Conceptually Informed AI/ML for South Korean Adolescent Smartphone Overdependency: Low-Risk Screening, Construct Exploration, & Place-Based Policy Implication Profiles

1. Introduction

Smartphone overdependency among adolescents in South Korea has become a critical public health issue, with over 98% of youth owning smartphones and engaging in unprecedented screen time levels [1]. The prevalence rate among Korean adolescents, ranging from 37% to 40%, is the highest among all age groups [2,3]. This raises concerns due to its strong associations with mental health issues, negative emotions, suicidal ideation, and substance use [4,5,6]. Excessive smartphone use is linked to psychological problems such as depression, anxiety, OCD, and ADHD, as well as elevated stress and loneliness [7]. Furthermore, a recent meta-analysis found that smartphone overdependence negatively affects academic achievement among adolescents and young adults [8].

In response, policies have emphasized early identification and intervention, with schools playing a central role in risk assessment [9]. These policies reflect global trends recognizing the impact of smartphones on educational settings. For instance, in France, using a smartphone in class is prohibited for students who are under 15 years old [10]. In the United States, at least 18 states have enacted laws or policies banning or restricting student smartphone use during the school day in K–12 settings [11]. In Korea, schools utilize structured self-report screening tools to classify students by risk level and deliver interventions, including psychological counseling, parental education, and behavioral therapy. These efforts reflect a coordinated public health strategy that integrates educational institutions, mental health services, and family support systems to address youth risk in a comprehensive manner. However, this traditional approach relies heavily on lengthy structured self-report surveys, requiring multiple reporters and asking psychologically intrusive items to determine intervention needs. These embedded survey concerns have been discussed in a large body of research.

First, the reliance on invasive psychological questions that probe deeply into adolescents’ emotional and private experiences can be problematic. Items commonly ask whether students feel hopeless without their phones, experience declines in academic performance, or feel depressed after extended smartphone use [3]. While these items aim to identify risk, their intrusive nature, particularly those involving hopelessness or depressive symptoms, may raise ethical concerns when used at scale in educational settings. Such invasiveness may be particularly distressing for adolescents, who may have a more limited capacity to tolerate the emotional stress associated with self-reflection on sensitive topics compared to adults. As a result, questions that may appear relatively benign to adults can feel psychologically intrusive to adolescents, potentially triggering anxiety, embarrassment, or emotional discomfort. According to the literature, intrusive surveys are associated with increased response bias, reduced accuracy, and, in some cases, participant disengagement [12], particularly among youth concerned about privacy or stigma [13,14]. According to major ethical guidelines such as the Belmont Report, researchers are obligated to respect the dignity and welfare of participants and adhere to the principle of beneficence, which emphasizes “doing no harm, maximizing possible benefits, and minimizing potential risks” [15]. Similarly, Standard 3.04 of the APA Code of Ethics [16] states that psychologists must avoid causing harm to research participants and minimize unavoidable potential harm. Hasking et al. [17] found that 15% of adolescents with poorer psychological functioning than their peers reported feeling upset during survey completion, and experienced a decline in psychosocial functioning afterward. This suggests that particular attention must be paid to the potential for emotional distress among at-risk youth in their smartphone related problems when designing and administering surveys.

Given these concerns, predictive modeling using artificial intelligence and machine learning (AI/ML) has recently gained attention as a promising alternative. AI/ML technologies are already well-established across various domains for enabling precise classification and prediction [18]. In the context of public health and mental health, they are increasingly recognized for their potential to serve as efficient and sensitive screening tools [19]. Unlike traditional approaches that rely on lengthy and often intrusive questionnaires, AI/ML models can identify high-risk individuals with high performance and potentially reduce the number of indicators needed. Kim et al. [20] reported a precision of 87.60% when predicting adolescent smartphone overdependence from full-length surveys across five years of data, while Lee and Kim [21] used a small set of loggable smartphone-related variables yielding high performance when stratifying by age groups between 3 to 69 years old. International studies further validate this approach, showing that models built on psychological and demographic variables can accurately predict smartphone overdependency, offering alternatives to conventional assessments [22,23]. However, to the best of our knowledge, no prior study has developed a low-dimensional (i.e., minimal input) AI/ML model to predict smartphone overdependence specifically among adolescents. Furthermore, no existing models have explicitly utilized conceptually informed feature selection or been designed with the goal of minimizing psychological discomfort and ethical risk in assessment.

A second concern in current assessments is the insufficient attention to contextual factors such as region or urbanicity, despite emerging evidence that this context significantly shapes smartphone usage patterns. For instance, Sapienza et al. [24] explored the impact of urbanicity on smartphone use among approximately 500,000 users across 22 countries. The result shows that urban residents spend more time on their smartphones overall compared to rural users. Specifically, rural users dedicate proportionally more time to entertainment activities like gaming and social media, while urban users favor utility applications such as navigation, business, and productivity tools. The researchers suggest these patterns reflect environmental differences where rural users may turn to smartphones for entertainment due to limited physical recreation options, while urban users employ them as tools to navigate complex city environments. A study in Bangladesh shows that the overall prevalence of smartphone addiction is higher among preschoolers in urban areas than rural ones. One of the characteristics of urbanicity, high population density, has been shown to influence smartphone addiction among Koreans [25]. On the other hand, research in China found that adolescents in rural areas showed higher levels of smartphone addiction compared to their urban counterparts, driven by heightened feelings of loneliness and anxiety when not using smartphones [26,27]. These differences suggest that the relationship between adolescents’ smartphone use and risk for overdependency may vary depending on urbanicity and regional contexts. Nevertheless, to our knowledge, no AI/ML-based models in smartphone overdependency are developed or evaluated with sufficient attention to urbanicity and regional contexts. This limits their ability to deliver place-sensitive predictions and policy implications.

Given these gaps, researchers have called for a shift toward predictive models that are not only accurate but are also ethically informed and sensitive to local variation. Such models should incorporate minimal yet high-yield features, reducing respondent burden while maintaining predictive power. Specifically, these predictors should strategically combine indicators and non-intrusive data to enhance predictive efficiency and ethical soundness. In addition, conceptually chosen predictors are crucial to not only improve interpretability but also ensure generalizability across different populations [28]. Additionally, carefully selected predictors reduce computational costs and data collection burdens, facilitating more sustainable and accessible implementations of predictive tools in real-world settings [29]. By combining these techniques with place-stratified eXplainability [30,31] evaluations, AI/ML-based models can effectively address public health imperatives to deliver tailored interventions while minimizing risk and burden.

1.1. Research Aims & Contributions

This study aims to advance the understanding, detection, and policy implications of adolescent smartphone overdependence in South Korea through the integration of conceptual frameworks and AI/ML methods. Grounded in public health and social sciences, we outline three aims:

1.

Develop a high-performing, low-risk screening tool to detect adolescents at risk for smartphone overdependence using conceptually informed AI/ML modeling;

2.

Leverage AI/ML toward construct exploration of contributing factors; and

3.

Provide place-based policy profiles that offer insights into local variation in risk factors and model performance, enabling targeted intervention.

The study contributes methodologically by introducing a framework that prioritizes conceptually meaningful constructs over purely data-driven feature sets. Practically, it informs the design of adolescent-friendly screeners aligned with ethical research principles. Conceptually, it shows how AI/ML can illuminate patterns in adolescent behavior that align with psychologically meaningful constructs, offering new directions for behavioral research. From a policy perspective, it advances place-sensitive approaches to digital well-being, providing insights that can support regionally adaptive public health interventions.

2. Materials and Methods

The following sections outline our data, approach to feature selection and grouping, and analyses.

2.1. Data

Secondary data, collected by the National Information Society Agency (NISA) in South Korea, was used. The full sample consisted of 22,844 household members who had used a smartphone within the past month and were present and available for the interview, which took place from September to November 2023. During the interview, each participant was asked approximately 192 questions covering demographic characteristics, smartphone usage, online video services, counseling and education, awareness of overdependency, and psychosocial characteristics. In the current study, we focused on a total of 1,873 adolescent participants aged 10-18 from the full sample.

2.2. Outcome

The outcome variable in this study is smartphone overdependence. NISA measured this by administering the validated Smartphone Overdependence Scale [32]. This scale has ten questions measuring three facets: 1) the level of difficulty in controlling smartphone use, 2) concentration disruption due to smartphone use, and 3) the incidence of health, academic, or interpersonal conflicts caused by smartphone use. The scale uses a 4-point Likert ranging from Strongly disagree (1) to Strongly agree (4). Based on total scores, adolescents are categorized as high-risk (≥31 points), potential risk (23–30 points), or no risk (<23 points). In our AI/ML models, we use the binary any-risk vs no-risk target as our label by combining the potential risk and high-risk groups. This is in line with our research aim to develop a low-risk screener for public health detection.

2.3. Conceptual Feature Selection & Construct Grouping

From the NISA survey, all individual items related to adolescents were considered candidates for inclusion. While the original NISA survey includes five broad domains (smartphone use, online video use, prevention and counseling, problem perception, and psychosocial characteristics), we regrouped items based on theoretical grounding. Two overarching domains (demographics and urbanicity) and ten theoretically grounded constructs were identified. Only items aligning with one of these domains or constructs were retained for analysis. This resulted in a total of 92 final items. After one-hot encoding, these 92 study items become 132 features.1 In Appendix A, Table A1 and Table A2 outline the 2 domains and 10 constructs, and their respective 132 features. Below, we define each construct and summarize its relevance to adolescent smartphone overdependence.

1.

Intent to Seek Smartphone Education. This construct captures adolescents’ intention to engage in preventive education related to smartphone use [33]. Higher intention to seek education has been associated with reduced risk of overdependence.

2.

Preventive Education Experiences. This reflects actual participation in educational or intervention programs designed to prevent smartphone addiction [34]. Although evidence for their effectiveness remains mixed, adolescents who have received such education often demonstrate greater awareness and are more likely to engage in help-seeking behaviors.

3.

Smartphone Use Cases. This measures the frequency of engagement with different smartphone functions (e.g., gaming, video streaming, navigation) [35]. Entertainment and social networking applications, in particular, have been linked to increased overdependence due to their reward structures and habit-forming properties [36].

4.

Home Environment. The home environment includes both physical and psychological aspects of the household context [37]. Prior studies have shown that conditions such as socioeconomic status, family structure, and parental support can shape adolescent behavior and potentially influence smartphone use [38,39].

5.

Parental Prevention Efforts. Parental prevention efforts refer to the strategies parents use to manage and guide their children’s media use, including rule-setting, supervision, and open dialogue [40]. Active parental prevention efforts have been found to reduce risk for smartphone addiction, while overly restrictive or absent parenting may increase it [41,42].

6.

Social Support. Social support refers to the perception that one is cared for, valued, and part of a network of mutual obligations [43]. Strong social support has been associated with reduced smartphone overdependence, potentially by mitigating negative emotions and enhancing self-regulation [44].

7.

Self-Regulation Assessments. Self-regulation is the ability to control emotions, behaviors, and cognition in pursuit of long-term goals [45]. A growing body of research suggests that higher self-regulation is associated with lower levels of smartphone overdependence in adolescents [46].

8.

Perceived Digital Competence & Risk. This construct refers to individuals’ perceived abilities to effectively navigate digital environments (e.g., digital content creation, privacy awareness) as well as their perceived smartphone issues (e.g., excessive use, difficulty controlling short-form video consumption). These perceptions may shape how individuals engage with digital technology and manage potential overdependence [47].

9.

Life Satisfaction. Life satisfaction refers to a person’s evaluation of their overall well-being and quality of life [48]. Studies have shown that higher levels of smartphone addiction are correlated with lower life satisfaction among adolescents [49].

10.

Smartphone Consequences & Dependence. Psychological distress is characterized by emotional suffering, typically involving symptoms of depression and anxiety [50]. Excessive smartphone use has been linked to increased psychological distress and related cognitive and neurological effects [7].

2.4. Artificial Intelligence/Machine Learning Models

This section describes our conceptual framework and development of AI/ML models, and then presents our evaluation methods. All analyses were conducted using Python version 3.10.13, with commonly used data science and machine learning libraries (e.g., scikit-learn, pandas, XGBoost, LightGBM). The codebase can be made available upon reasonable request to the corresponding author.

2.4.1. Conceptual Framework

We build various AI/ML models using two primary approaches: 1) nested modeling and 2) construct-based modeling. In the nested modeling, we begin with simple demographics and add sets of features one construct at a time. The sequence of nesting is based on the risk for psychological discomfort. In other words, we consider the ethical considerations with implementation-informed design. This conceptual approach was chosen to develop the low-risk screener. This would also identify the screener with the minimum features needed for high prediction accuracy. Next, we take a construct-based modeling approach to identify which construct feature sets have the most predictive power and examine the feature importance within. This provides a conceptual approach toward the construct exploration of contributing factors leveraging AI/ML.

2.4.2. AI/ML Classification Modeling

To examine smartphone overdependency among adolescents, we developed classification models to predict our binary outcome of any risk for adolescent smartphone overdependency. Here we discuss our data processing and AI/ML procedures.

Data Processing

Our dataset comprises of numerical and categorical variables. First, we one-hot encode the categorical variables in our dataset. To standardize our data, we then train a scalar on our training set using the StandardScalar function from the scikit-learn library [51]. This resulting scalar is then applied to the test set to prevent data leakage.

AI/ML Procedures

We applied widely used high-performing algorithms to maximize classification performance. They include logistic regression (LogReg), decision trees (DT), random forest (RF), extreme gradient boosting (XGBoost), multi-layer perceptron (MLP), and light gradient-boosting machine (LightGBM) [52,53,54,55,56]. We then split the data into an 80/20 training and out-of-sample test sets. The training set had 1,498 adolescents, and the test set had 375 adolescents, which were held constant across algorithms and all experiments. We use 5-fold cross-validation, nested within the training set, for hyperparameter tuning and out-of-sample testing for final performance evaluation. We tune hyperparameters specific to each algorithm using Optuna (an automatic hyperparameter optimization framework) with 50 iterations to minimize mean logarithmic loss across the 5-fold cross-validation (mean val loss) [57]. For MLP and gradient boosting algorithms, we utilize early stopping to prevent overfitting. We report the search spaces and hyperparameters for our reported models in Appendix B. In the final testing, model performance was assessed using three metrics: mean val loss, test loss, and the out-of-sample area under the receiver operating characteristic curve (AUC). The consistency between mean val loss and test loss was used to assess model generalizability, while the out-of-sample AUC was employed to evaluate the model’s ability to discriminate between classes on unseen data [58,59].

2.4.3. eXplainable AI (XAI) Techniques

We apply XAI techniques to improve the interpretability of the model by identifying features that meaningfully contribute to its predictions [30,31]. This ensures transparency and interpretability of our classification models for predicting adolescent smartphone overdependency. We use two XAI tools, Shapley Additive exPlanations (SHAP) and Partial Dependence Plots (PDPs) [60,61].

SHAP assess both global (mean feature impact) and local (instance-specific) contributions of each predictor, clarifying how features influence predictions [61]. Additionally, we employed PDPs to analyze key nonlinearities, inflection points, and interaction thresholds in predictor-outcome relationships [62]. Section 3 presents SHAP beeswarm plots and PDPs to illustrate SHAP value distributions and the functional effects of high-impact predictors. We compute SHAP values to quantify and examine the rank and directionality of features, providing insights into how predictors influence predictions. We generate PDPs to visualize global marginal effects on the average prediction, providing insights into nonlinear relationships and slope inflection points.

2.4.4. Place-Stratified Predictive Analysis

We conduct a place-stratified predictive analysis by evaluating the performance of our AI/ML models on sample subsets stratified by urbanicity: 1) Metropolitan City, 2) Medium or Small City, and 3) Town or Rural District. Group comparisons found 31 significant differences by urbanicity out of 128 one-hot encoded features (Table 1). This may lead to nationally trained models missing key place-based nuances or inconsistent performance. Using our out-of-sample test set and stratifying the predictions by urbanicity, we recalculate the AUCs, SHAP, and PDPs only for those sample subsets and examine differences in feature importance and trends (see Table A4-Table A6 in Appendix C for group comparisons across all study variables). We use these insights to provide place-based policy profiles in Section 4.1.

3. Results

The following sections present the results the sample descriptive results, an overview of AI/ML results from each modeling approach (nested and construct-based), and further explore four exemplar models.

3.1. Sample Descriptive

A total of 1,873 adolescents aged 18 years or younger were included in the analysis, with 1,498 in the training set and 375 in the test set. No significant group differences between the training and test sets were observed in any of the study variables (see Appendix Table A9-Table A13 for detailed analysis).

Within the total sample, the mean age was 14.34 years (SD = 2.40), and 51.9% of participants were female. 44.4% of participants residing in Metropolitan Cities, 38.1% in Medium/Small Cities, and 17.5% in Town/Rural Districts. A majority (71.8%) reported that both parents were working, while 7.8% lived in single-parent households and 0.7% identified as being from multicultural families i.e., having at least one foreign-born parent. About 21.8% lived in single-family homes, while the remainder resided in apartments or other multi-unit dwellings. The average monthly household income was 4.25 million KRW (SD = 1.12; approximately 3,100 USD).

3.2. Overview of AI/ML Results

The following sections present the AI/ML results from the nested and construct-based modeling frameworks.

Performance Comparison for Nested AI/ML Classifiers

In the nested framework, 12 nested AI/ML classifiers were trained. Overall, the results show generalizability across training and test sets with consistent mean val loss and test loss of each nested classifier. The baseline model (M0 [Demographics]) performed at chance (AUC = 0.51), and M1 (+Urbanicity) modestly improved performance (AUC = 0.57). Adding a single intent to seek smartphone education item in M2 increased the AUC to 0.63, but further smartphone prevention education-related features in M3 did not yield gains. A substantial improvement emerged in M4 (+Smartphone Use Cases), which added behavioral features related to smartphone use. This model achieved an AUC of 0.81, an 18-point increase, indicating it as a “good test" [58]. We identify the M4 (+Smartphone Use Cases) as our low-risk screener as it offers high performance while using low-burden, non-intrusive features and requires no psychological assessments.

Subsequent models added features related to social context (M5–M7), self-regulation (M8), and perceived digital competence & risk (M9), leading to incremental improvements. The highest performance was observed in M11 (+Consequences & Dependence), which reached an AUC of 0.92 and qualifies as an “excellent test" [58]. However, the inclusion of psychologically sensitive content and longer assessments makes later models less suitable for early-stage or population-level screening for public health surveillance goals. The strength of the M4 (+Smartphone Use Cases) model lies in its balance between predictive power and feasibility (see Table A14 in Appendix D for more details).

Performance Comparison Between Construct-Based AI/ML Classifiers

In our construct-based modeling framework, we train 12 individual, construct-specific AI/ML classifiers. Overall, we find stable performances between the mean val loss and the test loss, indicating generalizability. Three constructs – Smartphone Consequences & Dependence (Smart CD), Perceived Digital Competence & Risk (Perceived DCR), and Smartphone Use Cases (Use Case) – achieve high performances, indicating a “good test” while all others indicate “poor tests” [58]. Notably, we find that the Parental Prevention Efforts and constructs related to smartphone education result in “poor tests” despite qualitatively being related to smartphone overdependence.

3.3. Four Exemplar Models

We present four exemplar models: 1) the M4 (+Smartphone Use Cases) nested model, 2) the Smartphone Consequences & Dependence (Smart CD) construct-based model, 3) the Perceived Digital Competence & Risk (Perceived DCR) construct-based model, and 4) the Smartphone Use Cases (Use Case) construct-based model. For each model, we present the mean val loss, test loss, and AUC (see Table 2 and Figure 1). Next, using XAI techniques, we examine the feature importance across models (see Figure 2 and PDP plots below). We then present the place-stratified results regarding each model’s performance (see Table 3) and relative SHAP feature importance (see Figure 3).

3.3.1. M4 (+Smartphone Use Cases) Nested Model

Of the 12 nested classifiers, the M4 (+Smartphone Use Cases) model is the best candidate for a low-risk screener with detection and public health surveillance goals in mind. This model is parsimonious, using 59/131 features, and balances the tradeoff between high-performance (AUC=0.81; Figure 1) and minimizing psychological discomfort. The majority of the predictive power comes from the adolescent’s behavioral use cases of their smartphone requiring no psychological reflection, assessment, or discomfort. These features added a substantial improvement of 18 percentage points, going from an AUC of 0.63 in M3 to 0.81 in M4, achieving a “good test” designation [58]. In addition, this model shows good generalizability across training and test sets with stable mean val loss (0.58) and test loss (0.55) scores.

Feature Importance

The SHAP beeswarm plot for M4 (+Smartphone Use Cases) is presented in Figure 2, subplot (a). Overall, higher endorsement of these features predicted a higher likelihood of being at risk for smartphone overdependency. Some notable nuances are that "Edu Intent" (3-point Likert) was the most important and appears to have a thresholding effect, where only high endorsement (red) led to a higher likelihood of smartphone overdependency. In the PDP, this results in only the endorsement of 3 being associated with a likelihood of overdependency (Figure 4, subplot A.). In contrast, the "Gaming" (8-point Likert) and "Short Form Video Usage Portion" (4-point Likert) features indicate a clear positive relationship where higher endorsement (red) is associated with a higher likelihood of overdependency. The PDPs for both features show a linear trend (Figure 4, subplot B. and subplot C.). The PDPs also revealed a notable nonlinear trend in the marginal effect of "Phone Use Case - Navigation" (7-point Likert), where extreme low (0-1) and high (4-6) levels of endorsement led to a higher likelihood of overdependency while only medium (2-3) levels of endorsement led to a lower likelihood (Figure 4, subplot D.). This may be due to the integral use of smartphones for navigation, where not using it at all is out of the norm and may indicate overuse of other use cases, while overuse would increase reliance and dependency on one’s smartphone (See Appendix E for all other PDPs).

Place-Stratified Results

Place-stratified analyses revealed that Metropolitan City adolescents have the highest performance, with an AUC of 0.87 (compared to 0.81 in the National performance; Table 3). In contrast, the performance among Medium/Small City adolescents is the lowest, with an AUC of 0.75, which would only indicate a “fair test” [58]. Among Town/Rural District adolescents, the discriminative power remains about the same, with an AUC of 0.80.

Placed-Stratified Feature Importance

Place-stratified analysis of SHAP feature importance found that across all strata, "Edu Intent - Intent to Participate in Prevention Program", "Phone Use Case - Gaming", and "Phone Use Case Video - Short Form consistently ranked as the top three predictors of smartphone overdependency. "Phone Use Case - Navigation" ranked moderately across all regions, appearing between 7th and 8th in all strata, with no major fluctuations in relative importance. Additionally, “Phone Use Case-Online Learning” ranked higher in Metropolitan Cities (7th) compared to the National and Town/Rural District strata (both 9th), and was the lowest in Medium/Small Cities (10th). In Town/Rural Districts, "Phone Use Case - Webtoon" was ranked 4th—one position higher than in other regions—exceeding "Phone Use Case - Selling Goods and Services”, which ranked 5th. Overall, the place-stratified results show strong consistency in the top-ranked predictors, with minor differences in mid- and lower-ranked features by regional context.

3.3.2. Smartphone Consequences & Dependence Classifier Results

The Smart CD model had the highest performance across all construct-based models with an AUC of 0.89. However, this construct asks adolescents to do the highest degree of intrusive reflection specific to their potential problems and dependency issues related to their smartphone use. Therefore, it is not surprising that it is the most predictive, but it also carries the largest potential for psychological discomfort.

Feature Importance

The SHAP beeswarm plot for the Smart CD model is presented in Figure 2, subplot (b). We find that consequences are more predictive than the dependence on smartphones for "Stress Relief". PDPs revealed three patterns of marginal effects: 1) linear, step-wise increase, 2) thresholding effect, and 3) binary effect. "Social Isolation" (3-point Likert) followed the linear, step-wise increase. "Physical Effects" (4-point Likert) and "Depression" (4-point Likert) followed the thresholding effect, where there was an initial increase in predicted probability of overdependence that later flattens out. "Stress Relief" (4-point Likert), "Low Battery Nervousness" (4-point Likert), "Sleep Issues" (4-point Likert), and "Periodic Check Nervousness" (4-point Likert) followed the binary effect. Figure 5 shows examples of each pattern (See Appendix F for all other PDPs).

Place-Stratified Results

The Smart CD model performance between the National and stratified by urbanicity samples are presented in Table 3. The results show that the performance among Metropolitan City adolescents is the highest with an AUC of 0.91, indicating an “excellent test” [58]. Both the performances among Medium/Small City and Town/Rural District adolescents are the same, with an AUC of 0.87, performing slightly worse than the National model but still in the “good test” range [58].

Placed-Stratified Feature Importance

Place-stratified analysis of SHAP feature importance found that across all strata—National, Metropolitan City, Medium/Small City, and Town/Rural District— "Low Battery Nervousness", "Physical Effects", and "Periodic Check Nervousness" consistently appeared as the top contributors to the predicted probability of smartphone overdependence. In the Metropolitan City stratum, "Physical Effects" was the most important feature, followed by "Low Battery Nervousness" and "Periodic Check Nervousness". This ordering differs from the overall National model, where "Physical Effects" ranked slightly lower. In the Medium/Small City stratum, the top-ranked features mirrored those of the National sample, with "Low Battery Nervousness" and "Physical Effects" leading in importance. In the Town/Rural District stratum, "Low Battery Nervousness" and "Physical Effects" again ranked highest. Notably, “Sleep Issues” appeared slightly more prominently in this group than in the others. Across all regions, all Smart CD features contributed positively to predicted overdependence when endorsed at higher levels. Feature rank order varied only slightly between strata, indicating overall stability in the predictor structure across places.

3.3.3. Perceived Digital Competence & Risk Classifier Results

Next, the Perceived DCR model achieves an AUC of 0.84 and asks questions regarding one’s self-perceptions rather than experienced harms.

Feature Importance

SHAP analysis revealed mixed results where higher endorsement (red) of perceived phone issues was often associated with a higher likelihood of overdependency and higher endorsement of perceived ability was often protective (Figure 2, subplot (c)). Several deviant cases are discussed. "Perceived Phone Issues - Short Form Video Usage Control Difficulty" (4-point Likert) revealed that moderate endorsement was more protective than low endorsement. This may reflect instances of denial in use control difficulties. "Perceived Ability - Assess Online Information Reliability" and "Perceived Ability - Online Privacy Awareness" were not protective, and higher endorsement (red) predicted higher likelihoods for overdependency, perhaps reflecting illusions of control and cognitive distortions. "Perceived Ability - Online Social Issue Engagement" and "Perceived Ability - Digital Content Creation" were also not protective and may reflect a higher reliance on smartphones for socialization and creativity. Three notable parabolic or higher-order trends emerged for the Perceived DCR model PDPs in their marginal effects: 1) a positive to negative curve, 2) a negative to positive curve, and 3) a positive to negative to positive curve, revealing complex cognitive patterns in smartphone overdependency. Figure 6 shows examples of each pattern (See Appendix G for all other PDPs).

Place-Stratified Results

The performance of the Perceived DCR model stratified by urbanicity is presented in Table 3. Across all strata, the performance remains in the “good test” range. Among Metropolitan City adolescents, the AUC is 0.85, while among both the Medium/Small City and Town/Rural District adolescents, the AUC is 0.83.

Placed-Stratified Feature Importance

Across all strata, three perceived phone issues consistently ranked highest: “Smartphone Overdependence”, “Excessive Smartphone Using Time”, and "Short Form Video Usage Control Difficulty". These self-perceived issues demonstrated stable predictive strength across Metropolitan and Medium/Small cities. In contrast, the Town/Rural District stratum showed slight reordering in feature importance. While “Smartphone Overdependence” and “Excessive Smartphone Using Time” remained top predictors, "Negative side effect on gaming" ranked higher than"Short Form Video Usage Control Difficulty", and "Perceived Ability - Online Social Issue Engagement" appeared more prominently. Specifically, "Online Social Issue Engagement" ranked 5th in the Rural/Town stratum (vs. 7th in others), and "Negative Side Effects on Messenger" ranked 6th (vs. 10th in others). Overall, while the top self-perceived dependence and usage difficulty features remained stable across regions, the Town/Rural District model showed relatively greater importance for social and emotional factors—such as issue engagement and messaging-related effects—while cognitive-evaluative skills and digital production activities showed relatively reduced predictive influence.

3.3.4. Smartphone Use Cases Results

The final “good test” classifier is the Use Case model with an AUC of 0.80, which only uses behavioral tracking related indicators. The Use Case model suggests potential in leveraging raw behavioral tracking data directly from smartphones.

Feature Importance

SHAP analysis revealed nuances in which types of smartphone use cases are predictive of overdependency (Figure 2, subplot (d)). 13 of the top 20 important use cases had a predominantly positive relationship, where higher endorsement (red) was associated with overdependency. The three most important features of this pattern were "Gaming", "Short Form Video Usage Portion", and "Webtoon". Four of the top 20 features had a predominantly negative relationship, where lower endorsement (blue) was associated with overdependency. The three most important features of this pattern were "Navigation", "News", and "Health". This may reflect a bifurcation where smartphone use cases for entertainment are more likely to predict overdependency, while use cases that capture a non-entertainment utility value are more protective. The Use Case PDPs generally supported the SHAP findings with minor nuances or thresholding effects across Likert responses, but still held overall positive or negative trends. Notably, the PDP of the "Navigation" (7-point Likert) feature was consistent with the nested M4 (+Smartphone Use Cases) model, where the same trend and cutoffs were found. Extreme low (0-1) and high (4-6) levels of endorsement led to a higher likelihood of overdependency while only medium (2-3) levels of endorsement led to a lower likelihood. This consistency across models supports the robustness of this feature’s relationship. Figure 7 shows examples and presents the "Navigation" PDPs (See Appendix H for all other PDPs).

Place-Stratified Results

The performance of the Use Case model stratified by urbanicity is presented in Table 3. These results show that among Metropolitan City adolescents, there is a substantial jump in performance with an increase of 8 percentage points for an AUC of 0.88. However, the performance substantially drops for the Medium/Small City adolescents with a decrease of 10 percentage points for an AUC of 0.70 falling into the “fair test” range [58]. The performance among Town/Rural Districts decreases marginally to 0.79 but falls from a “good test” to “fair test” [58].

Placed-Stratified Feature Importance

Across all strata, four features consistently emerged as top predictors: "Gaming", "Short-Form Video Usage Portion", "Webtoon", and "Navigation". "Gaming", "Short-Form Video Usage Portion", and "Webtoon" consistently showed strong positive contributions to predicted risk. "Navigation" was the only feature with a mixed effect, consistently ranked in the top but associated with slightly lower predicted risk at moderate levels of use. In the Metropolitan City stratum, "Gaming" ranked first, followed by "Navigation", "Webtoon", and "Short-Form Video Usage Portion". Phone use for "News" showed negative contributions to predictions in this stratum. Overall, behavioral features with immersive or compulsive qualities dominated. In the Medium/Small City stratum, "Navigation" and "Short-Form Video Usage Portion" were the top predictors, followed by "Gaming" and "Webtoon". Notably, "Listening to Music" ranked 10th only in this region, indicating that more passive forms of phone use may be more salient here. In the Town/Rural District stratum, the top four features mirrored those in the National sample. However, there was greater diversity in the top 10 features, including "Photography" (ranked 9th), which did not appear in other regions. This suggests that expressive and leisure-oriented smartphone behaviors play a more prominent role in rural contexts. Features like "Selling Goods & Services", "Online Learning", "Messenger", "Online Shopping", and "Scheduling" were present across all strata with moderate contributions, while "Listening to Music" appeared in all regions except Metropolitan cities, ranking the lowest in predictive strength but showing regional relevance.

4. Discussion

This study is the first to develop a high-performing, low-risk, AI/ML-based screener for adolescent smartphone overdependence in South Korea using conceptually informed modeling and a place-stratified analysis approach. By combining ethical considerations with conceptually-informed design, we identified a parsimonious screening model that balances predictive power with psychological safety. The M4 (+Smartphone Use Cases) model, leveraging only behavioral use data and demographic characteristics, achieved an AUC of 0.81 nationally, and 0.87 in Metropolitan Cities, comparable to more invasive models using psychological assessments. This confirms our working hypothesis that a minimal set of non-invasive indicators can yield strong predictive power.

These results align with prior literature emphasizing the importance of smartphone use patterns, particularly entertainment-related activities, in predicting overdependence [20,63]. Our XAI analyses provided further insight, showing that use cases like gaming and short-form video were strong predictors of risk, while utilitarian use cases such as navigation exhibited a nonlinear pattern, where both low and high engagement signaled higher risk. This thresholding and parabolic trend in use mirrors earlier findings in digital addiction and behavioral tracking, suggesting that moderate use may reflect healthy regulation, whereas extremes indicate dependency [24].

In parallel, our construct-based models expand on the potential of AI/ML conceptual exploration. Our models confirmed the theoretical relevance of both perception and consequence-based constructs. The Smartphone Consequences & Dependence model achieved the highest performance (AUC = 0.89), but its predictive strength likely reflects conceptual redundancy with the overdependence outcome itself. Many items in this construct directly assess perceived harm and experienced behavioral symptoms that closely mirror the criteria used to define smartphone overdependence, limiting their theoretical distinctiveness despite high classification accuracy.

In contrast, the Perceived Digital Competence & Risk model (AUC = 0.84) provides more meaningful theoretical insight. It aligns closely with cognitive-behavioral frameworks by examining how adolescents evaluate and perceive their own ability and issues with smartphone use. Features related to perceived control, awareness of usage patterns, and beliefs about digital competencies correspond to key mechanisms in models of behavioral addiction and cognitive distortions. The nonlinear and parabolic trends observed in this model’s PDPs suggest that adolescents’ cognitive self-perceptions are complex, potentially involving both denial and overconfidence—patterns well-documented in cognitive-behavioral theory [64]. Notably, some indicators, such as overconfidence in digital literacy, were associated with increased risk, perhaps reflecting distorted cognitions or rationalizations, which merit further psychological investigation. In the following section, we present place-based policy implication profiles based on our place-stratified analysis.

4.1. Place-Based Policy Implication Profiles

We identify tangible implications that can guide place-based policy responses across regional contexts for the prevention and intervention of smartphone overdependence among adolescents.

Metropolitan Cities

In metropolitan areas, adolescents at risk for smartphone overdependency exhibit significant engagement in immersive digital entertainment activities such as gaming and short-form video consumption. These activities frequently coincide with comparatively higher perceived digital competency and varied smartphone uses. Policy approaches in urban contexts should emphasize enhancing self-assessment of abilities and balancing screen engagement with offline activities rather than implementing broad restrictions. School-based initiatives can effectively incorporate media literacy components focused on self-monitoring techniques and healthy smartphone usage discussions among peers. Additionally, considering the high association between anxiety-driven behaviors and smartphone overdependence, integrating targeted mental health services within school systems to directly address compulsive smartphone use behaviors is recommended.

Medium & Small Cities

Adolescents residing in medium and small cities who are at risk of overdependency exhibit diverse smartphone use patterns, including practical applications such as online shopping, scheduling, and educational platforms. Nevertheless, compulsive use of smartphones, notably for stress relief, remains a relevant factor. Policy and interventions in these contexts should deliver contextually tailored educational programs that recognize and address entertainment-driven risks and the normalization of smartphone-based emotional coping strategies. Offering adolescents practical alternatives to smartphone-based coping by enhancing emotional literacy, digital citizenship education, and providing habit-tracking resources could effectively mitigate the risks associated with overdependence.

Towns & Rural Districts

In towns and rural areas, at-risk adolescents predominantly use smartphones for expressive and leisure-related activities such as video streaming and engagement in online social interactions. These patterns may be related to limited offline social opportunities in rural settings. Policy measures for rural adolescents should thus extend beyond limitations on screen time to address underlying structural needs. These measures could include expanding access to offline youth facilities, promoting community-driven activities, and fostering healthy online interaction guidelines. Additionally, considering the moderate predictive value of smartphone use for stress coping, interventions should seek input from adolescents regarding their daily smartphone usage patterns, thereby informing the creation of meaningful and accessible offline alternatives.

4.2. Limitations & Future Research

This study has several limitations. First, all data were derived from self-report surveys, including behavioral indicators. Future research should explore the integration of passive smartphone tracking data, which may offer higher ecological validity and remove social desirability bias. In addition, Ecological Momentary Assessment methods can improve these studies. Second, while we stratified models by urbanicity, other contextual variables—such as school type or family environment—could further refine prediction. Third, while our models were interpretable using SHAP and PDP, further work using longitudinal data is needed to establish causal pathways.

Future studies should also investigate protective features and resilience factors, particularly in adolescents who report high usage but do not develop dependency. Additionally, given the complex, nonlinear relationships found in perceived digital competence & risk indicators, qualitative or mixed-methods research may help unpack underlying cognitive distortions of smartphone use.

5. Conclusions

This study’s contributions are both methodological and practical. Methodologically, this study demonstrates the utility of conceptually informed construct modeling over “kitchen sink” approaches. Psychologically salient constructs-particularly those relating to use patterns, perceived digital competence & risk, and consequences—outperform broader demographic or contextual predictors. Practically, our work supports the adoption of low-burden, adolescent-friendly screeners in school settings, aligning with ethical frameworks (e.g., the Belmont Report; APA Code of Ethics) that call for minimal harm in youth research. Conceptually, this study illustrates the use of AI/ML to uncover potential theoretical phenomena. In our case, cognitive patterns point to potential cognitive distortions. Policy-wise, our work contributes to the growing field of place-based AI/ML by demonstrating that predictive accuracy and feature importance vary significantly by urbanicity. Our results suggest that risk for overdependence is not uniform across regions. Local social, technological, and cultural contexts shape which factors matter most. These findings support the development of regionally tailored public health messaging and digital wellness interventions. Adolescents from medium or small cities consistently showed lower model performance across classifiers, suggesting that these communities may exhibit more heterogeneous or undercaptured patterns of smartphone use. This nuance emphasizes the need for locally adaptive tools and supports critiques that national models risk obscuring regional variation [65].

Appendix C.1. Place-Stratified Descriptive Comparisons

Table A4-Table A6 present the place-stratified descriptive comparisons. One-way ANOVAs are used for continuous and ordinal variables while Chi-square tests are used for categorical ones.

Appendix D.1. Performance Comparison for Nested AI/ML Classifiers

We present the model performance of the 12 nested AI/ML classifiers in Table A14, which presents the AUCs for several high-performing models.

Appendix D.2. Performance Comparison Between Construct-Based AI/ML Classifiers

We present the model performances of the 12 construct-based AI/ML classifiers in Table A15.