Connectome-based prediction of problematic use of social media in adolescents: Findings from the ABCD study

Abstract

Problematic use of social media (PUSM) is a major public health concern estimated to affect 35% of adolescents. However, data-driven research to identify neural networks predictive of PUSM in adolescents remains limited. The aim of this study was to utilize connectome-based predictive modelling (CPM), a machine-learning approach that employs whole-brain functional connectivity data, to predict PUSM severity and identify underlying neural networks in adolescents. We included 2294 participants from the Adolescent Brain Cognitive Development study (M_age = 10.03, 50.6% female) who had resting-state functional magnetic resonance imaging (fMRI) data at baseline and PUSM scores at the four-year follow-up. CPM with 10-fold cross-validation was applied to resting-state fMRI data and PUSM scores. CPM successfully predicted PUSM scores and identified connectivity within and between multiple large-scale neural networks predictive of PUSM severity, which could be categorized into two key systems: (i) a cognitive control and self-regulation system consisting of the default mode, frontoparietal, and medial frontal networks, and (ii) a perceptual-motor integration system consisting of the visual area 1 and sensorimotor networks. The large-scale networks identified in the present study provide mechanistic insight into PUSM vulnerability and represent potential targets for personalized interventions. Future research should aim to replicate and extend the current results to refine prevention and treatment approaches.

1. Introduction

Screen media activity (SMA) among developing youth is a significant public health concern (Hutton et al., 2024). Currently youth aged 13–17 years spend considerable time engaging in SMA, especially social media. The widespread use of social media has reshaped adolescent experiences, offering highly stimulating and rewarding platforms on which boundaries between healthy and problematic engagement are blurred. Though not formally recognized as a clinical diagnosis, problematic use of social media (PUSM) has been reported by 35 % of adolescents (Cheng et al., 2021). PUSM shares features with other behavioral addictions (Andreassen, 2015; Cataldo et al., 2022; Moretta et al., 2023), such as gaming disorder and gambling disorder, which are characterized in the ICD-11 by impaired control, increased prioritization, and continuation of gaming/gambling despite negative consequences (World Health Organisation, 2019a, 2019b). Other components, such as tolerance and mood modification, are relevant to PUSM as with other addictive behaviors (Fournier et al., 2023; Griffiths, 2005). Research indicates that the severity of PUSM may be greater in adolescents with attention-deficit/hyperactivity disorder and internalizing psychopathology, such as depression and anxiety (Boer et al., 2020; Dekkers and van Hoorn, 2022; Settanni et al., 2018; Shuai et al., 2021; Thorell et al., 2024; Wang et al., 2017). Despite growing recognition of the need to address PUSM in adolescents, including acknowledgement in the US Surgeon General’s Advisory on Social Media and Youth Mental Health (Office of the Surgeon General, 2023), there is limited research identifying neural networks predictive of PUSM within developing youth. Identifying at early stages of human development potential risk factors for PUSM could provide insight into PUSM development and inform intervention approaches.

Previous resting-state neuroimaging studies of internet-enabled addictive behaviors have provided preliminary evidence of neural correlates that may underlie PUSM. These studies have reported altered functional connectivity spanning several networks in individuals with problematic usage of the internet (PUI) and problematic smartphone use (PSU), including the default mode, salience, frontoparietal, visual attention, and cognitive control networks (Áfra et al., 2024; Ahn et al., 2021; Hong et al., 2013; Liu et al., 2022; Wadsley and Ihssen, 2023; Wang et al., 2017; Wang et al., 2019). However, PUSM has typically been considered within the broader frameworks of PUI and PSU, which are umbrella terms encompassing various behaviors such as online gaming, shopping, and pornography use. Thus, much of the current neuroimaging literature misses distinct features of PUSM and dilutes the specificity of the unique characteristics and experiences associated with social media (e.g., images that trigger social comparison), which may interact differently with underlying youth vulnerabilities to influence PUSM-specific outcomes, such as issues around body image in girls (Çimke and Gürkan, 2023; Saud et al., 2019).

Although numerous brain networks have been implicated in internet-enabled addictive behaviors, significant challenges remain in establishing clinically useful neuromarkers. The existing literature typically relies on seed-based functional connectivity analyses, which focus on predefined regions of interest and limit insights into whole-brain connectivity patterns (Shen et al., 2017; Whelan and Garavan, 2014). Moreover, traditional neuroimaging methods that use simple correlation or regression techniques often risk overfitting, reducing the generalizability of behavioral predictions. Connectome-based predictive modelling (CPM) is one method of addressing these issues by applying machine-learning techniques to whole-brain functional connectivity data to develop predictive brain-behavior models (Finn et al., 2015; Shen et al., 2017). By incorporating cross-validation procedures with model-testing in held-out samples, CPM protects against overfitting and enhances model rigor and generalizability. Moreover, CPM is a fully data-driven approach that does not require a priori selection of regions/networks and facilitates the identification of “neural fingerprints;” i.e., networks subserving specific behaviors, such as PUSM (Finn et al., 2015; Shen et al., 2017).

Previously, we have used brain measures obtained at earlier developmental stages to understand potential clinical concerns later in life, including with respect to screen media activity. For example, we have investigated how brain structural covariation patterns that had been previously linked to early initiation of alcohol use in two samples of adults related to SMA and associated concerns (internalizing and externalizing features and sleep duration and difficulties) cross-sectionally and longitudinally in late childhood to early adolescence in the Adolescent Brain Cognitive Development (ABCD) study (Zhao et al., 2021; Zhao et al., 2022; Zhao et al., 2023; Zhao et al., 2024). Similarly, resting-state functional magnetic resonance imaging (rs-fMRI) ABCD data at baseline (youth aged 9–10 years) related to high-frequency SMA cross-sectionally to multiple clinical concerns and potential risk factors, with patterns persisting over a several year period (Song et al., 2023). In this study, we also observed a shift from high-frequency video-centric SMA to high-frequency socio-communication SMA occurring around the ages of 12–13 years (Song et al., 2023). A better understanding of how brain features during childhood relate to clinically relevant features later in development may help inform early interventions, and this may be particularly relevant for PUSM that may emerge during early adolescence.

The aim of this study was to use CPM to identify neural networks predictive of PUSM in early adolescence using rs-fMRI data from late childhood. We hypothesized that patterns of whole-brain functional connectivity during late childhood could be utilized to predict the severity of PUSM in early adolescence. To test this hypothesis, CPM was applied to rs-fMRI data acquired at baseline (youth aged 9–10 years) in the ABCD study to predict PUSM severity at the four-year follow-up (aged 13–14 years). Although the ABCD study includes two- and three-year follow-up assessments of PUSM, we specifically chose to predict PUSM severity at the four-year follow-up for several developmental and methodological reasons: (i) prior research suggests increased prevalence and frequency of social media use in adolescents than in children, with problematic patterns more likely to emerge and show greater interindividual variability in adolescents, (ii) predicting the four-year PUSM outcome allows us to focus on longer-term developmental risk, whereas earlier follow-ups may capture more transient or constrained patterns of social media use, and (iii) most social media platforms in the United States restrict account creation to individuals aged 13 years or older, consistent with the Children’s Online Privacy Protection Act (Nagata et al., 2025; Paakkari et al., 2021; Rideout, 2015; US Surgeon General’s Advisory, 2023). Additionally, rs-fMRI was used because resting-state functional connectivity reflects relatively stable, intrinsic patterns that emerge during childhood and may shape later behavior, rather than task-evoked responses or exposure-driven effects (Plitt et al., 2015; Uddin et al., 2025). For example, individual-level differences in resting-state connectivity during childhood have been shown to predict various subsequent behaviors and concerns, including future substance use initiation, autistic tendencies, and depressive and anxiety disorders (Kardan et al., 2025; Pawlak et al., 2022; Plitt et al., 2015). In the context of the present study, individual differences in resting-state connectivity patterns at baseline may be best conceptualized as potential neurodevelopmental vulnerabilities that precede the onset of PUSM. Such patterns may help identify children who are neurodevelopmentally predisposed subsequently to engage with social media in a more persistent, dysregulated, seemingly automatic way once access and use increase during early adolescence. This framework resonates with models of behavioral addictions, including the I-PACE model, which also propose that person-level neurobiological characteristics may interact with later environmental exposures to confer risk for problematic engagement with addictive activities, such as use of social media (Brand, Müller, et al., 2025; Brand et al., 2019; Brand et al., 2016). Building on this framework, CPM allows us to use an advanced machine-learning approach to test our hypothesis by leveraging whole-brain resting-state connectivity patterns acquired in late childhood to explore the extent to which they may predict PUSM severity years later.

2. Materials and methods

2.1. Participants

We used the baseline and four-year follow-up data from the ABCD study (5.1 data release), which is a large ongoing longitudinal project that collects behavioral and neuroimaging data from 21 research sites in the United States. At baseline, 11,875 children aged 9 to 10 years were recruited for the ABCD study. Comprehensive details on recruitment and data collection have been published elsewhere (Karcher and Barch, 2021; Luciana et al., 2018). For the CPM analysis, we included participants who had rs-fMRI data at baseline and PUSM scores (measured with the Social Media Addiction Questionnaire) at the four-year follow-up. We excluded participants who did not have valid fMRI data or had missing data on PUSM scores or sociodemographic variables (see Fig. 1). IRB approval and informed consent were obtained at ABCD sites. We analyzed de-identified data and received exemption from review by the Yale IRB and the Yale Human Investigation Committee.

Fig. 1.

The process by which the final analytic sample was derived from the ABCD study.

2.2. PUSM measure

In the ABCD study, the 6-item Social Media Addiction Questionnaire (SMAQ), based on the components model of addictions (Griffiths, 2005), was used to measure salience, tolerance, mood modification, relapse, withdrawal, and conflict associated with PUSM (Bagot et al., 2022). Participants who reported having at least one social media account were prompted to complete the SMAQ. Rated on a 6-point Likert scale from 1 (never) to 6 (very often), the measure yielded a total score ranging from 6 to 36. Higher scores indicated greater severity of PUSM. The SMAQ was introduced at the two-year follow-up of the ABCD study and demonstrated good internal consistency reliability (McDonald’s ω = 0.86) (Bagot et al., 2022).

2.3. Neuroimaging data acquisition/preprocessing and functional connectivity

The ABCD study acquired fMRI data from four five-minute resting-state scans, during which participants passively viewed a cross hair with eyes open (Casey et al., 2018). With imaging parameters harmonized across the study sites, fMRI data were acquired on Siemens Prisma, Phillips, and GE 750 3T scanners. Simultaneous multi-slice/multiband echo-planar imaging rs-fMRI scans (high spatial and temporal resolution) with rapid integration distortion correction were obtained, and the following scanning parameters were used: repetition time (TR) = 800 ms, echo time (TE) = 30 ms, flip angle = 52°, field of view (FOV) = 216 × 216 mm, voxel size = 2.4 × 2.4 × 2.4 mm³, matrix = 90 × 90, 60 slices, multiband acceleration factor = 6 (Casey et al., 2018). The fMRI preprocessing, conducted by the ABCD Data Analysis, Informatics, and Resource Center (DAIRC), included corrections for gradient nonlinearity distortions, B0 inhomogeneity distortions, and head motion, as well as between-scan alignment and registration to T1-weighted structural images (Chaarani et al., 2021). A detailed description of image preprocessing by the ABCD DAIRC has been reported in prior work (Chaarani et al., 2021; Hagler Jr et al., 2019). From the four available rs-fMRI runs, we selected the single run with the lowest frame-to-frame motion for each participant for subsequent analysis.

Further image preprocessing was conducted using BioImage Suite, as described in detail elsewhere (Finn et al., 2015; O’Connor et al., 2025; Rosenberg et al., 2016). The Shen 268-node atlas was used to define network nodes, given that it provides whole-brain coverage (Shen et al., 2013). A concatenation of a series of linear and non-linear registrations between the functional images, anatomical scans, and the MNI brain was performed to warp the atlas from MNI space into single-subject space (Rosenberg et al., 2016). Each transformation pair was calculated independently, combined into a single transform, and inverted to warp the atlas into native space to minimize interpolation error. For each of the 268 nodes, mean time courses were calculated, and node-by-node pairwise Pearson’s correlations were computed. Then, correlation coefficients were Fisher’s z-transformed to generate symmetric 268 × 268 functional connectivity matrices for each participant. Edges within a matrix represented connectivity strength between two nodes (Shen et al., 2017; Yip et al., 2019).

2.4. Connectome-based predictive modelling

We utilized CPM to develop neural predictive models for PUSM in adolescents. Analyses were conducted with a validated MATLAB script (Shen et al., 2017), with connectivity matrices and SMAQ scores entered as inputs. In order to assess model performance, we used 10-fold cross-validation, where participants were randomly divided into 10 subsets or “folds,” with the model trained on 9 folds and tested (to predict PUSM scores) on the remaining fold. This process was repeated for each fold until all participants had a predicted SMAQ score. To mitigate potential bias from random fold assignment, the entire 10-fold cross-validation procedure was repeated 100 times.

All edges in the connectivity matrices of the training set were correlated with SMAQ scores using Pearson’s correlation in order to construct the CPM model. We used partial correlation when controlling for covariates, specifically age, sex, race/ethnicity, parental marital status, parental education, family income, and mean frame-wise displacement. Edges that showed correlations significant at p < 0.05 with SMAQ scores in the training set were selected to form positive and negative predictive networks. Positive networks were characterized by increased edge weights (i.e., connectivity) being associated with increased SMAQ scores while negative networks were characterized by decreased connectivity being associated with increased SMAQ scores. These suprathreshold edges were then aggregated into positive and negative network connectivity indices for each participant, and all subsequent modelling and inference were conducted at the network level rather than at the level of individual connections. A linear model was fit between SMAQ scores and the single-subject network connectivity indices derived from the training set. In the final step of CPM, the linear model generated in the training set was applied to the single-subject summary values (calculated using the positive and negative predictive networks also identified in the training set) of participants in the test set to predict SMAQ scores. To assess model performance, Pearson’s r correlation and partial correlation (controlling for the aforementioned covariates of age, sex, race/ethnicity, parental marital status, parental education, family income, and mean frame-wise displacement) between the observed and predicted SMAQ scores were used. Permutation testing was conducted with 1000 permutations to assess statistical significance.

2.5. Network anatomy

Predictive networks were summarized at multiple levels of data reduction, including edges, nodes, and networks (Ibrahim et al., 2022; Yip et al., 2019). This study summarised networks based on overlap with macroscale brain regions, such as the prefrontal area (Finn et al., 2015), and overlap with canonical networks, such as the default mode network (DMN) (Noble et al., 2017).

3. Results

3.1. Participants

As indicated in Table 1, the CPM analysis included 2294 participants who were around 10 years old on average (SD = 0.61) at baseline. Around 50.6 % (n = 1161) of participants were female. The majority of participants were White (n = 1374, 59.9 %) and had parents who were married (n = 1628, 71.0 %). At the four-year follow-up, participants had an average SMAQ score of 11.80 (SD = 5.56).

Table 1.

Baseline demographic characteristics of ABCD participants included in the CPM analysis (N = 2294).

Variable	N ( %)
Age
Mean (SD)	10.03 (0.61)
Sex
Male	1133 (49.4 %)
Female	1161 (50.6 %)
Race/ethnicity
White	1374 (59.9 %)
Black	209 (9.1 %)
Hispanic	452 (19.7 %)
Asian	46 (2.0 %)
Other	213 (9.3 %)
Family income
≥$200,000	304 (13.3 %)
$100,000-$200,000	724 (31.6 %)
$50,000-$100,000	686 (29.9 %)
<$50,000	580 (25.3 %)
Parental education
Postgraduate degree	836 (36.4 %)
Bachelor’s degree	641 (27.9 %)
Some college	608 (26.5 %)
High school diploma or equivalency	140 (6.1 %)
< High school	69 (3.0 %)
Parental marital status
Married	1628 (71.0 %)
Not married	666 (29.0 %)
Hours of use of social media (typical weekday)
None	1897 (82.7 %)
< 30 min	222 (9.7 %)
30 min	110 (4.8 %)
1 hour	36 (1.6 %)
2 hours	20 (0.9 %)
3 hours	2 (0.1 %)
4+ hours	7 (0.3 %)
Hours of use of social media (typical weekend)
None	1880 (82.0 %)
< 30 min	192 (8.4 %)
30 min	124 (5.4 %)
1 hour	49 (2.1 %)
2 hours	19 (0.8 %)
3 hours	11 (0.5 %)
4+ hours	19 (0.8 %)

3.2. Prediction of PUSM severity

CPM successfully predicted PUSM severity (SMAQ scores) when not controlling for the pre-specified covariates (combined positive and negative networks: r = 0.066, p = 0.003). The final model selected controlled for all of the pre-specified covariates and successfully predicted PUSM severity (combined positive and negative networks: r = 0.055, p = 0.016). The statistical significance of this finding was confirmed using permutation testing (p = 0.02).

3.3. Overlap with macroscale brain regions

Regarding the spatial extent of networks, a total of 2507 edges were identified, including 1212 positive edges and 1295 negative edges (i.e., 7.0 % of possible connections). Positive and negative PUSM networks are displayed in Fig. 2. In the positive network, the nodes with the most connectivity (i.e., highest-degree nodes) included (i) occipital nodes connected to limbic, cerebellar, parietal, temporal, prefrontal, subcortical, brainstem, and insular nodes, and (ii) a temporal node connected to parietal, limbic, subcortical, other temporal, motor strip, and insular nodes. Highest-degree nodes in the negative network included a limbic node with connections to prefrontal, temporal, subcortical, brainstem, cerebellar, other limbic, and insular nodes.

Fig. 2.

Positive and negative PUSM networks. A-B) The 268 nodes are arranged by macroscale brain regions, following an approximate anterior-to-posterior anatomical order (i.e., from the top to the bottom of the circle plots). Longer-range connections are represented by longer lines. In each plot, the left side indicates the right hemisphere, while the right side corresponds with the left hemisphere.

3.4. Overlap with canonical brain networks

Fig. 3 presents connections within and between canonical networks for the positive and negative networks. The comparison of networks indicated that the positive network included relatively more connections between the DMN and the medial frontal and frontoparietal networks (Fig. 3C). Considering the comparison of networks, the negative network included relatively more connections (i) between the DMN and the visual area 1 network, (ii) between the sensorimotor network and the medial frontal and visual area 1 networks, and (iii) within the sensorimotor and visual area 1 networks. Overall, the DMN emerged as the network with the most connections (both positive and negative).

Fig. 3.

Summary of PUSM networks based on overlap with large-scale neural networks. A-B) The number of edges connecting nodes within/between each network are represented as cells. Darker cells indicate more edges. C) Cells within this matrix show the number of positive edges minus the number of negative edges connecting nodes within/between each network (i.e., the positive minus negative network). Red indicates more edges in the positive network, while blue represents more edges in the negative network. MF, medial frontal; FP, frontoparietal; DMN, default mode network; Mot, sensorimotor; VI, visual area 1; VII, visual area 2; VAs, visual association; SAL, salience; SC, subcortical; CBL, cerebellum.

4. Discussion

This study is the first to use CPM to identify neural networks predictive of PUSM severity. Notably, baseline rs-fMRI data collected when participants were 9–10 years old successfully predicted PUSM severity four years later. These findings suggest that childhood patterns of whole-brain functional connectivity may capture vulnerabilities that predispose individuals to PUSM in early adolescence. The PUSM model presented several large-scale networks predictive of PUSM, including the default mode, sensorimotor, visual area 1, medial frontal, and frontoparietal networks. Advancing toward individual-level predictions for PUSM is an important step toward optimizing personalized health, as it may facilitate early detection and tailored preventative interventions for adolescents, which are currently limited for internet-enabled addictive behaviors (Park et al., 2022; Park et al., 2025). Such approaches not only align with principles of evidence-based psychiatry, but also with the expressed needs and priorities of individuals with internet-enabled addictive behaviors (Bzdok and Meyer-Lindenberg, 2018; Park et al., 2021). However, before translating results from the present study into public health practice, future research should work towards replicating the findings and further refining the predictive model (Yip et al., 2019).

Recent research on behavioral and substance addictions has described roles for large-scale brain network dysfunction, encouraging systems-level interpretations to understand addictions (Antons et al., 2023, 2024; Garrison et al., 2023; Sutherland et al., 2012; Yan et al., 2021). Accordingly, we have focused on interpreting our findings at the network- and systems-level rather than specific nodes with altered connectivity. The predictive PUSM network had connections between and within several well-established neural networks, with two key systems identified: (i) a cognitive control and self-regulation system consisting of the default mode, medial frontal, and frontoparietal networks, and (ii) a perceptual-motor integration system consisting of the visual area 1 and sensorimotor networks. We propose that in children, decreased connectivity between and within networks in the cognitive control and self-regulation system and increased connectivity between networks in the perceptual-motor integration system may predict greater PUSM severity during early adolescence, as shown in the theoretical network model of PUSM in Fig. 4. The theoretical model further proposes that greater connectivity between the two systems may be predictive of future PUSM severity. Given that machine-learning and neuroimaging research for PUSM are in early stages compared to substance use disorders and other behavioral addictions (Chhetri et al., 2023; Kuss and Griffiths, 2012; Mak et al., 2019), this theoretical model may facilitate interpretation of findings derived from patterns of whole-brain functional connectivity. The model may also guide the development of testable hypotheses in future theory-driven neuroimaging studies and network-informed interventions for PUSM (Yip et al., 2020).

Fig. 4.

Theoretical network model predictive of PUSM. The red lines show that stronger connectivity between networks is predictive of greater PUSM severity. The blue lines indicate that weaker connectivity within and between networks is predictive of greater PUSM severity.

As part of the identified cognitive control and self-regulation system, the DMN emerged as the most informative network in predicting PUSM severity in adolescents, consistent with findings from previous CPM studies of internet gaming disorder in adults (Song et al., 2021; Zhou et al., 2022). At rest, the DMN supports self-referential processes such as autobiographical memory, social cognition, and imagination of the future, and typically deactivates during externally-focused tasks (Buckner et al., 2008; Spreng et al., 2009). Altered DMN connectivity has been implicated in psychopathology (e.g., depression, anxiety, internet addiction), as demonstrated by the previously proposed “triple network model of major psychopathology” that includes the default mode, salience, and frontoparietal networks (Menon, 2011; Wang et al., 2017).

In the present study, increased DMN connectivity with the visual area 1 network and decreased connectivity with the medial frontal and frontoparietal networks at baseline were predictive of greater PUSM severity four years later. These patterns of resting-state functional connectivity at baseline suggest that as individuals develop from late childhood to early adolescence and become exposed more to social media, individual differences evident in late childhood may predispose developing youth to having visual cues on social media platforms capture internally focused thoughts while self-monitoring and executive control over self-referential processes may exert less influence. These findings complement prior findings that engagement with social media (reading social media posts and viewing personalized video content) may enhance DMN coupling with visual networks and decrease coupling with frontoparietal networks (Hu et al., 2022; Su et al., 2021), raising the possibility that particular youth may be more susceptible prior to substantial exposure to social media. Overall, results from the present study and the larger literature suggest a dual-disruption model of future PUSM severity, characterized by weaker control over self-referential thoughts and stronger sensory-driven input to processes during later childhood, when self-identity and internally directed cognitive systems are still developing and prior to substantial social media exposure. This situation may confer vulnerability in a step-wise or cyclical fashion. For example, an imbalance may manifest as heightened responsivity to self-relevant visual stimuli during childhood, which may elicit ruminative self-judgments, social comparison, and imagined interactions. As children increase their use of social media over time, stimuli such as profile pictures, notifications, and feeds (i.e., personalized, curated visual content) may disproportionately engage cognitive systems in individuals with these pre-existing differences in connectivity patterns (e. g., relatively increased DMN connectivity with the visual area 1 network identified in the present study) and promote sustained engagement or reinforce compulsive checking behaviors on social media platforms (Raichle, 2015; Veissière and Stendel, 2018). These currently speculative possibilities warrant future investigation.

Another key component of the PUSM model was the sensorimotor network, which also emerged as a central component in other CPM-related models of substance use disorders and behavioral addictions, including a cross-sectional study on internet addiction (Antons et al., 2024; Feng et al., 2024). Previous research investigating internet-enabled addictions has provided evidence of enhanced engagement of the sensorimotor network in individuals with internet gaming disorder compared to those without (Dong et al., 2012; Hong et al., 2015; Lee et al., 2021; Park et al., 2017; Wang et al., 2016; Wang et al., 2018; Zheng et al., 2019). Our findings further extend these prior studies by demonstrating that individual-level differences in resting-state functional connectivity involving the sensorimotor network during late childhood can predict subsequent PUSM severity in early adolescence (i.e., four years later), thus suggesting the possibility of individual vulnerability predating substantial engagement. As depicted in our theoretical model, the sensorimotor network was strongly connected to the visual area 1 network, which suggests that pre-existing individual-level differences in the perceptual-motor integration system may reflect a neurodevelopmental vulnerability that precedes the emergence or increased severity of PUSM. As access to and engagement with visually salient, interactive digital environments (via social media) increase during adolescence, individuals characterized by this connectivity profile may be more prone to rapid, stimuli-driven behavioral responses with limited cognitive mediation, such as habitual scrolling, tapping, or swiping. This interpretation is consistent with the I-PACE model of behavioral addictions, which proposes that individual vulnerabilities may lead to subsequent engagement in addictive behaviors that in later stages of addiction may be characterized by seemingly automatic behaviors (Brand, Müller, et al., 2025; Brand et al., 2019; Brand et al., 2016). Given that certain types of social media platforms may preferentially promote automatic motor engagement through visually salient cues, future research should investigate whether neural correlates of PUSM in youth may differ by the primary platform used (e.g., platforms dominated by short-form videos versus those emphasizing static content or text-based interactions).

4.1. Limitations

There are several limitations to consider. The study did not include an external replication sample, which limits the generalizability of findings. Future research should examine the generalizability and replicability of the predictive model for PUSM by using a second independent sample. Also, the potential influence of potential confounds (e. g., commonly co-occurring concerns, such as attention-deficit/hyperactivity disorder, depression, and anxiety) on connectivity was not ruled out (Boer et al., 2020; Keles et al., 2020; Şentürk et al., 2021). Although the CPM framework can, in principle, be applied to predict other domains of psychopathology (e.g., affective, anxiety, and attention-deficit/hyperactivity problems as assessed by the Child Behavioral Checklist), the present study focused specifically on PUSM. Future work should evaluate whether the identified connectivity networks uniquely predict PUSM or generalize across other forms of adolescent psychopathology, which would help distinguish disorder-specific versus transdiagnostic neural signatures. Additionally, as the CPM-identified contribution to subsequent PUSM has a small but statistically significant effect size, future studies may consider other neurodevelopmental, environmental, psychosocial, and genetic factors that may influence the development or increased severity of PUSM among adolescents (Brand, Antons, et al., 2025; Brand, Müller, et al., 2025; Brand et al., 2019). Furthermore, given that this study used p < 0.05 for edge selection, future research could explore the use of different significance thresholds. Finally, the study utilized SMAQ, which does not capture certain dimensions included in some other measures of PUSM, such as lying about usage to others (Van Den Eijnden et al., 2016). However, no single measure has yet been established as the gold standard for assessing PUSM (Bányai et al., 2017; Cataldo et al., 2022). The development of a standardized international measure could enable future validation of the current study findings and refinement of the predictive model.

4.2. Conclusion

We demonstrated that CPM using rs-fMRI data at about age 10 years can successfully identify neural networks predictive of PUSM severity in early adolescence. Several large-scale neural networks were implicated (i.e., default mode, sensorimotor, medial frontal, visual area 1, and frontoparietal networks), suggesting that individual differences in functional connectivity across networks associated with self-referential processing, cognitive control, self-regulation, and perceptual-motor integration during childhood can predict PUSM severity in early adolescence. Although these networks represent potentially promising targets for personalized interventions, future research should work toward replication and extension to translate findings into public health and clinical advances.

Data and code availability

Data analyzed in this study can be found in the ABCD data repository, which are available in the National Institute of Mental Health Data Archive (https://nda.nih.gov/). The CPM code is available at the following website: https://github.com/YaleMRRC/CPM. Data analyses utilized the Bioimage Suite (https://bioimagesuiteweb.github.io/webapp/).