Predicting IGD treatment outcomes across interventions using a whole-brain functional connectome biomarker

Abstract

Background

Although treatment approaches have been explored in internet gaming disorder (IGD), however, little is known about the specific neural activity that is associated with treatment outcomes. This study aims to establish a treatment-sensitive network that reliably predicts treatment outcomes across different interventions for IGD.

Methods

Resting-state fMRI data were collected from 90 individuals with IGD who received one of three different intervention protocols (mindfulness meditation, approach bias modification, or retrieval-aversive conditioning training). Connectome-based predictive modelling (CPM) was used to build a treatment-sensitive network to predict individual symptom reduction, then validated them on the external dataset.

Results

The negative (r = 0.401, p <.001) and combined network (r = 0.480, p <.001) can significantly predict a reduction of addiction severity (baseline minus post-test). Specifically, the negative network was primarily centred on the motor, subcortical, and salience networks. The positive network primarily involved cerebellar-related connections. Incorporating these into a combined model significantly improved predictive accuracy, suggesting that cerebellar-related connectivity may contribute to the prediction of intervention-related clinical improvement. Finally, the treatment-sensitive network was validated in a separate data.

Conclusions

This study established a treatment-sensitive network based on the whole-brain connectome. These findings provide valuable empirical evidence for identifying biomarkers for IGD treatment and advance personalised therapeutic strategies.

Clinical Trials Registration

The trial protocol was pre-registered with the Chinese Clinical Trial Registry (www.chictr.org.cn; ChiCTR2300077831(MM), ChiCTR2300075869(ApBm), ChiCTR2500095444(R-AC), ChiCTR2400090834(SAE)).

Introduction

Internet gaming disorder (IGD), a major form of behavioural addiction, has emerged a significant global public health concern (American Psychiatric Association 2013; World Health Organization 2018). A number of studies have found that IGD is not only associated with a wide range of psychological problems and cognitive deficits (Ostinelli et al. 2021; Gao et al. 2022; Meng et al. 2024), but also causes functional and structural neurological changes (Yao et al. 2017).

In recent years, a range of treatment approaches for IGD-including cognitive behavioural therapy (CBT), pharmacotherapies, non-invasive brain stimulation, and other methods-have demonstrated preliminary efficacy (Dong et al. 2024). However, challenges remain, such as pronounced inter-individual variability, limited understanding of long-term outcomes, and the unpredictability of relapse risk (Dong et al. 2024; Hayes et al. 2020; Mestre-Bach and Potenza 2023). Overall, IGD treatment research is still in its early stages.

Notably, alterations in neural activity and functional connectivity (FC) within the brains of individuals with IGD are not only associated with the development of the disorder but also provide insights into the neural mechanisms underlying treatment outcomes (Mestre-Bach and Potenza 2023). Interventions such as CBT and craving behavioural intervention (CBI) have been shown to modulate the FC between the prefrontal cortex and reward-related regions like the striatum, thereby enhancing cognitive control and reducing impulsive behaviours and gaming cravings. For instance, Zhang et al. (2018) found that IGD individuals who underwent CBI exhibited a significant decrease in resting-state functional connectivity (rsFC) between the ventral striatum (VS) and the left inferior parietal lobule (lIPL). This reduction suggests that the reversal of VS-lIPL connectivity may serve as a potential biomarker for effective intervention. Similarly, Liu et al. (2021) reported that IGD participants displayed diminished whole-brain connectivity involving the left angular gyrus and ventromedial prefrontal cortex compared to healthy controls. Changes in intrinsic connectivity distribution values between these regions were associated with alterations in craving and addiction severity six months post-intervention. Furthermore, CBI was linked to improved connectivity within the default mode network (DMN) and between the DMN and salience network regions. In terms of pharmacological treatment, Bae et al. (2018) observed that bupropion administration in IGD individuals led to decreased FC within the posterior DMN and between the DMN and cognitive control network. Additionally, a longitudinal study by Dong et al. (2019) found that, compared to persistent IGD patients, individuals who had naturally recovered from IGD showed lower dorsolateral prefrontal cortex activation in response to gaming cues at both pre- and post-gaming assessments. Collectively, these findings suggest that intervention-related changes in neural connectivity may underpin recovery from IGD. Identifying neural markers that prospectively predict treatment response could facilitate the development of more targeted and effective intervention strategies for individuals with IGD.

Although previous neuroimaging studies have documented widespread abnormalities in brain network dysfunction (Tolomeo and Yu 2022) and gray matter volume (Zhang et al. 2021) among individuals with addiction, these findings predominantly rely on “predictive” analyses associating self-reported symptom severity with features of specific brain regions. However, such prospective association-based statistical inference fails to meet the key standard for true predictive modelling-namely, validation on independent samples (Scheinost et al. 2019). Furthermore, existing neuro-prediction studies often conflate explanatory modelling with predictive modelling. While the former focuses on identifying statistical associations between variables, the latter emphasises generalisation to unseen individuals (Yarkoni and Westfall 2017). Thus, overcoming the limitations of current neuro-prediction approaches and developing efficient, generalisable individualised predictive models has emerged as a crucial direction in the field of addiction intervention.

In recent years, machine learning approaches based on whole-brain functional connectivity - particularly connectome-based predictive modelling (CPM) - have gained substantial attention due to their unique computational advantages (Finn et al. 2015; Shen et al. 2017). Unlike conventional approaches, CPM circumvents overfitting risks through cross-validation strategies and enables unbiased extraction of “neural fingerprints” from whole-brain networks. This methodology has demonstrated successful applications across diverse domains including cognitive abilities (Barron et al. 2021; Dong et al. 2023), emotion (Ren et al. 2021; Yang et al. 2024), and treatment outcome (Yip et al. 2019; Lichenstein et al. 2021; Lichenstein et al. 2023; Garrison et al. 2023; Misaki et al. 2024; Liu et al. 2025). Notably, meta-analyses by Tolomeo and Yu (2022) revealed that while substance addictions and behavioural addictions (such as IGD) share common neural substrates (e.g., prefrontal-limbic system imbalance), they exhibit distinct core biomarkers. Studies now applying CPM have found that cocaine and opioid withdrawal networks are anatomically distinct from each other (Yip et al. 2019; Lichenstein et al. 2021; Lichenstein et al. 2023). However, it remains unclear whether there exist consistent neural biomarkers across interventions targeting behavioural addictions.

To address these gaps, the present study proposes a cross-intervention connectome-based predictive framework for IGD, aiming to answer two pivotal questions: (a) Do FC biomarkers exist that can predict post-treatment changes in IGD severity and are consistent across diverse intervention modalities? (b) Are these neural predictors stable and generalisable? To this end, we selected three intervention approaches-mindfulness meditation (MM), approach bias modification (ApBM), and retrieval-aversive conditioning (R-AC)-each representing distinct cognitive and emotional regulatory mechanisms. In addition, an independent cohort receiving systematic aerobic exercise (SAE) training was included for external validation to further test the generalisability of the treatment-sensitive network. Based on previous findings, we hypothesised that functional connectivity within and between salience and subcortical networks would predict treatment effects in IGD under different intervention conditions.

Methods

Ethics

This study was approved by the Institutional Review Board of the Yunnan Normal University and was conducted in accordance with the Declaration of Helsinki. All participants provided written informed consent and received monetary compensation upon completion of the experiment. The trial protocol was pre-registered with the Chinese Clinical Trial Registry (www.chictr.org.cn; ChiCTR2300077831(MM), ChiCTR2300075869(ApBM), ChiCTR2500095444(R-AC), ChiCTR2400090834(SAE)).

Participants and assessments

The study cohort comprised 123 participants diagnosed with IGD (57 males; mean age = 20.68 ± 2.59 years), recruited from four distinct clinical trials (The method for calculating the sample size is described in the study by Ni et al. (2024). The sample distribution across interventions was as follows: MM (n = 30), ApBM (n = 30), R-AC training (n = 30) and SAE group (n = 33). Inclusion and exclusion criteria are described in the Supplementary Information (SI). All participants underwent two rounds of neuroimaging scans and completed questionnaire assessments (Addiction severity was assessed with the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition, Text Revision (DSM-5-TR) proposed criteria for IGD (Petry et al. 2014)). All individuals included in the present study met diagnostic criteria for IGD and had a total score ≥ 6 (ranging from 0 to 9) at baseline. For data analysis, dataset 1 was established using MM, ApBM and R-AC group as the discovery dataset, while dataset 2 utilised SAE as the external validation dataset. Table 1 summarises the demographic and clinical characteristics.

Table 1.

Demographic and clinical characteristics of the two datasets.

Variables	Discovery Sample (N = 90)	Validation Sample (N = 33)
Sex (Male/Female)	46/44	11/22
Age (years)	20.94 ± 2.87	19.97 ± 1.36
Handedness (right/left)	90/0	33/0
IAT	69.44 ± 11.03	62.88 ± 12.19
DSM-5-TR
Baseline	6.94 ± 1.32	6.48 ± 1.15
Post-treatment	4.86 ± 2.45	3.91 ± 2.14
t a	8.762***	6.17***

Notes: Data were presented as the mean ± standard deviation.

Abbreviations: DSM-5-TR, Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition, Text Revision., a: the t-value was obtained from the paired t-test between baseline DSM-5-TR and post-treatment DSM-5-TR, ***p < 0.001.

To determine the treatment effect in IGD patients, we used DSM-5-TR scores at baseline and after treatment. Specifically, we used the difference between the two (baseline minus post- treatment) for evaluation.

Behavioural interventions

We employed four distinct intervention protocols, which are briefly introduced as follows:

Mindfulness meditation (Ni et al. 2024): Participants underwent a 4-week group mindfulness training, consisting of two weekly sessions lasting 2.5 to 3.5 h each. Sessions were led by two certified mindfulness trainers with 10–20 participants per group. Trainers provided the participants and addressed inquiries. The goal was to enhance emotional regulation and self-awareness through systematic mindfulness practice to reduce gaming cravings and addictive behaviours.

Approach-avoidance bias modification (Wei et al. 2025): Participants completed a 10-day training involving five sessions per day. They performed approach-avoidance motor responses-pushing a joystick for neutral images and pushing it for gaming-related images. Frame ratios (square vs. circular) were manipulated (9:1 for gaming, 1:9 for neutral) to prevent pattern recognition. Each image appeared 10 times (200 trials per session), with balanced responses directions. This intervention aimed to reduce automatic approach tendencies towards gaming cues and decrease gaming-related cravings.

Retrieval-aversive conditioning training (Li et al. 2026): Over five consecutive days, participants first viewed gaming videos to activate related memories. After a 10-minute quiet period within the reconsolidation window, they rated the emotional valence and craving intensity of 20 gaming-related images. Subsequently, each gaming image was followed by an aversive image to form a conditioned “gaming–aversive” pairing. Participants then re-rated the original gaming images. This protocol aimed to weaken the positive emotional association with gaming cues via memory reconsolidation mechanisms.

Systematic aerobic exercise (Cui et al. 2025): Participants in the SAE group received a 20-session aerobic exercise program consisting of at least 4 sessions per week. Each session consisted of a 10-minute warm-up, 40 min of moderate-intensity aerobic exercise (46-55% of maximum oxygen consumption), and 10 min of stretching. Exercises included treadmill running, stationary bike, jump rope, and resistance. The intervention was designed to improve physical fitness, increase cognitive control, and reduce play-related cravings.

These four intervention protocols are effective, as demonstrated in our published work (Ni et al. 2024; Wei et al. 2025; Li et al. 2026; Cui et al. 2025), and their detailed procedures are given in the SI.

Neuroimaging data acquisition and fMRI pre-processing

Pre-treatment imaging data were used for analysis with acquisition parameters detailed in the SI. Data preprocessed and analysis were conducted using DPABI (http://rfmri.org/dpabi (Yan et al. 2016)). The first 10 volumes were removed to ensure signal stabilisation. Then, remaining volumes underwent slice-timing correction and realignment for head motion (exceeding 2 mm translation and 2° rotation were excluded). Functional volumes were normalised to MNI space using EPI templates (voxel size = 3 × 3 × 3 mm³) and spatially smoothed with a 6 × 6 × 6 mm³ Gaussian kernel. Subsequently, linear trends were removed, and band-pass filtering (0.01-0.1 Hz) was applied to reduce low-frequency drift and high-frequency noise. Finally, nuisance signals (24 motion parameters, white matter, cerebrospinal fluid) were regressed out.

Functional connectivity construction

Functional network nodes were defined using Shen’s 268-node brain atlas (Shen et al. 2013), which provides a graph-theory algorithm-based parcellation optimised for spatial coherence and functional homogeneity. The Shen atlas originally defines 8 large-scale functional networks, which were further refined into 10 subnetworks by Finn et al. (2015) (see Fig. S1), including medial frontal (MFN), frontoparietal (FP), default mode (DMN), motor (MOT), visual I (VI), visual II (VII), visual association (VAs), salience (SAL), subcortical (SCN), and cerebellum (CBL) network. To improve the reliability and sensitivity of network analyses, we constructed static whole-brain rsFC matrices for each participant by computing pairwise correlations (Pearson’s r) between nodes. Whole-brain rsFC matrices were constructed by in DPABI. Fisher’s r-to-z transformation was applied to normalise the correlation coefficients, yielding a symmetric 268 × 268 matrix, where each element represents the connectivity strength (or “edge”) between a pair of nodes.

Connectome-based predictive modelling

CPM was implemented in MATLAB to predict treatment effects based on whole-brain rsFC. This script is freely available online (https://www.nitrc.org/projects/bioimagesuite/). The CPM process consists of the following key steps (see Figure 1):

Figure 1.

Flowchart for prediction of treatment effect using whole-brain functional connectivity.

Abbreviations: MM, mindfulness meditation, ApBM, approach bias modification, R-AC, retrieval-aversive conditioning, SAE, systematic aerobic exercise training; FC, functional connectivity.

1. Data preparation: Connectivity matrices and behavioural data were loaded into MATLAB (training dataset: MM, ApBM and R-AC group, N = 90).

2. Cross-validation: Based on the ten essential guidelines for predictive modelling of rsFC data proposed by Scheinost et al. (2019), we implemented leave-one-out cross-validation (LOOCV) to determine the edge sets forming the network. In this approach, each iteration used N − 1 participants as the training set, while the remaining participant served as the validation set, where N represents the total number of participants. This process was repeated N times, ensuring that each participant was used once for validation.

3. Edges selection: For each participant, we calculated the correlation between each edge in the connectivity matrices and behavioural data. To account for potential confounding effects, partial correlation was performed using linear regression, with group, age, sex, and mean framewise displacement (mFD) included as covariates. Edges that met the predefined significance threshold were selected for model construction.

4. Network construction: Positively and negatively correlated edges were used to construct separate positive and negative predictive networks, respectively.

5. Single-subject summary values: For each participant, the values of all edges within the positive and negative networks were summed separately to derive single-subject summary values.

6. Model construction and prediction: Linear regression models were fitted to predict treatment effects from the derived network strengths. Additionally, a composite model was generated by combining both networks.

7. Model evaluation: The predictive model’s performance was assessed by calculating the correlation between predicted and observed values. The statistical significance of this correlation was determined through permutation testing.

For a detailed introduction to CPM, please refer to Shen et al. (2017).

Model validation and performance

We employed three validation approaches, each with a specific objective. First, we used LOOCV and tested different thresholds during feature selection to compare results. To evaluate predictive accuracy, we used partial correlation to assess the relationship between actual and predicted outcomes. After calculating the true prediction correlation, we applied a permutation test to disrupt the true brain-behaviour relationship. Specifically, we randomly shuffled the correspondence between connectivity matrices and behavioural variables 1,000 times, reran the CPM pipeline for each permutation, and generated null distributions. The statistical significance of the LOOCV prediction was assessed by computing the proportion of permuted correlations equal to or exceeding the observed correlation, across 1,000 permutations. Unless otherwise specified, all reported p-values were derived from nonparametric permutation tests. Second, considering that LOOCV provides prediction error estimates with higher variability compared to K-fold CV, we implemented K-fold CV to further validate treatment-sensitive models, using 5-fold, 10-fold, and 20-fold configurations. Each K-fold CV procedure was repeated 1,000 times with random splits to derive stable estimates of model performance. Finally, model generalisability was tested by applying the model trained on dataset 1 to an independent validation dataset.

Predicting treatment effect using demographic and clinical factors

As in previous work (Lichenstein et al. 2021; Ju et al. 2020), to compare with CPM performance, we applied the support vector regression method to investigate whether sex, age, and baseline DSM-5-TR scores could effectively predict treatment effects.

External validation

We assessed the generalisability of the treatment-sensitive network in an independent sample (SAE group, N = 33) that received an aerobic exercise intervention. fMRI data were processed using the same pipeline, and model predictions of addiction change were evaluated via partial correlation.

Results

Self-reported treatment results

As shown in Table 1, paired t-tests revealed a significant reduction in DSM-5-TR scores following treatment in both datasets (p < .001). There was no significant association between baseline DSM-5-TR scores and treatment outcomes. Detailed results are reported in the Supporting Information.

Prediction of treatment effect

Following Shen et al. (Shen et al. 2017) recommendations, we tested different thresholds (0.03, 0.05, 0.01, 0.005, 0.001) in the CPM to identify the optimal edge threshold (as sensitivity analysis). Because more liberal thresholds may include a larger number of weak or potentially non-informative edges, we focused on the relatively stringent threshold of p < .001 for the main network characterisation (see Table 2). As shown in Figure 2, the negative (r = 0.401, p < .001) and combined network (r = 0.480, p < .001) significantly predicted addiction severity. By contrast, the positive network did not significantly predict treatment effect (r = 0.206, p = .052). In addition, rerunning the CPM while controlling for covariates (age, sex, and mFD) produced similarly good predictive results (further details in SI).

Table 2.

Results of LOOCV with different thresholds (sensitivity analysis).

	Positive network			Negative network			Combined network
Threshold	r	p	N-edges	r	p	N-edges	r	p	N-edges
0.05	0.084	0.433	546	0.200	0.059	1254	0.419	<.001	1800
0.03	0.091	0.392	318	0.233	0.027	824	0.417	<.001	1142
0.01	0.130	0.222	92	0.240	0.022	350	0.392	<.001	442
0.005	0.154	0.146	46	0.281	0.007	194	0.416	<.001	240
0.001	0.206	0.052	12	0.401	<.001	59	0.480	<.001	71

Abbreviations: r, partial correlation coefficient in network; p, probability value in network.

Figure 2.

Scatter plots of correlation between observed and predicted treatment effect and permutation distribution of the correlation coefficient (r value) for the prediction analysis.

Notes: (A) Positive network. (B) Negative network. (C) Combined network. The line represents the fitted trend, with the shaded area indicating the 95% confidence interval.

Treatment-sensitive network

We applied a threshold (p < .001) and LOOCV to retain the most significant edges in the connectivity matrixes for the following analyses. The total possible edges were defined by the 268-node functional brain atlas used in this work (Shen et al. 2013), resulting in a total of (268 × 267)/2 = 35,778 possible connections. After LOOCV, 12 edges in the positive network and 59 edges in the negative network consistently appeared across all iterations and were defined as the contributing networks (see Table 2 and Figure. 3). Furthermore, there were no overlapping edges between the positive and negative networks. Given the lack of significant predictive performance in the positive network, subsequent analyses focused exclusively on the negative network. The results show that the MOT had the highest number of within-network connections, followed by the SC. For between-network connections, strong connectivity was observed between the SC and SAL, SC and VI, SC and VII, SC and MOT, and SAL and VI.

Figure 3.

Prediction models and contributing connections at the 0.001 threshold. Notes:

The results were visualised using BioImage Suite (BioImage Suite Connectivity Viewer).

(A) The model selects positive (purple lines) and negative (blue lines) networks; a larger sphere indicates more connections for that node. The positive network indicates that the higher the FC the more DSM-5 scores decreased, and the negative was vice versa. For the matrix of (B), cells represent the number of positive versus negative edges connecting nodes within (and between) each network. The larger the number of cells, the darker the colour indicates a larger number of edges. It is worth noting that positive and negative networks do not contain overlapping edges. (C) The chord plots for the positive, negative and combined networks are presented in order.

Abbreviations: Network Labels: MF, medial frontal; FP, frontoparietal; DMN, default mode; MOT, motor; VI, visual I; VII, visual II; VAs, visual association; SAL, salience; SC, subcortical; CBL, cerebellum.

To determine the relative specificity of the different network subcomponents, we performed virtual lesions (see SI ‘Computational virtual lesion prediction’ section). First, we conducted virtual lesion analysis by systematically removing edges corresponding to each specific network from the full network and rerunning the CPM. The results show that in the negative network, all networks except cerebellum were important predictive networks. Second, the results of retaining only a particular network showed that the predictions of MOT, SC, VAs and CBL networks were significant.

Additionally, the 5 most highly connected nodes in the negative network (i.e., nodes for which decreased connectivity is predictive of treatment outcomes in IGD) were predominantly located in the limbic regions, with additional nodes distributed in the subcortical, temporal, and prefrontal lobe (see Table S6, Figure. 4).

Figure 4.

Connectivity patterns of the top 5 nodes with the most connections.

Abbreviations: L, left hemisphere; R, right hemisphere; NTW, network; BA, Brodmann area; MNI, Montreal Neurological Institute; Neg, negative.

Validation analyses

To further evaluate the robustness of the findings, we re-estimated model performance using additional k-fold cross-validation schemes (5-fold, 10-fold, and 20-fold) at the 0.005 threshold. All k-fold results were further assessed using 1,000 permutation tests. The predictive effects remained statistically significant across all validation strategies (see Table 3).

Table 3.

Results of different CV schemes at the threshold of 0.005.

	Positive network			Negative network			Combined network
K-fold	r	p	N-edges	r	p	N-edges	r	p	N-edges
5	0.111	0.335	18	0.255	0.027	87	0.377	0.003	105
10	0.154	0.182	29	0.285	0.010	117	0.409	<0.001	146
20	0.162	0.139	35	0.287	0.013	153	0.416	<0.001	188

Abbreviations: r, partial correlation coefficient in network; p, probability value in network.

Associations between demographic and clinical factor and treatment effect

Support vector regression (SVR) incorporating baseline clinical variables did not predict treatment effect (details in SI).

Control and sensitivity analyses

Control analyses examining the effects of different covariate specifications are described in the “Control Analyses” section of the SI. We also conducted a series of sensitivity analyses to evaluate the robustness and specificity of the CPM findings. These analyses examined the relationship between the identified network and pre-treatment DSM-5 symptom severity, tested the trained CPM model in the sham-control group, reassessed external validation after ComBat harmonisation, and evaluated the potential influence of sex in the external validation dataset. Detailed methods and results for these analyses are provided in the “Sensitivity testing” section of the SI.

External generalisability

To evaluate external validity, we applied the predictive model to an independent validation sample, including mFD as a covariate. Results revealed the negative network (r = 0.454, df = 30, p = .008) and combined network (r = 0.428, df = 30, p = .013) significantly predicted treatment outcomes, whereas the positive network did not (r = 0.171, df = 30, p= .341) (Figure. S3).

Discussion

This study delineated a treatment-sensitive network using CPM, consisting of connections distributed across multiple functional regions of the brain, which can predict individual differences in symptom improvement. Across different intervention paradigms, the value of this treatment-sensitive network lies in its ability to capture functional architecture that moderates treatment outcomes-shedding light on why some individuals benefit more from interventions than others, and providing potential neural targets for future therapeutic strategies.

Mapping the predictive treatment-sensitive network

We found that the treatment-sensitive network is a highly complex system comprising interactive connections across multiple canonical brain networks (Yip et al. 2019; Lichenstein et al. 2023; Zhou et al. 2022; Lin et al. 2022; Chen et al., 2023). In recent years, a growing body of evidence has suggested that both the onset, maintenance and recovery of addictive behaviours involve widespread and dysregulated systems related to motivation, reward processing, sensorimotor integration, and attentional control (Zhang et al., 2018; Li et al. 2023; Skok and Waszkiewicz 2024).

Critically, the network’s strongest predictive signals came from the negative networks, where greater FC strength was associated with poorer treatment outcomes. Specifically, the MOT network exhibited the highest number of within-network connections, followed by the subcortical network, with the most strongly connected nodes concentrated in these two networks. Additionally, the SC network demonstrated strong connectivity with several key networks, including the salience network, primary and higher-order visual networks (VI and VII), and the motor network. Virtual lesion analysis further revealed that the MOT, SC, and cerebellar networks were the most important contributors to craving prediction. Notably, within the positive network, the cerebellum played a particularly prominent role. Although the standalone predictive effect of the positive network was not statistically significant, we observed that the best performance emerged when both positive and negative networks were combined; thus, the final model incorporated features from both. Based on these findings, we propose a predictive network model for addiction intervention responsiveness (see Figure. S4), which hypothesises that: (i) increased within-network connectivity in the MOT and SC networks may serve as a negative predictor, reflecting a functional decoupling of the motor-motivational system as a key neural marker of treatment efficacy; (ii) modulation of the reward circuitry-manifested as reduced connectivity between the SC and the salience and visual networks-may underlie decreased motivational drive and sensory over-reactivity to gaming-related cues, contributing to successful intervention outcomes; and (iii) the cerebellum may support cognitive restructuring and emotional regulation in addiction by modulating cortico-subcortical interactions.

Key networks that sensitive to treatments

In this study, the MOT emerged as a core component of the negative predictive network, with its intra-network connectivity significantly associated with poorer treatment outcomes. This finding aligns with previous research. For instance, Kosten et al. (2006) reported that sensorimotor system activity was predictive of relapse in individuals with cocaine dependence. Similarly, Yang et al. (2023), employing a connectome-based machine learning approach, found that reduced sensorimotor connectivity positively predicted craving scores at follow-up. Moreover, the heightened connectivity within the MOT may reflect a strong tendency towards automatised behavioural control (Ersche et al. 2020; Weinstein and Lejoyeux 2020). Such automatisation likely stems from the repeated reinforcement of perceptual-motor pathways through prolonged gaming behaviour, ultimately leading to rigid, game-related stimulus-response patterns (Dong et al. 2012; Wang et al. 2016; Yan et al. 2021). We speculate that this behavioural automatisation and rigidity constitute a key barrier in the therapeutic process. That is, a highly integrated MOT network may indicate the brain’s prioritisation of established behavioural routines (Wang et al. 2016), reducing the individual’s capacity to flexibly adapt to novel situations or cognitive demands. In other words, elevated MOT connectivity may represent a form of “neural resistance” to cognitive-behavioural intervention-individuals with IGD may find it more difficult to disrupt habitual neural circuits, incorporate new regulatory strategies, or execute alternative response patterns. This “neural rigidity” may limit adaptive changes during therapy, thereby undermining intervention effectiveness.

The SCN, which mediates the bidirectional flow of information between cortical and subcortical regions, plays a central role in reward processing, impulse control, and habit formation (Li et al. 2023; Schmidt et al. 2015; Keramati et al. 2017). In the present study, the SCN showed significant predictive value within the negative network, especially in terms of its cross-network connectivity with other functional systems. Our findings are in line with those of Yang et al. (Yang et al. 2023), who demonstrated that negative predictive networks-including the SCN-effectively forecasted treatment response in patients with opioid use disorder. In addition, tour findings partially correspond to previous studies (Weinstein and Lejoyeux 2020; Zong et al. 2022; Chen et al., 2023). Notably, increased SCN connectivity with the MOT and SAL may indicate a more efficient or robust bottom-up integration of sensory input with motivational and incentive control systems, which in some cases may reflect greater sensitivity to behaviourally relevant stimuli (Worhunsky et al. 2013).

The SAL, a key system for allocating attentional resources and evaluating the relevance of internal and external stimuli, is typically activated during bottom-up attention to salient or behaviourally meaningful cues (Seeley et al. 2007). Individuals with IGD often exhibit heightened reactivity and attentional bias towards addiction-related salient cues (Tolomeo and Yu 2022; Ryan 2002; Laird et al. 2011; Ma et al. 2019). Increasing evidence suggests that altered SAL connectivity is closely associated with the severity of addictive behaviours and intervention-related symptom changes (Fedota and Stein 2015; Geng et al. 2017; Steele et al. 2018; Moeller and Paulus 2018). For example, Moeller & Paulus (2018) reported that heightened SAL reactivity to drug-related cues prior to treatment predicted poorer clinical outcomes, while Geng et al. (2017) found that SAL dysfunction in chronic cocaine users could predict relapse in independent samples. Extending these findings, our study identified a significant association between increased SAL–VI connectivity and poorer treatment outcomes. This may reflect an unresolved attentional bias towards visually salient, gaming-related cues during therapy. In other words, the enhanced coupling between the salience and visual networks may signal a failure to suppress the automatic processing of addictive cues, thereby undermining the processing and internalisation of therapeutic content. Such persistent reactivity could result in continuous environmental distraction by gaming-related stimuli, hindering attentional shifts towards new cognitive-behavioural goals. Consequently, individuals with IGD may experience reduced engagement and focus on core therapeutic elements such as cognitive restructuring or alternative behaviours, ultimately compromising treatment efficacy.

Cerebellum -related connectivity may provide complementary predictive information within the treatment-sensitive network

Although the predictive performance of the positive network alone did not reach statistical significance, the combined model integrating both positive and negative network strengths outperformed the negative network alone. This finding should not be interpreted as evidence that the positive network is an independent predictor of symptom reduction. Rather, it suggests that cerebellar-related connectivity may provide complementary information when combined with the negative predictive network.

Notably, cerebellar-related connections accounted for the majority of the positive network features. Traditionally, the cerebellum has traditionally been regarded as primarily involved in motor coordination and procedural learning. However, accumulating evidence from addiction research suggests that the cerebellum may also contribute to habitual behaviour, craving-related processing, and the modulation of the salience of addiction-related cues (Moulton et al. 2014). The cerebellar predictive model further proposes that the cerebellum may function as a forward-modelling system that monitors internal states and predicts changes in craving or motivational states at a relatively automatic level (Moreno-Rius and Miquel 2017).

In this context, cerebellar-related connectivity may complement the negative network by capturing additional variance related to cue processing, habit learning, and the coordination between craving-related and control-related systems. For example, increased connectivity between the cerebellum and visual network may reflect enhanced integration of visually salient gaming-related cues (Antons et al. 2020; Ko et al. 2009), whereas cerebellar-DMN connectivity may be related to internally oriented craving, self-referential processing, or difficulty disengaging from gaming-related mental representations (Halko et al. 2014). Therefore, although cerebellum-related connectivity in the positive network was insufficient, it may improve prediction when integrated with salience, motor, and subcortical network features (Misaki et al. 2024; Moulton et al. 2014).

Nevertheless, this interpretation should remain cautious. Because the positive network alone did not reach statistical significance, the cerebellar contribution should be viewed as a preliminary and complementary signal within the integrated predictive model rather than as an independent biomarker or causal mechanism of treatment response. Future studies with larger samples and independent validation are needed to determine whether cerebellar connectivity reliably contributes to treatment-related symptom improvement in IGD.

Limitations

Several limitations should be noted. First, we focused on resting-state connectivity, while task-based fMRI might provide additional insights into dynamic neural processes related to addiction. Second, our sample included only IGD patients. Future studies could incorporate larger multi-site samples or transdiagnostic cohorts to determine whether a more generalised “treatment-sensitive network” can be established. Third, although we implemented multiple validation strategies, the overall sample and validation dataset size remain modest. Prior work has demonstrated that small test sets in machine-learning studies may lead to unstable or inflated performance estimates due to sampling variability (e.g., Flint et al. 2021). Therefore, in addition to viewing the present study as a preliminary exploratory investigation, replication in larger independent intervention cohorts is needed. Fourth, the mechanistic interpretation of cerebellum-related connections within the positive network should therefore remain cautious. Future studies with larger independent samples are needed to further validate the contribution of cerebellar connectivity to treatment-sensitivity prediction in IGD. Fifth, the sex imbalance in the external validation cohort should be acknowledged, and future studies should validate the model in larger, sex-balanced cohorts and examine the potential moderating role of sex in treatment-sensitivity prediction for IGD. Finally, while we accounted for common covariates in our analyses, it is impossible to control for every potential factor that may influence connectivity strength.

Conclusions

This study successfully developed a treatment-sensitivity predictive model based on whole-brain functional connectivity across multiple intervention approaches, revealing the potential role of multisystem neural interactions in the treatment of IGD. Our findings highlight the importance of the motor, subcortical, and salience networks, whose intra- and inter-network coupling strengths emerged as potential biomarkers of intervention-related symptom changes. Furthermore, the cerebellum’s bridging role within the positive predictive network suggests that its contribution to treatment outcomes may have been previously underestimated. In sum, these results deepen our understanding of the neurobiological mechanisms underlying IGD and provide a generalisable, cross-treatment neural marker to inform the development of personalised therapeutic strategies.

Funding Statement

This work was supported by the Yunnan Provincial Department of Science and Technology Basic Research Project (202405AC350075; 202501AS070042) and the Innovation Team Program in Philosophy and Social Science of Yunnan Province (Research on psychological adaptation and development of China’s ethnic minority students in border areas.

Disclosure statement

No potential conflict of interest was reported by the author(s).