Analyzing adolescent mental health: Correlates of depression and anxiety through big data analytics

Abstract

Objective

Conventional scale-based diagnostic approaches are increasingly insufficient for addressing the growing mental health challenges among adolescents. Leveraging advances in artificial intelligence, this study aims to develop an accurate, efficient, and scalable model for early identification of adolescent depression risk using large-scale census data, and to identify key daily life factors associated with mental health outcomes.

Methods

Data were obtained from the 2021 National Survey of Children's Health, including 50,892 adolescents and 463 variables. Based on prior literature, 60 relevant variables were selected. Three progressively structured hypotheses concerning the relationships between adolescent depression and developmental environments were proposed. Machine learning models, including decision trees, XGBoost, support vector machines, and neural networks, were applied to predict depression risk. Mediation analysis was conducted to examine the pathways through which living conditions influence mental health.

Results

The optimal model demonstrated strong predictive performance, achieving an accuracy of 0.85 and an AUC exceeding 0.87. Feature importance analysis identified several key predictors. Mediation analysis indicated that living conditions exerted a direct effect of 0.225 on mental health, while physical activity and diet quality partially mediated this relationship.

Conclusion

Living conditions are critical indicators for early identification of adolescent depression risk. The use of nationwide census data enables timely screening and targeted intervention. Improving dietary habits and increasing physical activity may serve as effective preventive strategies for adolescent mental health disorders.

Keywords

Adolescent depression census data disease prediction mental health big data

Introduction

In today's fast-paced society, mental health disorders have emerged as a critical public health issue. Adolescents are particularly affected, experiencing elevated levels of emotional distress, with prevalence rates ranging from 15% to 32% and showing a persistent upward trend.¹ This trend underscores the urgency of addressing psychological challenges in this demographic. Adolescents’ psychological and physiological immaturity, combined with adverse environmental factors such as parental neglect, lack of emotional support, tense family dynamics, harsh parenting styles, and strained relationships with peers, can significantly contribute to behavioral problems.^2–4 Addressing these underlying causes through early identification and targeted interventions is therefore essential not only for promoting adolescents’ holistic development and well-being but also for advancing broader societal health.⁵ This remains a critical global challenge.

In current clinical practice, mental disorders are typically diagnosed using inquiry-based scales guided by the International Classification of Diseases, 10th Revision (ICD-10), which outlines criteria for depressive episodes and recurrent depressive disorders.⁶ For diagnosing depression, commonly used tools include Zung's 1965 Self-Rating Depression Scale (SDS), which has demonstrated strong reliability and validity.⁷ In addition to the SDS, the Mini-International Neuropsychiatric Interview (MINI) is another prominent instrument, designed as a brief structured interview for diagnosing mental disorders.⁸ This tool was developed collaboratively by American and European researchers and aligns with both the Diagnostic and Statistical Manual of Mental Disorders (DSM-IV) and the ICD-10.⁹ Furthermore, the Symptom Checklist-90 (SCL-90), created by Derogatis, consists of 90 items that evaluate mental health based on self-reported symptoms.¹⁰ Although standardized scales offer a valuable basis for clinical assessment, most originate from specific cultural and regional contexts; even after translation and adaptation, they often exhibit limited applicability—especially among younger adolescents, for whom developmental and linguistic differences can compromise their validity and usability.¹¹ Prediction frameworks that utilize statistical models, regression analysis, support vector machines (SVMs),¹² neural networks, and Bayesian networks have demonstrated effective performance in mental health contexts.^13–16 Leveraging advances in diagnostic technology, researchers have adapted risk prediction models from general medicine to schizophrenia, with the goal of identifying individuals at high risk for psychosis and enhancing early intervention efforts.¹⁷ Early ultra-high risk models provided a new strategy for the early identification and intervention of mental disorders.¹⁸ But they have now been proven to have drawbacks such as a high rate of false positives and limited predictive ability.¹⁹ In addition to the aforementioned statistical approaches,²⁰ Hu et al. explored the association between well-being and depression among college students from both genetic and childhood abuse perspectives, providing a transition from objective biological factors to subjective environmental influences in understanding adolescent depression and its relationship with developmental contexts. Furthermore, prediction models for adolescent depression can be constructed by evaluating psychosocial factors and measuring biological markers, such as plasma cortisol and dehydroepiandrosterone sulfate levels, which provide valuable insights into the mental health of adolescents.²¹

However, reliance on these tools alone may lead to diagnostic inaccuracies, as assessments typically depend on adolescent self-reports or caregiver responses during clinician-administered interviews. Adolescents may struggle to articulate their emotional experiences due to limited insight or reluctance, while caregivers may lack sufficient understanding of the adolescent's internal state, resulting in biased or incomplete evaluations.²² Research has established that depression arises from a combination of biological, psychological, social, and lifestyle factors.²³ Notably, experiences of childhood abuse are strongly linked to the onset of clinical mental disorders during adolescence.²⁴ Drawing on psychological, public health, and social ecological theories, this study examines how measurable aspects of adolescents’ daily environments, such as physical activity, diet, sleep, and family economic conditions, affect mental health, as analyzing these factors can provide valuable supplementary information for diagnosing depression and related mental health conditions.^25,26 Unhealthy dietary patterns, particularly high intake of fast food and sugary drinks, are linked to greater depression risk in adolescents, while nutrient-rich diets show protective effects, underscoring the potential of dietary interventions in promoting mental health.²⁷ Key environmental and social factors, including socioeconomic status and parental mental health, play a critical role in child outcomes²⁸ and should be integrated into mental health risk prediction models. Based on this, we formulated three progressively deeper hypotheses at the outset of our study: (1) machine learning can be used to predict depression risk; (2) developmental factors are associated with adolescents’ risk of depression; and (3) the relationship between environmental factors and adolescent depression risk is mediated by specific intervening variables.

Adolescent mental health disorders are often difficult to diagnose due to limited communication skills and reluctance to express emotions, leading to underreporting and misinterpretation. Recent advances in artificial intelligence, with their capacity for consistent and objective pattern recognition, provide a promising alternative to reliance on subjective reports. To address this challenge, we developed a predictive model using data from the National Survey of Children's Health (NSCH) to examine how objective aspects of adolescents’ living environments relate to depression risk. The model not only identifies key contributing factors but also captures their interrelationships, offering insights into the underlying mechanisms. To model depression risk, we applied machine learning and deep learning techniques, demonstrating their effectiveness through experimental validation. Additionally, to investigate potential causal factors, we employed a mediating effect²⁹ model to examine both direct and indirect effects between depression and physiological factors, offering insights into the underlying mechanisms of adolescent depressive disorders. Based on our findings, we propose relevant recommendations for early intervention and policy development.

Materials and methods

Data information

This study utilizes official comprehensive screening data to ensure the reliability of all findings. The NSCH from the United States³⁰ provides extensive information on various intersecting aspects of children's lives, including physical and mental health, access to quality healthcare services, and the family, community, school, and social environments that influence children. The available database encompasses annual census data from 2016 to 2021, derived from revised surveys conducted by the U.S. Census Bureau via mail and online methods. In our research, we have selected the 2021 survey data that is comprehensive and the latest revised.

The 2021 NSCH database includes a total of 50,892 survey cases, each comprising 443 variables. These variables cover broad domains such as caregiver (respondent) characteristics, basic sociodemographic information of the adolescent, physical and mental health status (including current conditions and medical history), as well as family and community environments. Each domain contains multiple structured questionnaire items with predefined response options. A simplified description of the survey instrument is provided in Table A1 in the Appendix, while detailed variable definitions and specifications can be found in the 2021 NSCH Topical Variable List.

Guided by the SDS, ICD-10, ICD-11,³¹ DSM-5,³² SCL-90, and MINI, we selected corresponding or potential causes and outcomes relevant to this study. The potential causes encompass other psychological disorders, physical diseases, growth environments, and lifestyle habits, while the outcomes include mental health issues such as depression, anxiety, autism, hyperactivity, and Attention Deficit Disorders (ADD). Logically, identifying potential cause factors would help develop preventive and treatment strategies.

Based on the aforementioned guidelines and a review of historical research related to adolescent mental health, we have selected the key reference variables as follows: gender,²⁹ age,³⁰ family environment,³¹ and physical activity,^33–35 along with other influencing factors.^36–39 Based on a review of the literature, we identified 60 variables with potential impacts on mental disorders for further investigation. Descriptive statistics and sample variable details are presented in Table A2 in the Appendix.

Sample inclusion and exclusion criteria

The original dataset covers minors aged 0–17 years. As this study focuses on adolescents, we restricted the sample to individuals aged 12–17 years. After excluding cases outside this age range, the final analytical sample comprised 17,539 participants. Among these, 2212 had been diagnosed with depression, 3491 with anxiety, and 2769 with ADD/Attention Deficit Hyperactivity Disorder (ADHD), with some overlap across diagnostic categories. Notably, over 80% of all reported mental health diagnoses in the 2021 NSCH dataset were concentrated within the 12–17 age group, which is consistent with general epidemiological trends.

Machine learning dataset

We used a binary indicator of depression diagnosis (K2Q32A) as the target outcome variable. In the mediation analysis, 12 variables reflecting socioeconomic status, physical activity levels, and dietary habits were selected as observed predictors. All cases with missing values on any of the included variables were excluded to ensure data completeness and model reliability. After applying the data exclusion criteria and removing records with missing values, a total of 12,856 samples remained for machine learning analysis. Among these, 1778 were positive cases of depression. Additionally, the dataset was split into training and testing sets in an 8:2 ratio.

Hypothesis

Hypothesis 1: Machine learning can predict adolescent depression risk using daily-life data.

We hypothesize that machine learning models trained on objective daily-life factors from census data such as physical activity, sleep patterns, diet, and socioeconomic background—can effectively identify adolescents at risk of depression.

Hypothesis 2: Socioeconomic and environmental factors are associated with adolescent depression risk.

We propose that adolescents from lower socioeconomic or less supportive backgrounds are more prone to depressive symptoms, underscoring the strong influence of external living conditions on mental health—a foundation for our mediating effect model.

Hypothesis 3: The relationship between environmental factors and adolescent mental health is mediated by behavioral or psychological variables.

We further hypothesize that this association is not direct, but partially mediated by intermediate factors such as physical activity levels, dietary habits, or psychological capital (e.g. resilience, self-efficacy). This hypothesis is the basis for us to propose two mediating variables in the mediating effect model.

Method

This is an applied cross-sectional study using fixed-year census data, which does not permit tracking individual changes over time. We combine artificial intelligence prediction methods with mediation analysis theory to examine mental health associations. Prior to modeling, relevant variables were selected using Pearson correlation in SPSS 29.0, guided by diagnostic criteria from ICD-10 and ICD-11.

Prediction based on machine learning

In the context of machine learning modeling and depression risk prediction, we implemented various algorithms in Python, including linear regression method,⁴⁰ decision trees,⁴¹ XGBoost,⁴² SVMs, and neural networks.⁴³ The process involved first standardizing the dataset, followed by model development and training to establish a predictive framework. These models are capable of assessing the risk of mental disorders using general survey data, serving as effective tools for preliminary risk screening. In summary, the outcome variable in our prediction task is whether adolescents aged 12–17 have a history of depression, while the explanatory variables include other mental health conditions and objective daily-life factors derived from the corresponding samples. Consistent with methodological standards in predictive analytics, incomplete entries containing missing values were excluded to ensure the integrity and reliability of the model's input parameters.

Mediating effects statistical methods

Based on the experimental results and a review of related literature,^44,45 we found a significant association between living conditions and adolescent mental health. To further investigate this relationship, we constructed a mediation model based on the methodological framework proposed by Frazier et al.,⁴⁶ demonstrating that living conditions exert indirect effects on mental disorders through two key mediators: physical activity levels and dietary quality. These mediators jointly contribute to the observed impact on adolescent psychological well-being. The proposed mediation model is illustrated in Figure 1. The optimized calculation formulas implemented in the Python library for mediating effect analysis are as follows:

Y = c X + e_{1}

M = a X + e_{2}

Y = c^{'} x + b M + e_{3}

Here, we use the statsmodels library to calculate the mediating effects. To illustrate this, we can refer to the following model:

Y = β_{0} + β_{1} X + β_{2} M + ϵ_{1}

M = γ_{0} + γ_{1^{x}} + ϵ_{2}

Figure 1.

Receiver operating characteristic (ROC) curve.

Evaluation metrics

We employed distinct evaluation strategies for data screening and model assessment. In the initial variable selection, Pearson correlation coefficients (r) with p < 0.05 were used to identify factors significantly associated with depression.

For machine learning models, performance was evaluated on the held-out test set using multiple metrics. Accuracy (ACC) reflects the overall proportion of correctly classified cases. However, because the dataset is imbalanced—with far fewer depression cases than nondepression cases, we prioritized the area under the curve (AUC), which measures a model's ability to discriminate between classes across all decision thresholds and is insensitive to class distribution. We also reported the F1-score, the harmonic mean of precision and recall, which provides a balanced measure of performance on the minority (depressed) class.

Results

Descriptive statistics of the sample

After data cleaning and preprocessing, a final dataset comprising 17,539 valid entries was selected for analyses, representing the effective portion of the collected data. Table 1 presents the detailed information of the data employed in our study. Among the individuals within this large sample, 1963 adolescents were identified as having depression. The distribution of depression scores was 1.11 ± 0.316 (with 2 indicating depression and 1 indicating noncases). Additionally, 3369 adolescents were identified with anxiety, with an anxiety score distribution of 1.19 ± 0.395. The study focused on individuals aged 12 to 17, with an average age of 14.68 ± 1.697 years.

Table 1.

Descriptive statistics of the dataset.

Identities		n	Proportion (%)	Have ever been medically diagnosed with depression
Identities		n	Proportion (%)	n	Proportion (%)
Sex	Male	9103	51.90	852	9.36
Sex	Female	8436	48.10	1161	13.76
Age (years)	12–13	5181	29.54	404	7.80
	14–15	5762	32.85	621	10.78
	16–17	6596	37.61	988	14.98
Have ever been medically diagnosed with depression	Yes	1963	11.19	-	-
Have ever been medically diagnosed with depression	No	15,526	88.52	-	-
Have ever been medically diagnosed with anxiety	Yes	3369	19.21	1582	46.96
Have ever been medically diagnosed with anxiety	No	14,106	80.43	431	3.06
Have ever been medically diagnosed with autism	Yes	661	3.77	201	30.41
Have ever been medically diagnosed with autism	No	16,800	95.79	1804	10.74

As shown in Table 1, the distribution of adolescents by gender and age group is approximately equal. Notably, the proportion of individuals who have experienced depression is slightly higher among females than males, which aligns with findings from previous literature. Additionally, the prevalence of depression increases progressively with age, consistent with prior research studies. Descriptive analyses indicate that the dataset adheres to statistical patterns, making it suitable for further research and analyses. For the subsequent risk prediction experiments, the dataset was refined by removing all entries with missing values. The cleaned data was then divided into a training set and a validation set, using a ratio of 5:1. This resulted in 10,111 entries in the training set and 2000 entries in the validation set. The distribution of depression and nondepression cases in the processed dataset is detailed in Table 2.

Table 2.

Data samples used for prediction.

Samples	n
Samples	History of depression	Without depression
Train set	1023	9088
Val set	194	1806

Correlation analyses

In the intercorrelation assessment, depression diagnosis was treated as the dependent variable while all other factors were considered independent variables. The correlation coefficients are presented in Table 3. All the selected factors were included in the correlation matrix, categorized into four groups: factors related to psychological disorders, physiological issues, life-related problems, and supportive assistance. Due to space limitations, only the factors that showed significant correlations from each category are listed.

Table 3.

Correlation analyses of related factors.

Variable		Correlation with mental illness
Variable		Depression	Anxiety	Autism
Psychological	Depression	1	0.535	0.231
	Anxiety	0.535	1	0.284
	Autism	0.231	0.284	1
Physiological	Have experienced physical pain or trauma within the past year	0.195	0.221	0.065
	Have had (physiological) health problems within one year	0.193	0.178	0.134
	Frequent headaches	0.163	0.194	0.064
Life issues	Parents have emotional issues	0.22	0.227	0.132
	Work hours within 1 year due to health issues	0.206	0.189	0.122
	People living together have problems with alcohol/drug abuse, etc.	0.171	0.141	0.109
	Is the bedtime routine regular	−0.161	−0.141	−0.099
	Do family members communicate frequently	−0.159	0.15	−0.151
Subjective issues	Difficulty concentrating/remembering/making decisions	0.326	0.37	0.476
	Behavior problem	0.276	0.29	0.49
	Difficult to handle problems alone	0.201	0.26	0.196

Except for autocorrelation items, all other correlation tests showed p < 0.01.

Test of intercorrelation coefficients²⁹ indicate that depression is significantly associated with symptoms such as anxiety and autism, all of which fall within the category of psychological disorders. The study focused on general behavioral performance and external objective factors, revealing that behaviors or factors with a high correlation to depression include the following:

Physiological issues: associated with bodily pain and trauma, health-related issues, and headaches.

Life-related problems: related to emotional problems with parents, decreased work and study time due to health issues, interactions with individuals who have alcohol and drug abuse problems, shared bedtime routines, and frequency of family communication.

Subjective issues: related to difficulties in concentrating or making decisions, behavioral problems, and managing issues independently.

Prediction and diagnosis of depression

Based on the descriptive analyses presented, the samples from this study can be effectively utilized to identify external factors and daily behavior metrics for predicting depression. In this effort, we employ various models, including logistic regression, decision trees, SVMs, and multilayer neural networks, to enhance prediction capabilities. Through the creation of a comprehensive dataset and validation processes, we established that the factors identified in the section “Correlation analyses” are valuable for prediction, achieving a high level of accuracy in forecasting the disease using standard machine learning techniques. As illustrated in Table 4 and Figure 2, each model demonstrated a relatively consistent and objective predictive performance.

Figure 2.

XGBoost feature importance.

Table 4.

Logistic regression predicting the presence of depression.

Method	Logistic regression	Deep Neural Network (DNN) with 10 layers	XGBoost	Decision tree
Accuracy	0.856	0.852	0.860	0.914
AUC	0.884	0.878	0.873	0.614

AUC: area under the curve.

The dataset used in this study was collected directly from the general public, resulting in a class imbalance due to the higher prevalence of nonclinical (nondisorder) cases. In such a setting, prediction models may achieve an ACC of 0.8 or higher simply by favoring the majority class (negative), although such results lack meaningful statistical validity. To provide a more reliable evaluation of model performance, we employed the receiver operating characteristic (ROC) curve and the associated metrics. Results showed that all models achieved an ACC above 0.85, while the AUC exceeded 0.87 for all except for the decision tree model, indicating stronger predictive power in the first three methods. In terms of overall performance, the models ranked as follows: XGBoost, logistic regression, simple neural network, and decision tree. These findings suggest that models based on XGBoost, logistic regression, and simple neural networks are particularly well-suited for predicting psychological disorders using survey-based data. Moreover, in the XGBoost method, four research subjects received importance scores above 0.1, indicating a strong correlation with depression prediction, as shown in Figure 3.

Figure 3.

Schematic diagram of the initial model.

Analysis of mediating effects on adolescent depression

The correlation matrix and subsequent analyses reveal a strong association between various psychological disorders. Notably, the diagnosis and treatment of depression demonstrate a significant correlation with anxiety disorders (r = 0.552, p = 0.000) and tic disorders (r = 0.364, p = 0.001), both exhibiting significant positive correlations (p < 0.01). Building on these correlation results, the selected neural network prediction method can effectively illustrate the relationship between living standards and psychological disorders, using depression as a focal point to explore potential factors and mediating effects related to the development of depressive symptoms. Through these analyses, we establish statistically significant relationships among various factors affecting depression research, with meaningful correlations identified. In this study, individuals with depression represented 11.2% of the total sample, and these correspond to the basic patterns observed in respective age surveys.

Mediating effects statistical study

Mediating effects were analyzed using SPSS and Python. Intercorrelations among variables were assessed, as correlation is a prerequisite for mediation testing. Model fit was evaluated via maximum likelihood estimation, and indirect effects were tested using a bootstrap method with p < 0.01. Consistent with prior literature, anxiety and history of psychological treatment were strongly associated with depression, whereas other associations were weaker, likely due to large data volume, irrelevant information increasing entropy, missing values, and dichotomization of variables, which reduced correlation strength. We conducted multiple sampling and framed our evaluations based on Root Mean Square Error of Approximation (RMSEA) and Root Mean Square Residual (RMR) principles.⁴⁷ We considered a correlation of 0.1 or higher as indicative of a relationship, and a correlation of 0.2 or higher as significantly related. These correlations are deemed effectively relevant.^48,49

Mediating effect modeling in adolescent depression

As illustrated in Figure 4, we constructed the mediating effect model based on the analysis results presented in the section “Correlation analyses.” Model 1 establishes the relationship between living standards and depression. Model 2 introduces physical activity as a mediating variable. Building on model 2, model 3 incorporates both physical activity and dietary habits as dual mediating variables.

Figure 4.

Adolescent lifestyle and depression model based on the mediating effects of physical activity and dietary levels.

Mediators of the impact of adolescents’ living standards on depression tendency

All potential mediators in this study were identified based on existing literature on factors associated with adolescent depression and further informed by empirical analyses of the variables available in the dataset. Specifically, we constructed three composite groups: living standards, physical activity levels, and dietary levels.

Correlation between adolescents’ living standards and depression diagnosis

Through factor analysis and dimensionality reduction of multiple experiences related to living standards (adverse childhood experiences, ACE series), we calculated the living standards component for adolescents. As shown in Table 5, the correlation results indicate a positive relationship between adolescents’ living standards and depressive symptoms (K2Q32), with a correlation coefficient of r = 0.225 and p = 0.001. The variables associated with living standards include: the frequency of difficulty maintaining basic expenses (ACE1), parental divorce or separation (ACE3), the death of a parent or guardian (ACE4), parental or guardian incarceration (ACE5), and living with individuals with tendencies toward alcohol or drug abuse (ACE9).

Table 5.

Correlation matrix of living standards variables.

Variable	ACE1	ACE3	ACE4	ACE5	ACE9	K2Q32
ACE1	1.000
ACE3	−0.255**	1.000
ACE4	−0.067**	0.095**	1.000
ACE5	−0.208**	0.303**	0.145**	1.000
ACE9	−0.190**	0.326**	0.121**	0.374**	1.000
K2Q32	−0.161**	0.138**	0.065**	0.127**	0.171**	1.000

ACE: adverse childhood experience. **Denotes a higher significance level of p < 0.01, showing even stronger statistical significance at the 99% confidence level.

Mediating effects of adolescent physical activity levels on depression tendency

Using structural equation modeling, our analysis showed good performance following maximum likelihood estimation. However, the data from public databases restricted our ability to specialize in various aspects, such as levels of physical activity and dietary habits. Consequently, we could not conduct further investigation into specific factors, including exercise duration, types of physical activities, and individual exercise preferences. To address this limitation, we selected relevant exercise-related variables and tested them for dimensionality reduction, ultimately forming a composite factor for physical activity levels. The specific physical activity variables included: participation in extracurricular physical education courses (K2Q30), involvement in sports clubs outside of school (K2Q31), frequency of participation in physical activities over the past year (K2Q33), and the number of days engaged in physical exercise in recent weeks (PHYSACTIV). All variables exhibited a significance level below 0.01, demonstrating their suitability for inclusion in the physical activity factor. The path coefficient for the lifestyle-physical activity-depression model is given in Table 6.

Table 6.

Correlation matrix for the physical activity factor.

Variable	PHYSACTIV	K7Q30	K7Q31	K7Q33
PHYSACTIV	1.000
K7Q30	−0.321**	1.000
K7Q31	−0.149**	0.302**	1.000
K7Q33	−0.176**	0.203**	0.176**	1.000

**Denotes a higher significance level of p < 0.01 , showing even stronger statistical significance at the 99% confidence level.

The mediating effect of dietary levels on depression in adolescents

The nutritional levels obtained through dimensionality reduction based on multiple variables, including dietary habits, consist of the following components: food consumption levels over the past year (FOODIST), participation in food voucher or supplemental nutrition assistance programs (K11Q61), and utilization of free or low-cost breakfast or lunch provided by schools (K11Q62). The mediating effect indicates that in the dimensionality reduction analysis, all significance levels (p-values) were less than 0.01, making them suitable for integration into the dietary level variable. The path coefficients for the lifestyle-dietary level-depression model and the lifestyle-dietary level-physical activity-depression model are shown in Figure 1.

Discussion

Hypotheses and results analysis

In the “Method” section, we progressively proposed three hypotheses. Based on our experimental results, we now analyze and discuss the validity of each hypothesis.

Hypothesis 1: Machine learning methods can assist in predicting adolescent depression risk. The experimental results show that the models achieved relatively high performance in terms of both ACC and AUC, reaching levels that could be considered clinically informative. In particular, the XGBoost model performed best among all the tested models. Moreover, through variable importance analysis, we were able to interpret the key factors influencing model predictions. Therefore, we conclude that hypothesis 1 is supported by the results.

Hypothesis 2: Socioeconomic status influences the likelihood of adolescents developing mental health issues—in this case, depression. To test this hypothesis, we selected five variables from the census dataset representing socioeconomic status as independent variables, and used them to predict whether an adolescent had been diagnosed with depression. The overall effect size was found to be 0.225. While not extremely large, given the scale and geographic diversity of the census data, this represents a meaningful impact. Thus, we consider hypothesis 2 to be valid.

Hypothesis 3: Socioeconomic status affects depression risk through mediators such as physical activity and dietary habits. As shown in the section “Mediating effect modeling in adolescent depression,” socioeconomic status significantly influenced dietary quality, and to a lesser extent, physical activity levels—both of which align with common expectations: families with better living conditions tend to place greater emphasis on nutrition and are more likely to invest in sports-related activities. These two factors, in turn, were found to have an impact on mental health, although the effect of diet was relatively weaker. Based on these findings, we confirm that hypothesis 3 is also supported.

Taken together, the findings support all of our stated hypotheses. Lifestyle factors influence adolescent depression through both a direct pathway, which accounts for 66.4% of the total effect, and indirect pathways mediated by physical activity and dietary quality, which together constitute 33.6% of the total effect. Physical activity alone explains 15.6% of the total effect and represents 46.5% of the overall mediating effect.

Although the estimated path coefficients are relatively small, this is consistent with known limitations of large-scale survey data, such as measurement error, variable dichotomization, and contextual heterogeneity. Nevertheless, the statistically significant and directionally consistent mediation effects support the plausibility of the proposed behavioral pathways. These findings indicate that lifestyle conditions influence mental health both directly and indirectly. Dietary quality, for example, may promote mental well-being not only on its own but also by encouraging greater physical activity, thereby amplifying its indirect effect on depression risk.

Insights and comparison with existing work

Current approaches to identifying adolescent depression largely depend on standardized psychological scales or clinician-administered interviews. These methods, while clinically validated, face practical limitations in large-scale or preventive settings. They require trained personnel, rely heavily on subjective self-disclosure, and are often influenced by transient emotional states or social desirability bias. In contrast, our study demonstrates that machine learning models can achieve robust predictive performance using only routinely collected, nonclinical survey variables. Specifically, models such as XGBoost and logistic regression attained an ACC of 0.85 and an AUC exceeding 0.87 in distinguishing adolescents with and without a depression diagnosis. The key predictors included not only psychological indicators such as anxiety and difficulty concentrating but also physical health complaints, family dynamics, and aspects of daily functioning, all of which are easily observable and reportable without specialized training.

This finding highlights a key contribution of our work: depression risk can be reliably flagged through passive, population-level data rather than active clinical probing. By moving beyond symptom-based self-reports to objective markers of daily living conditions, our approach enables early, nonintrusive screening at scale, particularly valuable in school, community, or digital environments where traditional tools face feasibility, stigma, or resource barriers.

Equally important, our work moves beyond black-box prediction by incorporating mediation analysis to uncover the pathways linking environmental adversity to mental health outcomes. While many machine learning studies stop at feature importance rankings, we explicitly model how socioeconomic and familial stressors such as financial strain, parental separation, or exposure to substance use translate into depressive symptoms. Our results show that these living conditions exert both a direct effect and a substantial indirect effect mediated through lifestyle behaviors. Specifically, physical activity levels and dietary quality jointly account for over one-third of the total association, highlighting their role as modifiable buffers against environmental risk.

This integration of predictive modeling with causal inference represents a methodological advance. It bridges the gap between data-driven pattern recognition and theory-driven psychological understanding. Unlike purely correlational studies, our framework offers interpretable mechanisms; unlike purely mechanistic studies, it validates those mechanisms within a high-performing predictive system. Together, these contributions provide not only a tool for identification but also a roadmap for intervention, suggesting that promoting healthy behaviors may mitigate the mental health impact of adverse living conditions, even when those conditions themselves cannot be immediately changed.

The interventional role of physical exercise and activity levels in adolescent mental disorders

The second section of this study introduced a mediation effect modeling approach. The third section presented the results derived from applying this modeling framework to the empirical data, followed by a brief analysis. These findings provide partial support for our initial hypothesis regarding the mediating effect involved in adolescent depression. Through the use of factor analysis and dimensionality reduction with multiple items within the same dimension, we developed three descriptive variables: lifestyle, physical activity level, and dietary level. Our findings indicate that adolescent lifestyle has a direct impact on mental health, particularly regarding depression. Our mediation analysis shows that adolescent lifestyle is significantly associated with depression, with a direct effect accounting for about two-thirds of the total association and the remaining one-third mediated jointly by physical activity and dietary quality. These findings are consistent with the evidence mentioned in the Introduction, which highlights physical activity and healthy dietary patterns as protective factors against adolescent depression.

The results also suggest that adverse living conditions may indirectly increase depression risk by limiting engagement in these health-promoting behaviors. Notably, physical activity accounts for nearly half of the total indirect effect, underscoring its potential as a key modifiable factor, particularly for adolescents from disadvantaged backgrounds.

Limitations and suggestion for future studies

This study has several limitations. The data were drawn from a fixed, publicly available U.S. census-type database. As a secondary source, it did not permit access to additional details, refinement of variables, or longitudinal follow-up. Consequently, some project descriptions lacked completeness and methodological rigor. The sample encompassed the entire U.S. adolescent population, which enhances national representativeness but may limit applicability to specific regions or subgroups. Furthermore, the cross-sectional design, based on a single time point, precluded temporal analysis of symptom trajectories and made causal inference unfeasible. Many measures also relied on passively recorded items with limited response options, such as binary choices, which may reduce measurement sensitivity and weaken observed associations.

Future research should incorporate longitudinal data collection to track changes over time and enable more robust causal inference. Developing evaluative models using diverse time-series data would improve both accuracy and clinical relevance. Additionally, primary data collection with refined, multilevel items could address the limitations inherent in passive, pre-existing database records.

Conclusions

Establishing a census-based database plays a crucial role in supporting the physiological and psychological developments of adolescents. By developing machine learning models that leverage this census data, we can effectively assess and predict mental health status, facilitating risk evaluation and early intervention. The identification of mediating effect modeling through qualitative and quantitative analyses highlights how disparities in lifestyle significantly impact adolescent depression. Additionally, both physical activity and dietary levels contribute to mediating effects related to mental health. By concentrating on improving adolescents’ lifestyle choices and effectively guiding their exercise and dietary behaviors, we can help prevent and alleviate depressive symptoms to a certain extent. This strategic focus not only addresses existing issues but also promotes overall well-being in the adolescent population.

Footnotes

Acknowledgments

We are grateful to all participants in our study.

ORCID iDs

Qiang Li

Hefeng Zhou

Gang Xu

Ethical considerations

This study is based on data from a publicly available database, which has been previously deidentified and approved for academic research purposes. The data was released by the U.S. Census Bureau under approval number CBDRB-FY21-POP001-0161. The U.S. Census Bureau reviewed this data product for unauthorized disclosure of confidential information and approved the disclosure avoidance practices applied to this release. As the data are publicly available and contain no personally identifiable information, this study was exempt from institutional review board approval. As no live experiments were conducted and the data used are anonymized, this research does not require additional ethical approval or informed consent.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This study supported by Jiangxi Provincial Department of Education Teaching Reform Project—Research on Physical Education Curriculum and Teaching Models Based on the Current Situation of Physical Fitness among University Students in Jiangxi under the OBE Concept (Project No. JXJG-22-16-15).

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Availability of data and materials

Original data is available on the website: .

Appendix

Table A2.

The selected variables and their brief explanation.

Category	Count	Codes	Explanation
Basic information	10	A1_AGE, A1_MENTHEALTH, A1_EMPLOYED, A1_PHYSHEALTH, A2_AGE, A2_EMPLOYED, SC_AGE, SC_SEX, SC_ENGLISH, SC_RACER	The basic information of the sample, as well as the information of its guardian
Child health status	15	K2Q01, K2Q30A, K2Q31A, K2Q32A, K2Q33A, K2Q38A, K2Q40A, MEMORYCOND, STOMACH, TOOTHACHE, HEADACHE, PHYSICALPAIN, OVERWEIGHT, ALLERGIES, BREATHING	The health conditions (psychological and physiological) and disease history of teenagers
Family and community	15	ACE1, ACE3, ACE4, ACE5, ACE6, ACE7, ACE8, ACE9, K10Q12, K10Q13, K11Q60, K11Q61, K11Q62, FAMCOUNT, FAMILY_R	Factors related to the family's economic situation (including housing, dietary levels, etc.) and the community environment
Child development	12	K6Q15, K6Q20, K6Q60_R, K7Q02R_R, K7Q30, K7Q31, K7Q32, K7Q33, K7Q37, BEDTIME, PHYSACTIV, STARTSCHOOL, MAKEFRIEND	The development of teenagers’ comprehensive qualities such as education and sports
Healthcare utilization	8	K5Q40, K5Q41, DOCPRIVATE, NOTOPEN, HCCOVOTH, ALTHEALTH, ARRANGEHC, COVIDARRANGE	The situation of health security, such as medical insurance, regular medical treatment, physical examinations, etc.

References

Jiao

Liu

, et al. The current situation of child and adolescent psychological problems and intervention in China. Asian J Med Sci 2016; 7: 15–17.

Burns

Birrell

. Enhancing early engagement with mental health services by young people. Psychol Res Behav Manag 2014; 7: 303–312. DOI: 10.2147/PRBM.S49151

Wang

Peng

, et al. Parental attachment and fear of missing out among Chinese adolescents: a moderated mediation model. J Child Fam Stud 2024; 33: 2324–2336.

Romera

Delgado-Cohen

Perez

, et al. Factor analysis of the Zung Self-Rating Depression Scale in a large sample of patients with major depressive disorder in primary care. BMC Psychiatry 2008; 8: 1–8.

Weisz

Sandler

Durlak

, et al. Promoting and protecting youth mental health through evidence-based prevention and treatment. Am Psychol 2005; 60: 28.

del FArmaco

. Descrizione degli indicatori e fonti dei dati. Rapporto Osservasalute 2013; 2013: 373.

Zung

. A self-rating depression scale. Arch Gen Psychiatry 1965; 12: 63–70.

Lecrubier

Sheehan

Weiller

, et al. The mini international neuropsychiatric interview (MINI). A short diagnostic structured interview: reliability and validity according to the CIDI. Eur Psychiatry 1997; 12: 224–231.

Sheehan

Lecrubier

Sheehan

, et al. The mini-international neuropsychiatric interview (MINI): the development and validation of a structured diagnostic psychiatric interview for DSM-IV and ICD-10. J Clin Psychiatry 1998; 59: 22–33.

10.

Derogatis

Kathryn

. The SCL-90-R and brief symptom inventory (BSI) in primary care. In: Handbook of psychological assessment in primary care settings. Taylor & Francis Group, 2000, pp.310–347.

11.

Kohrt

Rasmussen

Kaiser

, et al. Cultural concepts of distress and psychiatric disorders: literature review and research recommendations for global mental health epidemiology. Int J Epidemiol 2014; 43: 365–406.

12.

Hearst

Dumais

Osuna

, et al. Support vector machines. IEEE Intell Syst Appl 1998; 13: 18–28.

13.

Matarazzo

. The reliability of psychiatric and psychological diagnosis. Clin Psychol Rev 1983; 3: 103–145.

14.

Gibson

Can

Georgiou

, et al. Attention networks for modeling behaviors in addiction counseling. In: Interspeech 2017, pp.3251–3255.

15.

Galiatsatos

Konstantopoulou

Anastassopoulos

, et al. Classification of the most significant psychological symptoms in mental patients with depression using Bayesian network. In: Proceedings of the 16th international conference on engineering applications of neural networks (INNS), 2015, pp.1–8.

16.

She

Jia

, et al. Research on the identification of college students’ mental health problems based on campus big data. In: Journal of Physics: Conference Series 2020, p.052029. IOP Publishing.

17.

Lin

C-H

Lane

H-Y

. Early identification and intervention of schizophrenia: insight from hypotheses of glutamate dysfunction and oxidative stress. Front Psychiatry 2019; 10: 93.

18.

McGorry

Yung

Phillips

. The “close-in” or ultra high-risk model: a safe and effective strategy for research and clinical intervention in prepsychotic mental disorder. Schizophr Bull 2003; 29: 771–790.

19.

McGorry

Mei

. Ultra-high-risk paradigm: lessons learnt and new directions. BMJ Ment Health 2018; 21: 131–133.

20.

Zeng

Peng

, et al. The effect of childhood maltreatment on college students’ depression symptoms: the mediating role of subjective well-being and the moderating role of MAOA gene rs6323 polymorphism. Eur J Dev Psychol 2022; 19: 438–457.

21.

Gorham

Sadeghi

Eisner

, et al. Clinical utility of family history of depression for prognosis of adolescent depression severity and duration assessed with predictive modeling. J Child Psychol Psychiatry 2022; 63: 939–947.

22.

Klein

Goldstein

Finsaas

. Depressive disordersChild and adolescent psychopathology. 3rd ed. Wiley, 2017, pp.610–641. DOI: 10.1002/9781394258932.ch18.

23.

Lopresti

Hood

Drummond

. A review of lifestyle factors that contribute to important pathways associated with major depression: diet, sleep and exercise. J Affect Disord 2013; 148: 12–27.

24.

Nelson

Klumparendt

Doebler

, et al. Childhood maltreatment and characteristics of adult depression: meta-analysis. Br J Psychiatry 2017; 210: 96–104.

25.

Steinsbekk

Skoog

Wichstrøm

. Symptoms of depression, physical activity, and sedentary time: within-person relations from age 6 to 18 in a birth cohort. J Am Acad Child Adolesc Psychiatry 2025.

26.

Luo

Liu

Sun

, et al. Gender mediates the mediating effect of psychological capital between physical activity and depressive symptoms among adolescents. Sci Rep 2025; 15: 10868.

27.

Davison

Stewart-Knox

Connolly

, et al. Exploring the association between mental wellbeing, health-related quality of life, family affluence and food choice in adolescents. Appetite 2021; 158: 105020.

28.

Rivenbark

Arseneault

Caspi

, et al. Adolescents’ perceptions of family social status correlate with health and life chances: a twin difference longitudinal cohort study. Proc Natl Acad Sci USA 2020; 117: 23323–23328.

29.

Cohen

Huang

Chen

, et al.

Pearson correlation coefficient

Noise reduction in speech processing. Springer Nature, 2009, pp.1–4.

30.

Finlay

Mueller-Smith

Street

. Measuring intergenerational exposure to the us justice system: evidence from longitudinal links between survey and administrative data. In: 2021 APPAM fall research conference APPAM, 2022.

31.

First

. Harmonisation of ICD-11 and DSM-V: opportunities and challenges. Br J Psychiatry 2009; 195: 382–390.

32.

Regier

Kuhl

Kupfer

. The DSM-5: classification and criteria changes. World Psychiatry 2013; 12: 92–98.

33.

Yang

. The mediating role of resilience in the effects of physical exercise on college students’ negative emotions during the COVID-19 epidemic. Sci Rep 2021; 11: 24510.

34.

Zhang

Zheng

. The effect of physical exercise on depression in college students: the chain mediating role of self-concept and social support. Front Psychol 2022; 13: 841160.

35.

Liang

Yue

Liu

, et al. How does social capital affect individual health among the elderly in rural China?—mediating effect analysis of physical exercise and positive attitude. PLoS One 2020; 15: e0231318.

36.

Zou

Sun

Yang

, et al. Associations between negative life events and anxiety, depressive, and stress symptoms: a cross-sectional study among Chinese male senior college students. Psychiatry Res 2018; 270: 26–33.

37.

Pang

Wang

Zhang

, et al. Self-concept clarity and loneliness among college students: the chain-mediating effect of fear of negative evaluation and self-disclosure. Behav Sci 2024; 14: 94.

38.

Ersöz

. The role of university students’ general self-efficacy, depression and psychological well-being in predicting their exercise behavior. J Educ Train Stud 2017; 5: 110–117.

39.

Kvam

Kleppe

Nordhus

, et al. Exercise as a treatment for depression: a meta-analysis. J Affect Disord 2016; 202: 67–86.

40.

Weisberg

. Applied linear regression. John Wiley & Sons, 2005.

41.

Quinlan

. Induction of decision trees. Mach Learn 1986; 1: 81–106.

42.

Chen

Guestrin

. XGBoost: a scalable tree boosting system. In: Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, 2016, pp.785–794.

43.

Anderson

. An introduction to neural networks. MIT Press, 1995.

44.

Sampasa-Kanyinga

Colman

Goldfield

, et al. Combinations of physical activity, sedentary time, and sleep duration and their associations with depressive symptoms and other mental health problems in children and adolescents: a systematic review. Int J Behav Nutr Phys Act 2020; 17: 1–16.

45.

Awais

Chaudhery

Khan

, et al. Factors contributing to distress among school and college-going adolescents during COVID-19 lockdown: a cross-sectional study conducted in Sibi Balochistan, Pakistan. J Educ Health Promot 2021; 10: 17.

46.

Frazier

Tix

Barron

. Testing moderator and mediator effects in counseling psychology research. J Couns Psychol 2004; 51: 15.

47.

Kenny

Kaniskan

McCoach

. The performance of RMSEA in models with small degrees of freedom. Sociol Methods Res 2015; 44: 486–507.

48.

Kircanski

LeMoult

Ordaz

, et al. Investigating the nature of co-occurring depression and anxiety: comparing diagnostic and dimensional research approaches. J Affect Disord 2017; 216: 123–135.

49.

Jiang

She

Dill

S-E

, et al. Depressive and anxiety symptoms among children and adolescents in rural China: a large-scale epidemiological study. Int J Environ Res Public Health 2022; 19: 5026.