Assessment of rTMS treatment effects for methamphetamine addiction based on EEG functional connectivity

Abstract

Methamphetamine (MA) addiction leads to impairment of neural communication functions in the brain, and functional connectivity (FC) may be a valid indicator. However, it is unclear how FC in the brain changes in methamphetamine use disorder (MUD) after treatment with repetitive transcranial magnetic stimulation (rTMS). Thirty-four patients with MUD participated in this study. The subjects were randomized to receive the active or sham rTMS for four weeks. Subjects performed electroencephalography (EEG) examinations and visual analogue scale (VAS) assessments before and after the treatment. The FC networks were constructed and visualized, and then the graph theory analysis was carried out. Finally, machine learning was used to classify FC networks before and after rTMS. The results showed that (1) the active group showed a significant enhancement in connectivity in the beta band; (2) the global efficiency, local efficiency, and aggregation coefficient of the active group in the beta band decreased significantly; (3) the LDA algorithm combined with the beta band FC matrix achieved an average accuracy of 82.5% in distinguishing before and after treatment. This study demonstrated that brain FC could effectively assess the therapeutic effect of rTMS, among which the beta band was the most sensitive and effective frequency band.

Supplementary Information

The online version contains supplementary material available at 10.1007/s11571-024-10097-x.

Introduction

Methamphetamine (MA) is a widely used stimulant that imposes significant burdens on society and individuals. A recent investigation by the United Nations Office on Drugs and Crime in 2022 revealed that MA is the second most prevalent drug globally, after opioids (UNODC 2022). Nevertheless, dependence on MA may lead to more severe impairments in brain cognition and communication functions than opioid drugs (Feil et al. 2010; Khajehpour et al. 2019). Addressing drug addiction issues, repetitive transcranial magnetic stimulation (rTMS) has shown promise due to the precise targeting of specific brain regions (Diana et al. 2017; Salling and Martinez 2016). However, progress in applying rTMS for treating Methamphetamine Use Disorder (MUD) has been slow, primarily due to the lack of objectively quantifiable assessment techniques (Saini et al. 2022).

Previous imaging studies have demonstrated that rTMS can activate distant nodes within specific functional networks, resulting in sustained connectivity changes associated with clinical improvement (Philip et al. 2018). Apart from the global effects, rTMS induces cortical oscillations at the site of stimulation (Thut et al. 2011). This modulation can enhance attentional bias and processing in individuals with MUD by influencing the secretion of GABA and dopamine, without reported adverse effects (Chen et al. 2021; Ding et al. 2014). Functional magnetic resonance imaging (fMRI) during resting state has indicated that rTMS can produce either excitatory or inhibitory effects depending on the stimulation frequency and brain region (Grefkes et al. 2010; Watanabe et al. 2014). Despite these findings, the development of effective rTMS protocols for treating MA addiction has been sluggish. Currently, most assessments of rTMS intervention effects in MUD rely on subjective methods like questionnaires and cognitive ability tests, lacking objective and precise evaluation metrics or biomarkers (Saini et al. 2022). Hence, there is an urgent need for further research on objective assessment methods to advance addiction treatment approaches.

Electroencephalography (EEG) is an ideal tool for detecting changes in cortical neuron activity due to the simplicity and high temporal resolution. It has been extensively used in evaluating various cerebral disorders (Guo et al. 2016; Liu et al. 2023). The human brain consists of a complex network of functional regions that rely on information transmission for proper functioning (van den Heuvel and Pol 2010). Increasing evidence suggests that many psychiatric disorders exhibit severe disruptions in brain functional connectivity (FC) compared to healthy individuals (Bassett et al. 2012; Bunse et al. 2014; Cha et al. 2012; Hanson et al. 2013). Drug addiction, as a branch of psychiatric disorders, has also been associated with abnormal brain impairments (Devoto et al. 2020; Tolomeo and Yu 2022). In the field of addiction, research has been conducted on reward processing (Luijten et al. 2017), inhibitory control(Le et al. 2021; Luijten et al. 2014), executive function (Quaglieri et al. 2020), and other relevant neural mechanisms. These findings can support the diagnosis of disease-specific behaviors, such as craving, compulsive drug use, and relapse (Ceceli et al. 2022; Yang et al. 2022). In recent years, EEG-based FC networks have emerged as powerful tools for evaluating brain connectivity and interactions between brain regions. EEG-based FC networks have emerged as powerful tools for assessing brain connectivity and interactions between regions. Significant differences have been observed in the FC networks of individuals with MUD compared to healthy subjects (Bel-Bahar et al. 2022; Yan et al. 2023). However, the potential of these brain FC networks as indicators of addiction treatment effects remains unknown (Yucel et al. 2019). Moreover, these brain systems exhibit disruptions in multiple mental disorders, making the development and validation of addiction-specific biomarkers challenging (Garcia-Gutierrez et al. 2020; Niculescu and Le-Niculescu 2022). Currently, the lack of effective EEG FC-based indicators or biomarkers hampers the accurate monitoring and efficacy evaluation of MUD treatments.

In this study, EEG data in resting state were collected from 34 MUD patients after 4 weeks of active and sham rTMS stimulation to analyze the effects of rTMS on MUD patients. The study involved the construction of FC matrices in six frequency bands, followed by topological feature analysis and machine learning classification. Correlations with VAS scores were calculated as well. The primary focus of this research is to evaluate the therapeutic effects of rTMS on MUD using brain functional networks, encompassing three key objectives: (1) using FC to provide objective evidence of improved brain function following four weeks of rTMS treatment for MUD; (2) proposing an assessment metric that effectively reflects reduced craving through graph theory analysis; and (3) employing machine learning to identify individual differences in response to rTMS intervention. Hence, the direct measurement and quantification of changes in the FC network in this study bear significant research implications for promoting the development of rTMS treatment strategies for MUD.

Materials and methods

Subjects

Thirty-four male subjects with MUD were recruited from the Shanghai Qingdong Compulsory Isolation Drug Rehabilitation Center and randomly assigned to the active group (mean age = 42.75 years, SD = 8.51 years) and sham group (mean age = 45.58 years, SD = 8.48 years). All subjects were not taking addiction-like drugs and were regularly checked by urine tests. The inclusion criteria were as follows: (1) diagnosis of MUD according to DSM-5 diagnostic criteria, including at least one year of MA use; (2) no use of any stimulant or narcotic drugs within one year; (3) subjects had no history of neurological disorders, brain surgery, epilepsy, brain injury, or metal implants in the brain; (4) no history of substance use other than MA, no other mental illness, as confirmed by a structured clinical interview conducted by a professional licensed psychiatrist. The study was conducted by the Declaration of Helsinki and was approved by the Scientific Ethics Committee of Shanghai University. Participants voluntarily signed a written informed consent before participation and were free to withdraw from the study at any stage.

rTMS procedure

All subjects were stimulated with rTMS for four weeks, twice a week (consistent time intervals), for a total of 8 sessions by a specialized physician (Wang et al. 2022). The motor threshold of the subjects’ left motor cortex was determined by looking for the lowest intensity that generated a motor response in the right abductor pollicis breve (APB)—guaranteed to generate five motion-evoked potential reactions of at least 50 mV in 10 trials. During the treatment, the coil is placed in the left prefrontal region, 5 cm in front of the scalp position, where the movement threshold is determined. Subjects in the active rTMS group received high-frequency (10 Hz) rTMS at a 100% resting exercise threshold (rMT), 5s on and 15s off for 8 min, with 1200 pulses on the left DLPFC. The sham rTMS group received the same parameters, except that the coil was 90° away from the skull. This study used an 8-shaped coil for precise target stimulation (Su et al. 2017).

Craving assessment

The visual analogue scale (VAS) was used to assess the subjects’ level of craving. The VAS is now a commonly used tool to assess the level of craving in patients with substance use disorders and has been used in several studies (Moeller et al. 2015; Sayette et al. 2000). Although the desire may be related to the indicators found by EEG, skin conduction detectors and other tools, it still cannot replace subjects’ desire to self-report. The VAS ranges from 0 cm (corresponding to ‘no craving’) to 10 cm (corresponding to ‘the highest intensity of craving’) (Zhang et al. 2022).

Data recording and preprocessing

Two 5-min periods of closed-eye resting-state EEG were recorded: one before the first rTMS session (pre-EEG) and the other after the last rTMS session at 4 weeks (post-EEG). The EEG data were acquired using a 64-channel EEG cap (Neuracle Inc.) with channel electrodes arranged in rows based on the internationally recognized 10–20 system. The EEG acquisition system had a sampling rate of 1000 Hz, and all electrode impedances remained below 5 kΩ throughout the recording period. During data collection, the subjects sat in comfortable chairs in a quiet room with their eyes closed.

Data preprocessing was performed using the EEGlab toolbox, which included several steps. These steps comprised average referencing, bandpass filtering from 0.1 to 45 Hz, de-baselining, removal of bad segments, ICA artifact removal, and downsampling to 200 Hz (Hu et al. 2017). Additionally, the initial 60 s of data were discarded to minimize the influence of non-calm brain fluctuations during the early stage of the experiment. To ensure the reliability of FC analysis results, a segment length of 5 s was chosen based on previous literature (Fraschini et al. 2016). To maximize the available data for analysis, a sliding window technique was applied, as shown in Fig. 1 A. This technique divides the remaining 240-s of data with a window size of 5s and a sliding step of 1s. Due to variations in the lengths of the remaining clean data after preprocessing, the number of epochs ranged between 200 and 240 for each participant following the sliding window procedure. To maintain a consistent final sample size, each participant uniformly retained 200 valid epochs.

Fig. 1.

Overall research flow. A Data recording and processing, B Functional connectivity analysis, C Graph theory analysis, D Machine learning recognition

Functional connectivity matrix construction

Coherence (Coh) measures the linear relationship between two EEG channels at a given frequency, and this method has been widely used to calculate the FC matrix (Si et al. 2019). The imaginary part of coherence (iCoh) is calculated by dividing the imaginary part of the cross-spectrum of the electrodes by the square root of the product of the power spectra of the individual electrodes (Nolte et al. 2004). Both Coh and iCoh can be obtained from complex coherence. In this study, iCoh at a specific frequency was utilized to represent the strength of the connection between the two channels. Compared to Coh, iCoh reduces the effect of bulk conductivity and electrode reference (Bouchard et al. 2020). The complex coherence is calculated as follows.

$$C_{\mathit{xy}}\left(f\right)=\frac{P_{\mathit{xy}}(f)}{\sqrt{P_{\mathit{xx}}(f)P_{\mathit{yy}}(f)}}=Real\left\{C_{\mathit{xy}},\left(f\right)\right\}+Imag\left\{C_{\mathit{xy}},\left(f\right)\right\}$$ 1

where P_xy (f) represents the crossover power spectral density between electrode signal x and electrode signal y at frequency f, and P_xx (f) and P_yy (f) represent the power spectral density of electrode signal x and electrode signal y, respectively, at frequency f.

and iCoh is defined as the imaginary part of the complex coherence.

$${\mathit{iCoh}}_{\mathit{xy}}\left(f\right)=Imag(C_{\mathit{xy}}\left(f\right))$$ 2

The values of iCoh range from [0,1], where 1 indicates maximum linear interdependence, and 0 indicates nonlinear interdependence. Finally, brain FCs were constructed between each pair of electrodes in each of the following frequency bands: delta (0.5–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz), gamma (30–45 Hz), and all spectrum (0.5–45 Hz), as shown in Fig. 1 A.

On these bases, all 200 FC matrices for each subject were calculated by superimposing averages by frequency band separately to obtain the matrix for each subject. The FC matrices of the active/sham group before and after rTMS were then obtained by superimposing and averaging all subjects by the group. This study used the BrainNet Viewer toolbox (Xia et al. 2013) for 3D display. Finally, the connectivity differences between the four groups (active-pre, active-post, sham-pre, sham-post) were calculated and counted across the six frequency bands, as shown in Fig. 1 B.

Graph theory analysis

The graph theory indicators were further calculated after constructing the FC network to understand the changes in the topological characteristics of the brain, as shown in Fig. 1C. Clustering coefficient (C), global efficiency (Ge), local efficiency (Le), and characteristic path length (L) are four network properties that describe the ability of a resting-state network to process information (Si et al. 2019). In this study, these four weighted network properties were formulated using the following definitions and computed by the Brain Connectivity Toolbox (www.nitrc.org/projects/bct/BCT):

$$C=\frac{1}{N}\sum _{i\in \theta }\frac{{\mathrm{\Sigma }}_{j,l\in \theta }{\left(C_{\mathit{ij}}{C}_{\mathit{il}}{C}_{\mathit{jl}}\right)}^{1/3}}{{\mathrm{\Sigma }}_{j\in w_{\mathit{ij}}}\left({\mathrm{\Sigma }}_{j\in \theta }{C}_{\mathit{ij}}-1\right)}$$ 3

$$L_e=\frac{1}{N}\sum _{i\in \theta }\frac{{\mathrm{\Sigma }}_{j,l\in \theta ,j\ne i}{\left(C_{\mathit{ij}}{C}_{\mathit{il}}{\left[d_{\mathit{jl}}\left({\theta }_i\right)\right]}^{-1}\right)}^{1/3}}{{\mathrm{\Sigma }}_{j\in \theta }{C}_{\mathit{ij}}\left({\mathrm{\Sigma }}_{j\in \theta }{C}_{\mathit{ij}}-1\right)}$$ 4

where C_ij is the strength of the connection between nodes i and j, N is the node number, and θ is the set of all nodes in the resting-state network. c is defined as the fraction of triangles around a single network node. L_e is the average efficiency of the local subgraph. Both C and L_e are related to estimates of potential functional separation between brain lobes and reflect the local information processing capacity of the brain network.

$$G_e=\frac{1}{N}\sum _{i\in \theta }\frac{{{\mathrm{\Sigma }}_{j\in \theta ,j\ne i}{d}_{\mathit{ij}}}^{-1}}{N-1}$$ 5

$$L=\frac{1}{N}\sum _{i\in \theta }{L}_i=\frac{1}{N}\sum _{i\in \theta }\frac{{\mathrm{\Sigma }}_{j\in \theta ,j\ne i}{d}_{\mathit{ij}}}{N-1}$$ 6

where G_e is the average efficiency of the relevant brain network, and L is the average of the shortest path lengths between all pairs of network nodes. These metrics are used to assess the potential for functional integration between brain regions. Furthermore, these metrics are defined as the efficiency of global information processing in the brain network.

Machine learning recognition

The neuroscience field increasingly emphasizes the network patterns of FC in the brain. Leveraging the powerful information mining capabilities of machine learning, it becomes possible to capture more differences in FC networks before and after rTMS intervention, thereby identifying treatment effects (Hasanzadeh et al. 2019). In this study, Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), and Logistic Regression (LogReg) classifiers were employed to individually classify 17 participants in the active group. The sliding window technique was utilized to increase the sample size for machine learning, aiming to improve classification performance. Additionally, grid search algorithms were used to optimize model hyperparameters for better fitting.

Firstly, define the data of each subject before and after rTMS as two types of labels, with 200 samples in each class, for a total of 400 samples. Then, the dataset is divided using tenfold cross validation. That is, shuffle all 400 samples randomly and divide them evenly into 10 parts (with a balanced number of samples in both categories), with 9 parts used as the training set and 1 part used as the testing set. Under the premise of ensuring the number of test sets, tenfold cross-validation can expand the number of samples and test times of the training set, so that the test results can better reflect the real performance of the algorithm. On this basis, input the training set data into LDA, SVM, and LogReg models for training, obtain the corresponding classification models, and test them with the test set. Next, the results of each fold, such as accuracy, sensitivity, and specificity, are calculated using three models. And calculate the average of the 10 folds results to obtain the classification results of the current subject. Finally, each subject in the active group traverses once to obtain the average classification results of the three models. The identification process in this study was conducted using the NBS-Predict toolbox (Serin et al. 2021) within the MATLAB environment.

Statistical analysis

Student’s t-tests were used to test the difference in the continuous variable between the two groups, and chi-square tests were used for comparison of dichotomous variables. If the assumption of normality and/or homogeneity of variance was violated, non-parametric tests were used. Repeated-measures ANOVA was calculated to test the effect of group (active group vs. sham group) and time (baseline vs. post rTMS) on patients’ clinical performance. Paired-sample t-tests were used to compare within-group differences in pre-stimulation and post-stimulation brain network mean iCoh and graphological indicators. Independent samples t-tests were used to compare between-group differences between the two groups in pre-stimulation and post-stimulation brain network mean iCoh and graphological indicators. Network-based statistics (NBS) toolbox (Zalesky et al. 2010) was used to determine significant differences in iCoh data between the subjects before and after the rTMS intervention. Pearson correlation analysis was used to assess the correlation between critical indicators of the brain FC network and VAS data. All data analyses were completed on SPSS 26.0 and MATLAB.

Results

Demographic and clinical characteristics

There were no significant differences in age, sex, education years, duration of illness or clinical history between the active group and the sham group at baseline (Table 1). The paired t-test results showed that in the activity group, VAS scores significantly decreased after 10Hz rTMS stimulation (p < 0.001). Among them, the appetites of 15 patients improved to varying degrees, while there was no change in appetite for one patient and an increase for another. In the sham group, all VAS scores showed no significant change after stimulation (Fig. 2 and Supplementary Table S1).

Table 1.

Demographic and clinical characteristics statistics (± SD) of participants

Characteristic	Active(n = 17)	Sham(n = 17)	χ2 / F	P-value
Age, year	42.75 ± 8.51	45.58 ± 8.48	-1.682	0.107
Sex, male/female	17/0	17/0	-	-
Education, year	10.25 ± 2.70	9.58 ± 3.26	0.545	0.591
Onset age, year	26.58 ± 9.13	29.00 ± 8.10	-0.686	0.500
Withdrawal, month	20.42 ± 2.15	19.50 ± 2.37	0.866	0.396
Usage duration, year	10.75 ± 6.84	14.83 ± 10.18	-1.153	0.261
Daily dose, g	0.56 ± 0.36	0.58 ± 0.34	-0.145	0.886
VAS scores (baseline)	6.06 ± 2.63	5.59 ± 2.27	0.312	0.580

Fig. 2.

VAS score changes of the active and sham groups before and after the intervention. There was a significant decrease in the active group after the rTMS intervention but no significant change in the sham group

Functional connectivity analysis

Four groups (active-pre, active-post, sham-pre and sham-post) were calculated by averaging the FC matrices of all 200 epochs of each subject according to the connectivity difference of six frequency bands, and the matrices of each subject were obtained. Then all the subjects were calculated by grouping and averaging to get the FC matrices of the active and sham groups. Figures 3 and 4 show that the active group showed significant changes in the beta band after the rTMS intervention. There was an overall trend of enhanced FC in the beta band, with significant differences mainly in the frontal, parietal, and occipital lobes, with the most significant changes in the connectivity of C3 and FC4 channels. Additionally, the sham group showed little difference before and after the intervention, with no significant increases or decreases found in all frequency bands.

Fig. 3.

The left column represents the difference in FC of the brain in the active group after FDR correction, where blue means connectivity value of pre > post and red means pre < post. The middle/right column represents the FC matrices of the active/sham group before and after the intervention. All graphs are presented by delta, theta, alpha, beta, gamma and all bands

Fig. 4.

Bar plot of mean ± SD connectivity values in the six frequency bands. The star flags significance level (P < 0.05)

Graph theory analysis

Furthermore, the graph theory analysis of the brain network showed that the graphical indicators of the active group change significantly in the beta band. After receiving the rTMS intervention, the global efficiency (Ge Pre/Post, p = 0.005), local efficiency (Le Pre/Post, p = 0.035) and aggregation coefficient (C Pre/Post, p = 0.042) of the beta band in the active group showed significant decreases (Fig. 5 a-c and Supplementary Table S2). At the same time, the characteristic path length (L Pre/Post, p = 0.713) was not statistically different (Fig. 5 d and Supplementary Table S3). Except for the above changes, the changes in other frequency bands of the active group and all the sham group frequency bands were insignificant (Supplementary Tables S2 and S3).

Fig. 5.

Changes in the active group before and after the intervention were concentrated in the beta band: (1) significant decreases in global efficiency, local efficiency, and aggregation coefficient, and (2) no statistical difference in characteristic path length

To evaluate the relationship between the indicators (Ge, Le, and C) and actual craving, correlation analysis was conducted between the difference between the two results of each subject and the difference in the VAS. As shown in Fig. 6, the global efficiency (Ge), local efficiency (Le), and aggregation coefficient (C) of the beta band were positively correlated with the VAS, with correlation coefficients of 0.6314, 0.5194 and 0.5198, which were statistically significant (p < 0.05).

Fig. 6.

Correlation between the active group’s graph theory index and VAS score change. Ge, Le, C change (pre-post) and VAS score change (pre-post) have significant positive correlations in the beta band

Machine learning recognition

To further evaluate the treatment effects for each participant, the 17 individuals in the active group were classified before and after rTMS using LDA, SVM, and LogReg classifiers. Beta band FC was used as the input, and the classification results are presented in Table 2, Supplementary Tables S4, S5 and S6, and Fig. 7. Overall, all three models achieved effective classification, with LDA yielding the best performance, attaining an average accuracy of 0.825 (mean: sensitivity = 0.823, specificity = 0.824, precision = 0.829, F1 = 0.817, AUC = 0.826). From Fig. 7, it can be observed that all participants were distinguishable, although some differences among them existed. Additionally, similar analyses were conducted on other frequency bands, but the overall classification performance was lower, and those results are not included here.

Table 2.

Classification results statistics of three machine learning models

Features	Accuracy	Sensitivity	Specificity	Precision	F1	Auc
LDA	0.825	0.823	0.824	0.829	0.817	0.826
SVM	0.751	0.757	0.753	0.757	0.704	0.755
LogReg	0.724	0.732	0.723	0.713	0.667	0.729

Fig. 7.

Classification accuracy of three models before and after rTMS treatment for 17 subjects in the activity group

As shown in Fig. 7, there is some variation in classification accuracy between subjects. The relationship between classification accuracy and the VAS was further analyzed to determine the source of differences among different subjects. As shown in Fig. 8, there is a significant and strong correlation (p < 0.05) observed between accuracy and VAS across all three models. The correlation coefficients between LDA, SVM, and VAS exceed 0.6, with consistent statistical significance (p = 0.01).

Fig. 8.

Correlation between recognition accuracy and VAS difference in the active group. Accuracy was significantly correlated with improvement in VAS craving

Discussion

In this study, an analysis was conducted on the effects of rTMS stimulation in individuals with MUD using EEG brain networks and machine learning techniques. After 4 weeks of rTMS intervention, improvements in drug addiction symptoms (decreased VAS score) were observed in the active group. The main findings of this study were as follows: (1) Increased beta band FC was noted as a result of rTMS treatment, indicating a positive effect; (2) Beta band global efficiency (Ge), local efficiency (Le), and aggregation coefficient (C) demonstrated decreases, with significant correlations observed between these indicators and VAS performance. (3) The LDA classifier, which was based on the FC matrix, exhibited a relatively high recognition effect in the beta band. This band displayed a significant correlation with VAS performance and showed substantial inter-individual variability. These findings contribute to the understanding of the mechanisms underlying rTMS for drug addiction and provide a valid and objective assessment of treatment effects.

Demographic and clinical characteristics analysis

The two groups of participants included in this study did not show significant differences in age, sex, education years, duration of illness, or clinical history. This indicates that their physical condition and level of addiction were generally similar. It is worth noting that there are differences in the trend and speed of changes in craving levels among different withdrawal periods, and they will remain relatively stable after 12 months (Zhao et al. 2021). Therefore, all subjects recruited in this study underwent approximately 20 months of standardized mandatory isolation and rehabilitation management. This measure aimed to ensure that all participants were in the same withdrawal period and to minimize potential differences caused by variations in the stages of rehabilitation. Subjects with stable mental states can help demonstrate the positive effects of rTMS and exclude factors such as natural recovery. In terms of baseline subjective VAS scores, both groups of participants demonstrated high consistency (6.06 ± 2.63 and 5.59 ± 2.27), indicating a high level of addiction severity and risk of relapse. After receiving rTMS treatment, the active group showed a significant decrease in VAS scores, while the sham group did not exhibit significant changes. This finding confirms the positive role of rTMS, which is consistent with previous studies (Ding et al. 2023; Pan et al. 2021; Su et al. 2017). However, it is important to note that a small proportion of participants in the active group did not show improvement in VAS scores. This aligns with the general understanding of rTMS treatment and clinical experience, as it is not effective for everyone (Gold et al. 2022; Pan et al. 2021). Regardless of effectiveness, VAS assessment is overly simplistic and subjective. Therefore, to provide a more comprehensive evaluation, further objective analysis based on EEG was conducted.

Functional connectivity alterations

The utilization of brain FC methods contributes to the objective characterization and accurate assessment of neural features in individuals with MUD, thereby optimizing treatment (Khajehpour et al. 2019). In this study, it was found that the active group exhibited a significant increase in beta band FC (p < 0.05), while the sham group showed no statistically significant differences before and after intervention. The state of the brain’s electrical activity is associated with the oscillatory frequency, and although different frequency bands of EEG components are generated synchronously, most bands have distinct spatial dominance regions and neural generators. Previous studies analyzing addiction using resting-state EEG have indicated that higher frequency bands, such as the beta band, may contain more crucial information (Hu et al. 2017). In a characterization study comparing MUD with healthy individuals, the average FC value in the beta band was significantly lower in MUD, indicating an association between abnormal beta band activity and inhibitory dysfunction in the MUD brain (Khajehpour et al. 2019). Cross-sectional studies examining different stages of abstinence have also revealed a direct correlation between beta band oscillatory activity and patients’ craving latency period(Zhao et al. 2021). The literature suggests that the beta band is an important EEG oscillatory band in the context of MA addiction, with lower FC values compared to healthy individuals. However, in this study, an increase in FC values was observed, which may indicate improvements in certain brain functions. Another study involving a 4-week intervention on MUD patients demonstrated that applying transcranial magnetic stimulation (TMS) to the left dorsolateral prefrontal cortex (DLPFC) could modulate attention biases and beta oscillations during attention processing in MUD patients, as demonstrated by the Stroop Task paradigm (Chen et al. 2021). While experiment paradigms based on cognitive tasks offer insights into the behavioral and cognitive mechanisms of drug addiction, resting-state paradigms are better suited for evaluating longitudinal changes in craving intensity (Sutherland et al. 2012). By eliminating the influence of task performance on brain functionality, resting-state paradigms provide a more accurate assessment and minimize potential confounding factors. In this study, a 4-week rTMS treatment and resting-state EEG data collection experiment was designed. The active group reported a subjective VAS reduction in craving intensity and exhibited a significant increase in beta band FC (p < 0.05), while the sham group showed no statistically significant differences before and after intervention. Based on the validation of rTMS enhancing inhibitory craving capability (Chen et al. 2021), this study visually demonstrated that external neural modulation can enhance the brain’s synchronous information flow. These findings confirm the potential of beta band FC as a biomarker for evaluating rTMS treatment of MUD.

Topological characteristic alterations

This study further investigated the changes in brain network topology in individuals with MUD using graph theory analysis. These features reflect the alterations in information transmission patterns within the brain networks (Dugre et al. 2023). Compared to the sham group, the active group showed statistically significant differences in various measures mainly in the beta frequency band, consistent with the findings on changes in brain FC discussed earlier. Previous literature suggests that lower frequency bands are associated with large-scale long-range oscillations, while higher frequency bands are more locally specific (Bel-Bahar et al. 2022). The beta frequency band is particularly sensitive to cognitive processes and information processing, typically involved in short-range information transfer (Hu et al. 2017). In the active group, a decrease in overall efficiency, local efficiency, and clustering coefficient was observed in the beta band, indicating a reduction in the overall high synchrony of functional connections in the proximal regions of the brain and alleviation of the hyperexcitability state of the neural system (Ahmadlou et al. 2013). This may be related to improvements in motivation, reward, and memory recovery deficits in MUD patients (Motzkin et al. 2014; Zimmermann et al. 2017). Previous studies have shown that long-term traditional abstinent treatment leads to reduced dopamine receptor activity, as well as decreased local (short-range) cerebral blood flow and information transmission in MUD patients (Polesskaya et al. 2011; Yan et al. 2023). This suggests that abnormal brain arousal caused by MA dependence is inhibited in the MUD brain (Taebi et al. 2022). In this study, the analysis of changes in functional network topology revealed the crucial role of the beta band within high frequencies, which was significantly correlated with VAS scale reports, further demonstrating the effectiveness of rTMS in improving MUD. From an evaluation perspective of rTMS treatment, the overall efficiency, local efficiency, and clustering coefficient of the beta band hold promise as assessment indicators with interpretability and quantifiability.

Analysis of recognition results

The results above indicate that after receiving rTMS treatment, brain FC can demonstrate improvements in the FC of individuals with MUD. However, it is currently unclear whether FC can serve as a practical marker for assessing the effectiveness of rTMS treatment using machine learning (Chen et al. 2023). Previous studies utilizing machine learning techniques primarily focused on identifying raw EEG signals while neglecting the topological relationships between electrodes (Hasanzadeh et al. 2019; Yang et al. 2023). In this study, the beta band FC matrix of the active group was used as input, and LDA, SVM, and LogReg algorithms yielded good classification results, demonstrating distinct separability of participants before and after rTMS intervention, confirming the effectiveness of rTMS.

Among these methods, LDA achieved the best classification performance, with an average accuracy of 82.5% (mean: sensitivity = 82.3%, specificity = 82.4%, precision = 82.9%, F1 = 81.7%, AUC = 82.6%). LDA aims to maximize the separation between different classes and minimize within-class scatter, making different class data more easily distinguishable in the feature space (Lotte et al. 2018). This makes it more likely for LDA to find effective classification boundaries, thus improving classification accuracy (Blankertz et al. 2008). LDA considers the distribution of each class during classification and incorporates class information into the model. In contrast, SVM and LogReg classify based on the relative positions of instances and may overlook the impact of class distribution (Lotte et al. 2018). In this study, SVM and LogReg performed slightly worse, indicating the presence of interfering features in the FC matrix that affected classification performance. LDA can reduce the dimensionality of high-dimensional features through projection, helping to remove redundant features and improve classification performance. Previous research has demonstrated good results using LDA for classification in patients with schizophrenia and major depressive disorder (Jang et al. 2021). Additionally, LDA generally performs better when the sample size is relatively small (Wu et al. 2013), which is also evident in this study.

Individual machine learning classification further supplemented the analysis of FC and graph theory, revealing significant individual differences. Similar observations can be found in Supplementary Table 3 and Fig. 7, where there are notable differences between high and low classification accuracy among all participants. For example, participants S6 and S11 had classification accuracies exceeding 90%, while participants S3 and S17 had classification accuracies of only around 75%. To identify the sources of these differences, further analysis was conducted on the relationship between accuracy and VAS scores. Significant correlations were observed between the classification accuracies of the three algorithms and the actual VAS scores, as shown in Fig. 8. Participants with better VAS performance (i.e., greater reduction in VAS score after rTMS) had higher classification accuracies before and after intervention, especially with LDA and SVM (correlation coefficient of approximately 0.6, significance of approximately 0.01). Therefore, it can be inferred that classification accuracy can reflect the actual therapeutic effects of rTMS. This can be helpful in the future for screening patients with different treatment outcomes to facilitate the development of targeted stimulation protocols and parameters. In conclusion, the combination of LDA algorithm and beta band FC matrices shows promising accuracy and sensitivity in predicting the effectiveness of rTMS treatment. It can serve as an effective method for identifying the effects of rTMS treatment on MA addiction. It has the potential to be used in the future for personalized treatment planning, promoting the development of rTMS and other therapeutic techniques.

Limitations

Due to objective difficulties in recruiting MUD patients and continuing rTMS for four weeks, the sample size included in this study is relatively small. In future work, we will gradually expand the sample size. The graphical analysis included many indicators, and only the four commonly used indicators were chosen for this study, potentially overlooking the potential value of other indicators. The study focused on how to assess the effectiveness of the rTMS rather than how to improve the treatment effect. Therefore the frequency of the intervention cycles was low, and although it lasted for 4 weeks, only 2 interventions were conducted per week, which may have had a weak actual treatment effect.

Conclusion

EEG brain networks and machine learning techniques were utilized to analyze the effects of rTMS stimulation in MUD. This study demonstrated that rTMS has positive implications for the treatment of MUD: (1) significant changes are found in the beta band, demonstrating the positive ameliorative effect of rTMS; (2) the global efficiency, local efficiency and aggregation coefficient of the beta band can effectively reflect changes in craving; (3) the LDA classifier based on FC matrix showed a relatively high recognition effect in the beta band, which has a significant correlation with VAS performance. These findings have guiding significance for exploring different treatment options in the future.

Supplementary Information

Below is the link to the electronic supplementary material.

Authors contributions

Yongcong Li, Banghua Yang, and Jun Ma contributed to the conception, experimental design and wrote the manuscript. Yongcong Li, Jie Zhang, Yunzhe Li, and Hui Zeng acquired and analyzed experimental data. All authors critically reviewed the content and approved the final version for publication.

Data availability

The data supporting this study’s findings are available from the corresponding author upon reasonable request.

Declarations

Conflict of interest

The authors have no conflicts of interest to declare.