses

The Shapiro-Wilk test was used to test for the normality of distribution. Independent samples t-test was used to compare the differences between the ID and GSC groups (basic information, scale scores, microstate indicators and PSG indicators). The Wilcoxon rank-sum test was performed on the gender of participants in the ID and GSC groups. Two-way repeated measure analysis of variance and post hoc Tukey tests were used for differences before and after active and sham rTMS (scale scores, PSG indicators and microstate indicators).

The topographies of the different microstate classes between the groups were compared using topographical analysis of variance (TANOVA) implemented in the Ragu software.24 TANOVA was based on powerful and assumption-free randomisation statistics and could be used for the statistical comparison of EEG scalp field maps between two or more conditions.

Pearson’s correlation analysis was used to explore the potential significant correlation between abnormal microstate parameters and subjective and objective sleep quality in patients with ID. Data were analysed using SPSS V.26.0 (SPSS, Chicago, Illinois, USA). Graphs were plotted using GraphPad Prism V.8.0.2 (GraphPad Software, California, USA).

Construction of efficacy prediction model

PSQI scale scores were collected from patients with ID before and after treatment. Then, the percentage of improvement in the subjects’ PSQI scale scores was calculated using the formula shown below:

The percentages of improvement were sorted in ascending order, and the subjects were divided into optimal and suboptimal groups using the median 54 baselines. EEG microstate characteristics (Duration_X, Mean Duration, Occurrence, Mean Occurrence_X, Contribution_X, Mean GFP_X, OrgTM_X->Y, ExpTM_X->Y, DeltaTM_X->Y) were used as the features to build the dataset. Details of the characteristics are given in the ‘microstate analysis’ section. The dataset (10n) was divided into a training set (9n) and a test set (1n) using a hierarchical 10-fold cross-validation strategy. First, feature preprocessing of normalisation was performed for the features, and then principal component analysis dimensionality reduction was carried out to remove redundant features. The specific strategy retained 99% of the explainable variance. The remaining features were then input to a logistic regression classifier for model training. The test set was input to the trained model according to the same feature preprocessing and feature selection strategies for model evaluation. The performance metrics (accuracy, precision, recall, F1-score, area under the curve (AUC)) were calculated for each fold of the model, and the average performance metrics were averaged for each fold of the model to obtain the average performance metrics of the model.

Demographic information and behavioural results

Demographic information and behavioural results of the ID and GSC groups are presented in online supplemental table S1. There were no significant differences in age, gender, ESS or MoCA between the two groups (p>0.05). Significant differences were observed in PSQI, ISI, BDI, BAI and MMSE scores between the two groups (p<0.001).

Of the total 60 participants, one in the GSC group withdrew for personal reasons. PSG data from seven active and six sham group patients were excluded from the analyses due to poor data quality. The final PSG data used for analysis included 23 participants in the active treatment group and 24 in the sham treatment group. Significant impairments in both subjective sleep quality (PSQI: t₉₇=19.13, p<0.001; ISI: t₉₇=18.53, p<0.001) and objective sleep quality (SE: t₈₄=6.324, p<0.0001; NREM 3 duration: t₈₄=6.828, p<0.001; SOL: t₈₄=4.774, p<0.001) were observed in patients with ID when compared with the GSC group at baseline (online supplemental figure S2). No side effects were reported during or after brain stimulation.

Active rTMS treatment effectively enhanced sleep quality in patients with ID

Following active rTMS treatment, there was a noticeable improvement in subjective sleep quality in patients with ID, as indicated by reduced PSQI and ISI scores (figure 2A). Similarly, improvement in objective sleep quality was evidenced by increased SE, NREM 3 duration and shorter SOL (figure 2B). In contrast, no significant changes in subjective or objective sleep quality were observed in the sham group (figure 2). No significant differences were observed in MMSE, MoCA, BDI and BAI scores in patients with ID, both before and after treatment with active or sham rTMS.

Request permission

Abnormal EEG microstate spatial topographies and temporal parameters in patients with ID

The microstate analysis results demonstrated that microstate classes A–D corresponded well to the canonical microstate maps reported in the literature (figure 3A).25 No statistical difference was detected in the global explained variance between the ID group (mean (standard deviation, SD): 0.79 (0.06)) and the GSC group (mean (SD): 0.77 (0.06)) (t₈₃=1.713, p=0.090). The topographies of microstates A and C were significantly different between the ID and GSC groups (p=0.003, p=0.001). Additionally, at baseline, there were significant differences between patients with ID and the GSC in the occurrence (t₈₃=2.999, p=0.021) and contribution (t₈₃=0.0003, p=0.003) of class D microstate (figure 3B, online supplemental table S3). Furthermore, we found that the occurrence (n=23, p<0.0001, r=−0.74) and contribution (n=23, p<0.0001, r=−0.72) of class D microstate were negatively correlated with SOL (figure 4B). The differences of the other microstate features did not reach significance.

Request permission

Request permission

rTMS reversed the partial abnormal microstate parameters in patients with ID

The independent sample t-test result showed no statistical difference in the global explained variance between the ID group after active treatment (mean (SD): 0.77 (0.06)) and the GSC group (mean (SD): 0.77 (0.06)) (t₆₂=0.1723, p=0.860). After active rTMS treatment, the difference between patients with ID and the GSC on the microstate C topographic map became no longer significant (p=0.604, figure 3A). The occurrence and contribution of class D microstate in the ID group were improved, approaching the levels observed in the GSC group (t₆₂=0.1913, p=0.849 figure 4A). However, no such improvement was found in the sham group. The change in the occurrence of the class D microstate also showed a significant negative correlation with the improvement of SOL (n=23, p=0.012, r=−0.51, figure 4B) in the active rTMS group.

Treatment outcome prediction from baseline EEG microstates

Here, the EEG microstate parameters (Duration, Incidence, Contribution, MeanGFP, OrgTM, ExpTM, DeltaTM) of 86 ID patients (the EEG data of 4 patients were excluded due to quality issues) were downscaled using PCA, and then a ten-fold cross-validation strategy was implemented to partition the dataset, which was then fed into a logistic regression classifier for training and testing.

The final average prediction accuracy was 80.13% (figure 5A). The top 10 most important features for the prediction model are shown in figure 5B. The average AUC area was 0.81, and the average F1 score was 0.62. The specific performance of each fold of the efficacy prediction model (accuracy, precision, recall, F1 score and AUC) is shown in online supplemental table S4.

Request permission

Main findings

The findings of this study have several important implications. First, we observed that the baseline occurrence and contribution of class D microstate were significantly lower in the ID group compared with the GSC group. The topographic maps of microstates A and C in the ID group were significantly different from those in the GSC group, potentially serving as a novel biomarker for SOL for patients with ID. Second, we found that active rTMS treatment improved the occurrence and contribution of class D microstate and partially reversed the topography of microstate class C in patients with ID. Finally, our findings supported that the baseline microstate characteristics of patients with ID could predict the improvement of sleep quality following rTMS treatment.

The four canonical microstate classes exhibited spatial patterns of well-known resting-state networks.26 Specifically, intra-individual fluctuations of class A, B, C and D microstates were linked to the activation of the auditory, visual, salience and attention networks, respectively. Abnormal EEG microstates have been detected in various mental and neurological disorders, such as depression, schizophrenia and Lewy body dementia.27–29 These abnormalities might reflect a disruption of the intricate dynamic properties of the brain, leading to a loss of flexibility and adaptability that was crucial for normal brain functioning. Our study found a decreased occurrence and contribution of microstate class D in patients with ID when compared with the GSC group. This suggests that patients with ID may not get sufficient rest during the night, which can affect daytime functioning and impair both attention and cognitive control. It is worth noting that a previous study reported a shorter mean duration of class C microstate and a higher frequency of class D microstate in patients with ID.19 However, these results require further validation, considering the differences in the timing of resting-state EEG acquisition and the number of clusters. In addition, we employed PSG to assess the objective sleep quality in patients with ID, which was not included in the previous insomnia disorder EEG microstates research. Our correlation analysis showed that the baseline occurrence and contribution of class D microstate were correlated with the SOL in patients with ID (figure 4B). In summary, we extended this microstates analysis to reveal the pathology of ID and found that EEG microstates may serve as a biomarker for the sleep quality of patients with ID, especially for the SOL. These hypotheses need to be tested by including cognitive assessments in future studies.

In this study, we observed significant improvements in the sleep quality of patients with ID following rTMS treatment in both subjective and objective dimensions, which is in line with the remarkable efficacy of rTMS in the treatment of ID reported previously.30–33 However, the mechanism of rTMS in ID treatment requires further investigation. Our group previously reported that active rTMS may re-establish the disrupted cortico-hippocampal interactions in patients with ID.13 Meanwhile, we observed normalisation of the hyper EEG coherence in theta and alpha bands after rTMS treatment, with changes in theta band coherence correlating with changes in SE.8 The hyperarousal model hypothesised that the brains of patients with ID exhibited hyperactivity during both day and night,34 which may disrupt the transitions of EEG microstates. However, it remains unknown whether rTMS can improve the pattern of microstate abnormalities. EEG microstates can respond to the transformation of the overall collaborative pattern of functional brain activity. This study provided scientific evidence that the occurrence and contribution of abnormal class D microstate in patients with ID improved after active rTMS (figure 4A). Moreover, the change in the occurrence of microstate class D showed a significant negative correlation with the shortening of SOL, which may serve as an indicator of SOL changes induced by active rTMS in patients with ID.

Although rTMS can effectively improve sleep quality in patients with ID, the effectiveness of rTMS treatment varies largely between individuals.35 Therefore, the distinction between high-responders and low-responders to rTMS is a crucial consideration for subsequent treatment strategies. We used a machine learning method to predict efficacy in an independent dataset and found that baseline microstate indicators in patients with ID could accurately predict whether their sleep quality would significantly be improved after 1 month of rTMS treatment (figure 5). For low-responders to rTMS, changing treatment modality or adopting rTMS combination therapy (eg, rTMS in combination with medication or cognitive-behavioural therapy) may improve treatment outcomes.36 The top 10 most important features demonstrated that the transitions among different microstates contributed to the prediction model, suggesting that more attention should be paid to other microstate indicators in future studies, such as DeltaTM_X->Y and OrgTM_X->Y (figure 5B).

Limitations

This study had several limitations. First, we did not conduct a follow-up for patients with ID in this study, preventing us from assessing the long-term effects of rTMS on ID. Second, we did not fully explore the higher-order indexes of PSG or the relationship between the coupling of various frequency bands of PSG and the EEG microstates. Lastly, we used the MMSE scale for clinical assessment, which is not a comprehensive cognitive assessment.37 In future studies, we aim to address these limitations and further investigate the pathogenesis of ID as well as the mechanism of action of rTMS to provide additional insights and improve ID treatment strategies.

Conclusion

In conclusion, this study is the first to investigate the temporal dynamics and spatial topography of four typical EEG microstates in patients with ID before and after rTMS treatment. We also explored the relationship between microstates and objective sleep indicators in patients with ID using PSG. Our findings revealed the abnormal occurrence and contribution of class D microstate in patients with ID, which correlated with SOL. 20 days of 1 Hz rTMS over L_DLPFC could modulate the occurrence and contribution of the EEG class D microstate. Moreover, baseline microstate temporal indexes (eg, the transitions among different microstates) aided in selecting potential high-responders to rTMS treatment. These results provided new electrophysiological insights on the pathology of ID and the possible mechanisms of rTMS for ID treatment. They highlighted the clinical significance of EEG in guiding rTMS treatment for ID.