Introduction
As a typical nonsubstance related addiction, Internet gaming disorder (IGD) is defined as people who are unable to control their desires to play Internet games excessively. Studies have revealed that IGD can lead to severe negative consequences such as impaired physical and psychological states, social deficits, and poor academic performance.Reference Petry, Rehbein and Gentile1, Reference Meng, Deng, Wang, Guo and Li2 In 2013, the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) committee included the IGD in Section III of the DSM-5 as an issue warranting further study,Reference Király, Nagygyörgy, Griths, Demetrovics, Rosenberg and Feder3–Reference Gentile, Choo and Liau5 and a diagnostic criteria for IGD was proposed, which facilitated potential studies.6
Most people like games, however, only a few of them are addicted to them; most of the players play games in a recreational way.Reference Montag and Reuter7 They can play games in a controlled manner without showing psychiatric symptoms of addiction such as craving, conflict, and withdrawal.Reference Viriyavejakul, McFerrin, Weber, Carlsen and Willis8–Reference Wang, Wu, Wang, Zhang, Du and Dong10 These people are defined as recreational Internet gaming users (RGU).Reference Dong, Li, Wang and Potenza11
DSM-5 introduced IGD in the research appendix;Reference Bae, Hong, Kim and Han12 and at the same time, the International Classification of Diseases, 11th edition (ICD-11) is also considering it as a behavioral addiction.Reference Petry, Zajac and Ginley13 Studies have found that the IGD participants share similarities with traditional addictions including higher impulsivity, unsuccessful cognitive control, deficit in decision-making,Reference Ko, Hsieh and Wang14–Reference Dong and Potenza16 higher sensitivity to rewards,Reference Dong, Wang, Du and Potenza17 and neglect their losses in daily-life situations.Reference Wang, Wu, Wang, Zhang, Du and Dong10, Reference Dong, Hu and Xiao18 Studies have also found that IGD participants are more likely to make decisions with greater risk tendencies than healthy controls (HC).Reference Dong and Potenza16, Reference Wang, Wu and Lin19 Comparing to RGU, the IGD participants showed decreased frontal brain responses during processing of losing outcomes, suggesting their decreased loss sensitivity during decision-making.Reference Dong, Li, Wang and Potenza11 These findings were consistent with the results using substance use disorders or pathological gambling participants.Reference Gilman, Calderon, Curran and Evins20–Reference Madden, Petry and Johnson22
Rational decision-making describes one’s ability to make the best choice from multiple alternatives, which is important in exploring ones choice features in risky situations. Probability discounting (PD) tasks measure the trade-offs between reward magnitudes and probabilities to examine the decision-making features under risky situations. Participants need to make a choice between a large amount of money with a low probability of winning and a small amount of money with a high and fixed probability of winning. Choosing the larger amount of money with lower probability of winning showed a higher subjective value for probabilistic rewards, which reflects the tendency to take risks. Although most people would choose the fixed one to avoid getting nothing, the IGD participants preferred to choose the larger amount of money with lower probability of winning.Reference Wang, Wu and Lin19, Reference Madden, Petry and Johnson22 With this feature of PD, in this study, we would use PD task to explore the neuronal network underlying decision-making process of IGD participants.
Previous IGD studies mostly focused on the specific functions of a brain region, but how these brain regions interact together is still lacking in understanding. In this study, we combined two analyses (independent component analysis (ICA) and graph theoretical analysis (GTA)) to explore the differences between the neuronal network of IGD participants and that of RGU using a probability discounting task.
ICA is a data-driven technique that can extract the information about the intrinsic neuronal functional connectivity (FC) without cognitive tasks.Reference Calhoun, Adali, Pearlson and Pekar23 Previously, we found that IGD participants showed impaired executive control network (ECN) when compared with the HC,Reference Wang, Wu and Lin19 and they also exhibited enhanced sensitivity to rewards and an impaired executive control ability when performing a delayed discounting task.Reference Wang, Wu and Zhou24 GTA regards the whole brain as a complex network, composed of a set of nodes and edges.Reference He and Evans25 It quantitatively measures each node by incorporating connectivity information from the complete network, reflecting the integrated nature of local brain activity. Previous studies have demonstrated that human brain anatomical and functional networks have small-world properties,Reference Newman26, Reference Achard and Bullmore27 that is, high level of clustering coefficient and short path length linking all nodes.Reference Kim, Manoach and Mathalon28 Recently, a study has found that there are significant differences in regional nodal characteristics of IGD participants comparing with that of healthy participants during resting state.Reference Wang, Wu and Lin29 The ICA focuses on a specific group of brain areas, while the GTA measures the nodes and edges of the whole brain. These two approaches are complementary in that they look at different aspects of the brain network.Reference Ye and Doñamayor30
The second strength of this study is the subject selection. However, most previous studies have focused on the differences between the IGD participants and HC, which have some limitations. The IGD participants played online games frequently; however, the HC are none or low-frequent game players, who have limited experience with online gaming. Hence, we selected the RGU as comparison group to overcome these limitations. Thus, the current study used RGU as control group to find some features of IGD participants in a PD task.
In our previous studies and other studies, we found the IGD participants’ group show impaired decision-making in probability-discounting task. They were also more impulsive in decision-making and could not effectively control their impulsivity.Reference Wang, Wu, Wang, Zhang, Du and Dong10, Reference Dong, Li, Wang and Potenza11, Reference Dong and Potenza16, Reference Dong, Hu and Lin31 Based on these results, in this PD task, we hypothesized that the IGD participants would chose lower PD when making decisions, which may be related to the impairment of reward circuits and executive control ability. We combined the ICA and GTA methods to explore dysfunctional networks in IGD to gain a better understanding about the neural mechanism of IGD and to provide insights for better treatment strategies for people with IGD.
Methods
Participants
This study conformed to The Code of Ethics of the World Medical Association (Declaration of Helsinki) and was approved by the Human Investigations Committee of Zhejiang Normal University. All participants provided written informed consent before the formal scan session.
Forty right-handed male university students (20 IGD, 20 RGU) were recruited in this study. Only males were included because of the high IGD prevalence in males. There was no significant age difference between the two groups (Table 1). All participants underwent structured psychiatric interviews (mini international neuropsychiatric interview (MINI)) conducted by an experienced psychiatrist.Reference Lecrubier, Sheehan and Weiller32 The MINI results revealed that all participants were free from psychiatric/neurological disorders. No participants reported previous gambling or illicit drugs (e.g., marijuana and heroin) experiences. All participants were instructed not to use any medicine or substances including coffee, tea, and alcohol on the day of scanning.
Table 1 Demographic information and group differences
Note. values are presented as mean ± SD. IAT, Internet addiction test; IGD, Internet gaming disorder; RGU, recreational Internet game users.
IGD participants were selected based on a Young’s Internet addiction test (IAT) and the nine DSM-5 criteria of IGD.Reference Petry, Rehbein and Gentile1, Reference Young33 Participants were diagnosed with IGD if they satisfied the following three criteria: (1) scored 50 or higher on the IAT scores; (2) reached at least five DSM-5 criteria; and (3) spent at least 80% of online time playing games and spent at least 2 hours per day in online games during the last 2 years.
For RGU, first, most of them scored less than 50 in the IAT, reached less than five DSM-5 criteria and were not affecting their daily life. Second, RGU should be a minimum of 2 years and without showing any symptoms of physical or psychological dependence. Third, the RGU played online games more than 14 hours per week and a minimum 5 of 7 days in a week.Reference Dong, Wang, Du and Potenza17 The two groups showed significant differences in IAT and DSM scores (Table 1).
Task and procedure
During the experiment, participants first completed an informed consent form and the Matters of Attention of functional magnetic resonance imaging (fMRI). Second, they were provided with a sample PD task. To motivate participants making their choices seriously, we told them that they would be paid after the experiment according to their choice in a randomly selected trial of the task. If they chose the certain option, he would receive the money in cash. If they selected the probabilistic option, he could select a card from many cards of two colors (red and black) reflecting the probability of receiving the money. Finally, they completed the PD task in the fMRI scanner. During the task, participants were told to make choices between a fixed but small amount of money and a probabilistic but larger amount of money available based on a probability value (i.e., 10 Yuan 100% versus 14 Yuan 40%) (Figure 1).
Figure 1 The timeline of one trial in the PD task in the present study. The fixed (e.g., 10 Yuan 100%) options were presented on the left of the screen and the probabilistic (e.g., 14 Yuan 40%) options were presented on the right of the screen. In the examples, ‘元’ is the monetary unit of currency in China.
The probabilistic choices ranged from 10% to 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, and monetary amounts ranged from 11 to 12.5, 14, 17, 20, 25, 33, 50, and 100 Yuan. There were 81 trials in total and took approximately 15 minutes. Stimuli were randomly presented, and the behavioral data were collected using the E-prime software (Psychology Software Tools, Inc.).
Behavioral data analysis
Two participants were excluded from data analysis because of large head movements and/or choosing the same options in all trials. As a result, the remaining data of 38 participants (18 IGD, 20 RGU) were included in data analysis. PD task contains a large but probabilistic reward against a small but fixed reward. In PD task, discounting means the devaluation of an outcome when the outcome is obtained probabilistically. PD rates were calculated by using a hyperbolic function,Reference Madden, Petry and Johnson22, Reference Yi, Chase and Bickel34 which represented as follows:
$$ {{V}} = {{{A}} \over {\left( {1 + h{\rm{O}}} \right)}},{{O}} = \left({1 \over P} - 1\right) $$
The V represents the discounted value, and the A is monetary reward. P represents the probability to receive rewards, and O represents the odds against receiving the rewards. The h value is a subject specific discounting constant. Notably, smaller values of h might suggest a tendency to take risks and more impulsivity. It can represent cognitive impulsivity features affecting decision-making process. One important procedure for calculating h is to determine indifferent points, which are points that the subjective value of a probabilistic option is equivalent to that of the other fixed option. The analysis procedure was composed of two steps: first, a nonlinear curve-fitting program was applied to data in Origin 7.0 to calculate each subject’s best-fit-h value from Eq (1).Reference Rachlin, Raineri and Cross35 Second, a log 10 transformation of the h values was performed. By this transformation, the distribution of values was normal.Reference Young36
Image acquisition and pre-processing
Functional MRI was performed on a 3T scanner (Siemens, Prisma). The sequence parameters were as follows: repetition time (TR) = 2000 ms; echo time (TE) = 30 ms; flip angle 90°; interleaved sequence; 33 slice per volume; 3 mm thickness; field of view 220 × 220 mm2; and matrix 64 × 64. All stimuli were presented by the Invivo synchronous system (Invivo Company; www.invivocorp.com/) through a monitor in the head coil. Structural images covering the whole brain were collected through a T1-weighted three-dimensional spoiled gradient-recalled sequence (176 slices, flip angle = 15°, echo time = 3.93 ms, slice thickness = 1.0 mm, skip = 0 mm, inversion time 1100 ms, field of view = 240 × 240 mm2, in-plane resolution = 256 × 256).
Imaging analysis was performed using SPM8 (http://www.fil.ion.ucl.ac.uk/spm). Images were slice-timed, reoriented, and realigned to the first volume. Then, T1-co-registered volumes were normalized to an SPM EPI template and spatially smoothed by using an 8-mm full-width-half-maximum Gaussian kernel. No participant was eliminated due to large head motion coefficients based on the criteria (head motion <2 mm and 2°).
Independent component analysis
The group ICA was applied to the preprocessed fMRI data within a toolbox (GIFT v2.0) implemented in Matlab R2012a (http://icatb.sourceforge.net). The fMRI data were reduced through two principal component analysis (PCA) stages.Reference Mitchell37, Reference Reynolds, Richards, Horn and Karraker38 The default component number was 20, and a spatial ICA was conducted to estimate the 20 mutually independent components using Fast ICA algorithm, which is a stochastic process.Reference Calhoun, Adali, Pearlson and Pekar23 ICASSO algorithmReference Bell and Sejnowski39 was performed to remedy the problem, which repeated the ICA analysis multiple times and then output a final set of independent components, providing a measurement of consistency between different ICA runs. Eventually, a single ICA time course and an independent functional spatial map for every subject were obtained.
Component selection
There are two steps for selecting components of interest. Each ICA component spatial map was correlated with probabilistic maps of gray matter (GM), cerebral spinal fluid (CSF), and white matter (WM) within a standardized brain space provided by the MNI templates in SPM5. First, components showing relatively high correlations with WM and CSF and low correlation with GM were considered as artifacts and should be discarded. Components that satisfied the threshold of r 2 < 0.025 for CSF and WM and r 2 > 0.05 for GM were reserved and considered as meaningful. Thus, this analysis would exclude noise related components that represented eye movement, head motion, ventricular pulsations, and other signal artifacts.Reference Himberg, Hyvärinen and Esposito40 The second step was to select the components that were highly correlated with the experimental tasks from the remaining components based on the first analysis. A multiple temporal regression was performed on the ICA time courses with the GLM design matrix to estimate the association between the experimental task and the independent components, which resulted in a set of beta weights for each subject and each condition (probability and certain). Then, an independent one-sample t-test (p < 0.05) on beta weights for the IGD and RGU groups were examined under each task condition. The beta weight of components that differed significantly from zero indicated a significant association with the experimental tasks, whereas components failed to show a significant relationship were diagnosed as task-irrelevant and discarded.
Between-group comparison of components
Components that passed the two criteria were subjected to between-group task-related activity comparison analysis. An independent two sample t-test (p< 0.05) on the beta weights of each remaining component between the RGU and IGD groups was performed in the GIFT toolbox.
Correlations between behavioral performances and brain activities
We analyzed the correlation between behavioral performances and the beta weights for IGD participants and RGU separately. Specifically, we analyzed correlations among beta weights, reaction time (RT), and PD rate (the h values). Furthermore, the correlation between beta weights and the addiction severity (the IAT and DSM scores) was also examined.
Graph theoretical analysis
Components and voxels selection
For this method, ICA was used to detect the strongest PD task-related components. One sample t-tests of beta values for each component were performed to define task-related components.Reference Meda, Stevens, Folley, Calhoun and Pearlson41 According to the above results and the features of the task, we selected three components to perform the remaining analysis. In the study, we used 95 voxels to assess their small-world properties for three brain networks separately. All voxels in each component were sorted from high to low according to their Z scores and the top 95 voxels (most task-related) constituted the three task-related networks.
Estimation of the partial correlations
We used partial correlation to measure connectivity between a given pair of voxels and built undirected graphs respectively for three networks. The partial correlation matrix is a symmetric matrix, filtering out the contributions of all other variables; each off-diagonal element is the correlation coefficient between a pair of variables. The first stage was to estimate the sample covariance matrix S from the data matrix Y= (xi), i = 1,…,95, for each individual. Here, xi was the time series of each ith voxel. If we introduce X= (xj xk) to represent the observations in the jth and kth voxels, Z = Y/X represents the other 93 time series matrices. Each component of S contains the sample covariance value between two voxels (say j and k). If the covariance matrix of [X, Z] was
$$ S = \left( {\matrix{ {\mathop S\nolimits_{11} } \quad {\mathop S\nolimits_{12} } \cr {S_{12}^T} \quad {\mathop S\nolimits_{22} } } } \right)$$
In which, S11 was the covariance matrix of X, S12 was the covariance matrix of X and Z, and S22 was the covariance matrix of Z, then the partial correlation matrix of X, controlling for Z could be defined formally as a normalized version of the covariance matrix, 
${\rm{S}_{xy}} = {\rm{S}_{11}} - {\rm{S}_{12}}S_{22}^{ - 1}S_{12}^T$. Finally, a Fisher’s r-to-z transformationReference Salvador, Suckling, Coleman, Pickard, Menon and Bullmore42 was used on the partial correlation matrix to induce normality on the partial correlation coefficients.
Constructing brain network
An N × N (N = 120 in the present study) binary graph brain network, G, consisting of nodes (brain voxels) and undirected edges (connectivity) between nodes could be constructed by applying a predefined threshold T to the partial correlation coefficients:
$${e_{ij}} = \left\{ {\matrix{ \;{1 \quad{if\;\left| {z\left( {i,j} \right)} \right| \ge T}} \cr {\hskip-1.3pc 0 \quad{otherwise}} }} \right.$$
If the absolute value of correlation between i and j was larger than the predefined threshold T, an edge was said to exist; otherwise it did not exist. The selection of threshold T will be discussed in following sections.
Small-worldness
The small-world properties including clustering coefficient (C net), normalized clustering coefficient (γ), characteristic path length (L net), normalized characteristic path length (λ), global efficiency (E global), local efficiency (E local), and small-worldness (δ). Small-world networks have lower path lengths but higher clustering coefficients, that is, γ = C net/C random > 1, λ = L net/L random = 1.Reference Fisher43, Reference Liu, Liang and Zhou44 Combining these two conditions, we can get a scalar quantitative measurement of small-world networks, small-worldness, δ = γ/λ, which is typically > 1.Reference Watts and Strogatz45 We applied a wide range sparsity threshold T to all correlation matrices. The range of T was determined according to the following criteria: (1) the average degree of all nodes of each network was larger than 2×log (N),Reference Zhang, Wang and Wu46 N is the number of nodes (here, N=120); and (2) the small-worldness scalars of each threshold network was larger than 1.0 for all participants. These criteria guaranteed that the threshold networks had a few spurious edges in sparse properties as possible.Reference Achard and Bullmore27 Thus, the range of our generated sparsity threshold of C9 was 0.07 < T < 0.4 with an interval of 0.01. The range of our generated sparsity threshold of C10 was 0.10 < T < 0.28 with an interval of 0.01 (one RGU subject was excluded because most δ were smaller than 1.0). The range of our generated sparsity threshold of C20 was 0.23 < T < 0.4 with an interval of 0.01 (one RGU subject was excluded because most δ were smaller than 1.0). Next, we calculated global network metrics at each sparsity threshold. Moreover, the area under the curve (AUC) that independent of single threshold was calculated for global network metrics in order to provide a summarized scalar for the topological organization of brain networks.Reference Achard and Bullmore27
Statistical analysis
To determine the differences between the IGD and RGU groups in small-world properties, a two-sample two-tailed t-test with a threshold of p < 0.05 was performed on the AUC of each metric for three task-related networks (C net, L net, E global, E local, γ, λ, and σ) separately. Also, we performed the two-sample t-test on each metric of three networks at each sparsity threshold level.
Relationships between network metrics and behavioral scores
We calculated the correlations between small-world properties that group-level differences of each network and behavioral data and IAT and DSM scores.
Results
Behavioral performance
The independent sample t-test on the logged h values indicated that the IGD group showed lower log (h) values than the RGU group, t = 2.81, p = 0.009. The mean h values for IGD and RGU groups were 2.03 and 4.21, respectively (Figure 2a), suggesting a more rapid rate of PD for IGD than the RGU group (Figure 2b). The R 2 value for PD functions (0.90 for IGD and 0.87 for RGU) represented the accounted variance by Eq. (1). Finally, the data of RT (RT probability - RT certain) were subjected to an independent sample t-test, and the result showed that the RT (RT probability - RT certain) of IGD is much shorter when compared with that of RGU (IGD = 47 ± 132 ms, RGU:129 ± 255 ms, t = 1.22, p = 0.23). Correlation analysis between the log (h) values and the RT showed that they were positively correlated (r = 0.467, p = 004) (Figure 2c).
Figure 2 Probability discounting value differences between the IGD participants and RGU. (a) The IGD participants showed the lower h value than the RGU group. (b) Probability discounting functions for IGD and RGU. Points show medial indifferent points for monetary rewards as a function of the odds against receiving the rewards. (c) Correlation between the log (h) value and reaction time in possible minus fixed option. IGD: Internet gaming disorder; RGU: recreational Internet gaming users. The log (h) represents the probability discounting rate.
ICA results
Independent component
Six out of twenty components (C4, C9, C10, C11, C15, and C20) passed our selection criteria. These six components had a relatively low spatial correlation with cerebral spinal fluid and white while a high correlation with gray. Also, these components showed significant correlation with the experimental task.
Between-group differences
A two-sample t-test was used to determine whether the beta weights that were produced by a regression showed statistically significant difference under probability or certain conditions. C9, C10, and C20 showed significant difference under the two conditions (Table 2).
Table 2 Components that showed significant differences in two-sample t-tests of beta weights
Note. Probability means the probability larger option in the PD task, certain means the certain smaller option in the PD task; RGU: recreational Internet gaming users; IGD: Internet gaming disorder; C: component.
Components that showed significant difference between IGD group and RGU group in beta weights (p < 0.05). The beta weights of C10 under both two conditions showed marginally significant difference across the two groups, and the beta weights of C9 and C20 were modulated in different directions under the certain condition across the two groups.
Under the probability condition, C9 was positively modulated by the condition and the IGD group showed higher task-related activity than RGU. Under the probability and certain conditions, C10 was negatively modulated by the two conditions and the IGD group showed marginally lower task-related activity than RGU. For C20, the IGD and RGU groups were modulated in different directions under the two conditions. C20 was positively modulated in IGD group and negatively modulated in RGU group under the probability and certain conditions. To identify which networks the three components belong to, we contrasted each component of these brain regions with the 14 brain networks.Reference Widyanto, Griffiths and Brunsden47 Consequently, component 9 was involved in left executive control network (ECN), which mainly consists of the prefrontal and parietal cortices. Component 10 was involved in anterior salience network (ASN), primarily including anterior insula and dorsal anterior cingulate cortex. Component 20 was involved in basal ganglia network (BGN), primarily including the striatum and the thalamus (Figure 3).
Figure 3 Brain areas showing differences in IGD participants compared with RGU group ((IGD possible-IGD fixed)-(RGU possible-RGU fixed)).
Correlation analysis results
We analyzed the correlation between task behavioral performances and the beta weights of C9, C10, and C20. There was a significant positive correlation between the beta values (betaprobability – betacertain) of C10 and the logged h values only in RGU (r = 0.481, p = 0.032) (Figure 4).
Figure 4 Correlation between the log (h) value and beta values (probability-certain) of C10. The log (h) represents the probability discounting rate. C: component.
GTA results
Small-world properties
The results showed that the three functional networks of all participants had similar characteristic path lengths (λ≈1) but higher clustering coefficients (γ>1), and the small-worldness, δ=γ/λ (> 1). Hence, the IGD group and RGU all showed typical features of small-world topology in these networks. Despite common small-world architecture, the results revealed the significant differences in small-world attributions between IGD participants and RGU. A two-sample t-test on AUC values showed group differences in these three networks. For the BGN, the IGD participants showed significant higher clustering coefficient than RGU on AUC values (t = 2.388, p = 0.022). Although other properties did not reach the significant, the trends are in existence. The IGD participants showed higher λ (t = 1.749, p = 0.088) and lower E global (t = −1.742, p = 0.089) than RGU. For the ASN, the IGD participants showed lower clustering coefficient (t = −1.807, p = 0.079) and higher E global (t = 1.755, p = 0.088) than RGU (Table 3).
Table 3 Small-world properties that showed significant differences in two-sample t-tests of AUC values
Note. For the C20, the IGD participants showed significant higher clustering coefficient than RGU on AUC values (t = 2.388, p = 0.022). AUC: the area under the curve; RGU: recreational Internet gaming users; IGD: Internet gaming disorder; C: component; C net: clustering coefficient; E global: global efficient; λ: normalized path length.
For each threshold level, the IGD participants showed significantly lower clustering coefficient and higher E global, and shorter path length in the ASN (Figure 5a). In the BGN, the IGD participants showed significantly higher cluster coefficient and higher E local (Figure 5b) and lower E global and longer path length and longer normalized path length in each 18 threshold levels (Figure 5c). In the ECN with 34 threshold levels, the IGD group showed lower E global and longer path length than RGU group (not significant).
Figure 5 (a) The small-world properties of anterior salience network over a range of sparsity threshold. a (1) The IGD participants showed lower clustering coefficient than RGU; a (2) The IGD participants showed shorter characteristic path length than RGU; a (3) The IGD participants showed higher global efficient than RGU. Black dots above indicate significant group difference (p < 0.05). (b) The small-world properties of reward network over a range of sparsity threshold. b (1) The IGD participants showed higher clustering coefficient than RGU; b (2) The IGD participants showed higher local efficient than RGU. Black dots above indicate significant group difference (p < 0.05). (c) The small-world properties of reward network over a range of sparsity threshold. c (1) The IGD participants showed longer characteristic path length than RGU; c (2) The IGD participants showed lower global efficient than RGU; c (3) The IGD participants showed longer normalized path length than RGU. Black dots above indicate significant group difference (p < 0.05).
Correlation between network metrics and behavioral measures
We explored the relationships of network metrics with both behavioral and addiction severity (the DSM and IAT scores) of all participants. Correlation analysis demonstrated that the observed global abnormalities were correlated with the IAT scores and behaviors. Significant positive correlations (p < 0.05) were found between IAT scores and AUC values of clustering coefficient in the BGN (r = 0.324, p = 0.047) (Figure 6a). The AUC value of E global was negatively correlated with the logged h values in the ASN (r = −0.319, p = 0.054) (Figure 6b). In the ECN, the AUC values of normalized cluster coefficient negatively correlated with the RT (r = −0.284, p = 0.085) (Figure 6c). Although some of these correlations did not reach the statistical significance, the trends are in existence.
Figure 6 Correlation between small-world properties and behavior and IAT scores of three brain networks across all participants. (a) The significant positive correlations (p < 0.05) were found between IAT scores and the AUC values of cluster coefficient in reward network. (b) The AUC values of E global was negatively correlated with the log (h) values in the anterior salience network (p < 0.05). (c) The AUC values of normalized cluster coefficient negatively correlated with the RT (p < 0.05). AUC: the area under the curve; the log (h) represent the discounting rate; RT: reaction time. E global: global efficient.
Discussion
The current study explored the potential changed neuronal networks of IGD participants using ICA and GTA. The behavior results showed that the IGD group was associated with lower PD of risky options and spent less time during decision-making compared to RGU. The lower h values might suggest a tendency to take risks and increased level of impulsivity.Reference Dai, Harrow, Song, Rucklidge and Grace48–Reference Holt, Green and Myerson50 This suggested that the IGD group was poor in risk evaluation and impulse control compared to RGU. Positive correlation between PD rate and RT indicated that the ones who were higher impulsivity with more quickly in making decisions without extra thinking. In imaging results, the ICA and GTA results both showed dysfunction networks connectivity in ASN and BGN.
Executive control network
The ECN was involved in cognitive control and goal-directed behaviors.Reference Shirer, Ryali, Rykhlevskaia, Menon and Greicius51 Previous studies have revealed that brain training games could improve executive functions for individuals.Reference Krmpotich, Tregellas, Thompson, Banich, Klenk and Tanabe52, Reference Sutherland, Mchugh, Pariyadath and Stein53 However, it was not the same for excessive Internet games. A study using the ReHo method found that IGD participants showed higher nervous activity of ECN compared with HCs.Reference Rui, Taki and Takeuchi54, Reference Rui, Yasuyuki and Hikaru55 Our previous study found that the IGD group exhibited stronger FC in ECN when selecting small but immediate options.Reference Wang, Wu and Zhou24 In this study, the ICA results showed that the IGD participants enhanced FC in ECN compared to RGU in probability and certain conditions, which may suggest that the ECN engaged more in IGD participants than RGU in making decisions. Even so, the IGD participants still failed to control their risky behaviors and preferred lower probability and higher risky choices. In behavioral performance, the IGD group showed lower PD than RGU, which suggests that the IGD group is more impulsive than RGU.Reference Dai, Harrow, Song, Rucklidge and Grace48, Reference Reynolds, Richards, Horn and Karraker49 Numerous studies had detected the possible neural mechanisms of the IGD and suggested that it may be related to the impaired cognitive control.Reference Wang, Wu and Zhou24, Reference Dong, Lin, Hu, Xie and Du56 Hence, we speculated that the IGD participants impaired the ability to inhibit their impulse and to make better choice under a risky circumstance.
Anterior salience network
The ASN plays an important role in identifying relevant internal and external stimuli in order to guide and modulate cognitive behaviors.Reference Bonnelle, Ham and Leech57 One theory is that the salience network, which includes the anterior cingulate cortex and insula, regulates dynamic changes in other brain core networks to modulate cognitive behaviors.Reference Sridharan, Levitin and Menon58, Reference Uddin, Supekar, Ryali and Menon59
For ICA results, the IGD participants showed lower beta weights in ASN under the PD task, which suggested that the IGD participants decreased FC of ASN compared to RGU. A traumatic brain injury study has found the disruption of this network could lead to inefficient cognitive control.Reference Uddin, Supekar, Ryali and Menon59, Reference Menon60 Additionally, reduced structural and effective connectivity within the salience network has been found to be related to impaired individuals’ cognitive control function.Reference Sridharan, Levitin and Menon58, Reference Menon60 We found that the beta weights of ASN were positively correlated with the log(h) values, which suggested that the higher risk, the weaker FC of the ASN. It was consistent with our ICA results. These results suggested that the IGD participants used less time to consider the different condition and could not efficiently control impulsivity and guide their behaviors.
For small-world results, significant group differences were observed in the clustering coefficient, characteristic path length, and E global showed at each threshold level. IGD participants had lower values in clustering coefficient, lower characteristic path length, and higher E global values than RGU. A small amount of long-range connections are good not only for connection separation of local nerve and connection cost constraint but also for long distance information transmission and integration for brain regions.Reference Delbeuck, Linden and Collette61 A study has showed when the E global of MCI patients has increased, their ability to process long-distance information ability has also improved.Reference Wang, Zhao and Jiang62 In our study, the small-world results showed that the inner brain regions’ connections have difference in IGD, with lower clustering coefficient than RGU. This may reveal that the ability of IGD participants to process local information is impaired. The IGD participants showed shorter path length and higher E global, which suggested that the ability to process local information was decreased, under the compensatory mechanism, the ability to process the long distance information would have been improved. The E global had negative correlation with the PD of the ASN. In addition, the lower logged h values, the higher global efficient also suggested that the ASN of IGD was impaired; they cannot effectively identify the external risks and failed to guide their behavior in a more rational way when making decisions. Combining these two results, we concluded that the disturbed connectivity within ASN may be related to the failure of risky behavior regulation, and this inefficient regulation was contributed in the impaired cognitive control in IGD participants.
Basal ganglia network
Neuroimaging studies suggested that the BGN plays a critical role in mediating the subjective reward effects.Reference Power, Goodyear and Crockford63, Reference Dong, Huang and Du64 Reward circuits mainly involve in evaluating the values of both stimuli and rewards before making a decision.Reference Zhao, Tang, Feng, Li and Sui65
For ICA results, the IGD participants showed stronger FC in the BGN than the RGU, and this phenomenon was congruent with previous studies.Reference Yuan, Wei and Yu66, Reference Xing, Yuan and Bi67 The different levels of FC in BGN might be explained by the different sensitivities to the rewards.Reference Moussa, Steen, Laurienti and Hayasaka68, Reference Schmidt, Denier and Magon69 A review reported significant hyperactivity in BGN of gambling disorder.Reference Meng, Deng and Wang70 Some studies have suggested that the reward circuits have been changed in IGD participants.Reference Volkow, Wang, Fowler, Tomasi, Telang and Baler71–Reference Dong, Wang and Potenza73 The IGD participants showed more sensitivity toward reward and reduced sensitivity of punishment.Reference Dong, Hu and Xiao18, Reference Wang, Wu and Zhou24, Reference Lorenz, Gleich, Gallinat and Kühn74 A study has shown that the IGD is associated with enhanced FC in BGN, which suggests IGD participants have enhanced sensitivity to rewards when making decisions.Reference Wang, Wu and Zhou24 In comparison with RGU, the participants with IGD showed shorter RT when choosing the risky options, which might suggest that the IGD participants who made their decisions hastily irrespective of the potential loss.
For small-world results, the IGD group showed higher clustering coefficient, higher E local, longer path length, and lower E global when compared with the RGU group at each threshold level. It has been suggested that lower clustering coefficient and lower E local mean relatively sparse local connectivity; a short path length and high E global represent the high synchronization of brain functional networks.Reference Kaiser and Hilgetag75 Higher clustering coefficient reflects disrupted neuron integration between distant regions, which has relatively sparse long-distant and relatively dense short-distant functional connections in the Internet addiction disorder group.Reference Wee, Zhao and Yap76 In this study, the clustering coefficient of the reward network has positively correlation with IAT scores, the higher IAT scores, the higher the clustering coefficient. This reflects disrupted neuronal integration between distant regions with the severity of addiction. The addictive behaviors may lead to the disconnection of long distance connections and may encourage the establishment of short distance connections within clusters as an alternative path to keep information transmission between two distant regions. However, establishment of short distance connections may introduce abnormal clusters, which increases the risk of creating an uncontrolled information flow through the entire network. The IGD groups were more likely to choose the lower PD and spend less time making decisions in PD task. The IGD participants preferred the large monetary rewards with low probability to the certain small monetary rewards. Taken together, the dysfunction network connectivity in IGD participants might suggest that the IGD groups are more sensitivity to reward and neglect the potential risky when making choices.
Limitations
Several limitations should be mentioned for this study. First, a nongamer group with no experience of gaming can be included to look at difference between the three groups (healthy control, IGD, and RGU). The results will be more interesting if a nongamer group is included. In addition, only male college students were recruited for this study. It will be necessary for further research to include female participants and to explore gender effect in IGD. Second, the IAT is not a specific test for IGD and we did not measure the impulsivity by Barratt Impulsiveness Scale-11 (BIS-11) or the UPPS Impulsive Behavior Scale . The lack of behavioral support for many of the results only provides a possible explanation that needs to be further explored. In future research, we should collect more behavioral data, which could provide more support to imaging results. Third, undirected, unweighted networks are built in the present study.Reference Lynall, Bassett and Kerwin77 A weighted network could provide more information in the future studies.Reference Van Den Heuvel and Hulshoff Pol78 Besides, data smoothing during preprocessing on network analysis needs to be improved. Spatial smoothing could result in artificial correlations between voxels by previous voxel-based network analysis studies.Reference Hayasaka and Laurienti79 Partial correlation will remove local correlations (including the correlation caused by smoothing) but preserve the unique voxel variance. Further research is needed to determine the full effect of smoothing on correlations and partial correlations, including their relationship to variability between participants.
Conclusions
The present study examined the functional organization of the brain networks for the IGD participants under a PD task using ICA and GTA. The results revealed connectivity of three brain networks (ASN, ECN, and BGN) were altered in IGD participants compared with RGU. Taking the role of these networks into consideration, the current study suggests that Internet game addicts are more impulsive in decision-making and cannot effectively control their impulsivity because of their impaired executive control ability. The changes in these networks were involved in decision-making and cognitive control and may be a key mechanism preventing recreational game players from the risk of developing addiction.
Disclosures
The authors declare that they have nothing to disclose.
This research was supported by the National Science Foundation of China (31371023). The funders have no further role in the study design; in the data collection, analyses, and interpretation; the decision to publish, or preparation of the manuscript.
Data accessibility
Data are available upon request from the corresponding author at dongguangheng@zjnu.edu.cn.
