Introduction

Health-related quality of life (HRQoL) is a key indicator of adolescents’ health and overall well-being, encompassing physical, psychological, and social functioning as well as general health perceptions [1]. The World Health Organization defines quality of life as individuals’ perceptions of their position in life within the context of their culture and value systems. Given that lower HRQoL is consistently associated with a range of adverse outcomes, it has increasingly been incorporated into health assessment and decision-making [2]. Adolescence is a critical developmental period during which physical, psychological, and social functioning, as well as overall health, undergo substantial change. Multinational studies have consistently documented an age-related decline in HRQoL among children and adolescents, beginning as early as 10 years of age [3, 4]. This pattern highlights the need to identify modifiable factors linked to declining HRQoL. However, the existing literature has often examined determinants within isolated domains (e.g., individual, family, school, or broader social contexts) [5], which may obscure the relative importance of predictors and limit understanding of how multiple factors jointly relate to HRQoL. Moreover, much of the evidence has focused on children and adolescents with illness [6, 7], potentially underrepresenting socio-ecological and behavioral determinants relevant to the well-being of generally healthy adolescents.

Ecological systems theory posits that adolescent development is shaped by ongoing interactions within nested ecological contexts. In this process, school, family, and lifestyle behaviors together constitute the most immediate microsystems and directly shape individual developmental trajectories related to health [8]. School is the primary institutional setting in which adolescents encounter academic demands, behavioral norms, teacher expectations, and peer interactions. Family provides the most immediate source of emotional support, relational context, and resource background. Lifestyle behaviors, including diet, physical activity, sleep, screen time, smoking, and alcohol consumption, are viewed as important behavioral pathways through which the influences of school and family environments are translated into daily health and well-being outcomes. Therefore, incorporating school, family, and lifestyle into a unified analytical framework may facilitate a more comprehensive ecological understanding of the mechanisms underlying adolescent HRQoL.

Previous studies have shown that better school adjustment is associated with higher academic performance, more positive social integration, better mental health, and greater subjective well-being [9, 10]. In contrast, poorer academic performance and strained peer relationships have been linked to more antisocial and aggressive behaviors, as well as poorer mental health and HRQoL [11], and this issue may be particularly pronounced in the context of China’s highly competitive academic culture [12]. Supportive family environments, better parental relationship quality, higher parental educational attainment and income, and more abundant developmental resources have also been associated with higher levels of well-being and HRQoL [13,14,15]. In addition, previous evidence suggests that, compared with males and senior high school students, females and junior high school students are more likely to report stronger associations of poor school adjustment or family conflict with adverse outcomes [16, 17]. Meanwhile, multinational studies have supported cumulative associations between healthy lifestyle behaviors and adolescent HRQoL [18,19,20,21,22].

However, most existing studies have examined these factors separately within a single domain or have focused only on the independent associations between individual variables and HRQoL. This makes it difficult to determine the relative importance of family, school, and lifestyle factors, and it also limits the ability to identify their synergistic effects on adolescent HRQoL. In addition, the existing evidence has certain limitations in both study populations and analytical approaches. On the one hand, a large proportion of previous studies have been based primarily on clinical or diseased samples of children and adolescents, which may not adequately reflect the complex social-ecological and behavioral environments of the general healthy adolescent population. On the other hand, most studies on adolescent HRQoL have mainly relied on traditional statistical methods. Although these methods are useful for examining independent associations between variables, they are limited in their ability to simultaneously address the relative importance, nonlinear relationships, and interaction patterns of predictors across multiple ecological domains.

Therefore, this study is the first to apply interpretable machine learning to develop predictive models of HRQoL among Chinese children and adolescents across three domains: school adjustment, family characteristics, and healthy lifestyle. We further employed permutation feature importance, PDPs, and SHAP analyses to systematically examine the relative contribution, direction of effect, and interaction patterns of the predictors. In addition, we explored whether the contributions of these factors differed by sex and grade. Compared with traditional statistical methods, machine learning offers several advantages [23]. First, it can simultaneously integrate a large number of relevant predictors from different ecological domains to improve predictive accuracy. Second, it can detect nonlinear relationships and higher-order interaction patterns. Third, when combined with interpretability analyses, machine learning can provide not only prediction results, but also insight into which factors contribute most to model performance, whether their effects are positive or negative, whether threshold or interaction patterns exist, and how these patterns differ across subgroups. By incorporating school adjustment, family characteristics, and healthy lifestyle into a unified interpretable analytical framework, this study aimed to move beyond the limitations of single-domain and purely correlational analyses, and to provide a more comprehensive, interpretable, and developmentally informed framework for understanding adolescent HRQoL, while also offering evidence to inform targeted interventions in public health and educational practice.

Methods

Study design and population

This study used a combined sampling approach incorporating probability proportional to size (PPS) sampling and simple random sampling. Between 2017 and 2024, we annually and randomly selected 3 high schools and 7 middle schools from 10 prefecture-level cities in Shandong Province, China. Within each selected school, approximately 100 students per grade were randomly sampled, yielding an annual survey sample of approximately 30,000 students. The present analysis used data collected from 2021 to 2023. Overall, 28,149 questionnaires were obtained; after excluding questionnaires with missing key information or obvious logical inconsistencies, 20,236 questionnaires were retained for analysis. Eligibility criteria included (1) the ability to complete the questionnaire in full and (2) provision of informed consent from both the participant and their guardian. Participants with severe mental or physical illness that precluded completion of the self-report questionnaire were excluded. The study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of Shandong University (Approval No.: 20180517).

Measures

School adjustment

School adjustment was assessed using the School Social Behavior Scales (SSBS) developed by Kenneth W. Merrel [24]. The SSBS comprises 65 items, such as “Is able to cooperate with other classmates most of the time.” It consists of two domains and six dimensions: social competence and antisocial behavior. Social competence captures academic behavior, self-management, and peer relations, whereas antisocial behavior includes hostile–irritable behavior, antisocial–aggressive behavior, and defiant–disruptive behavior. All items were rated on a 5-point Likert scale, ranging from 1 (“never”) to 5 (“frequently”). Antisocial behavior items were reverse-coded, and total scores were calculated for each dimension; higher scores indicated higher levels of the corresponding dimension. In the present sample, the SSBS showed excellent internal consistency (Cronbach’s α = 0.970).

Family characteristics

Guided by family systems theory and socioeconomic status theory, we selected family-related predictors, including only-child status, interparental relationship quality, frequency of exercise with parents, parental educational levels, parental occupational status, and family economic level. Only-child status was assessed using a single-item question (“Are you an only child?”; yes/no). The interparental relationship quality was assessed using three questions addressing parental alcohol-related behaviors, parental quarrels, and the overall quality of the parental relationship. All responses were coded as yes/no and summed, with higher scores indicating better relationship quality. The frequency of exercising with parents was assessed using a single-item question: “How many times per week do you engage in physical activity with your parents?” Responses were coded on a 5-point ordinal scale (1 = 0 times, 2 = 1 time, 3 = 2 times, 4 = 3 times, 5 = 4 times or more), with higher scores indicating more frequent parent–child physical activity. Parental educational level was reported on a 9-level scale ranging from no formal education to a master’s degree or above. Parental occupational status was reported using a 13-level classification. Family economic level was self-assessed on a 5-point scale ranging from “very difficult” to “very affluent,” scored from 1 to 5. Detailed questionnaire items and coding schemes are presented in Table S1 [see Additional file 1].

Healthy lifestyle

Healthy lifestyle variables were informed by prior empirical research and included dietary habits, sleep duration, physical activity, smoking, alcohol use, and screen time. Diet was assessed using a food frequency–based score that covered both healthy and unhealthy food categories, with higher scores indicating a healthier dietary pattern. Sleep duration was measured as the average number of hours of sleep per night. Physical activity was assessed by the number of days in the past 7 days on which adolescents engaged in more than 30 min of physical activity per day, with higher scores indicating more frequent physical activity. Smoking and alcohol consumption were determined using multiple items regarding lifetime initiation of smoking or drinking, age at first initiation, and recent smoking or drinking status, with higher scores indicating greater exposure to smoking or alcohol consumption. Screen time was assessed by the amount of time spent on watching television and using a computer/playing video games for non-school purposes on weekdays, with higher scores indicating longer screen time. Detailed variable definitions and coding schemes are provided in Table S1 [see Additional file 1].

Health-related quality of life

Adolescents’ HRQoL was assessed using the Quality of Life Scale for Children and Adolescents (QLSCA) developed by Huazhong University of Science and Technology. This instrument has been incorporated into the Chinese national developmental assessment guideline (GB/T 31178 − 2014) [25] and has been used to measure HRQoL among Chinese children and adolescents [26]. The QLSCA comprises 49 items across four domains: psychosocial functioning, physical and psychological health, living environment, and life satisfaction. An example item is: “Are you satisfied with your life?” Each item is rated on a 4-point Likert scale (1 = “never” to 4 = “always”), with higher total scores indicating better HRQoL. In the present sample, the QLSCA demonstrated high internal consistency (Cronbach’s α = 0.947). As there is currently no unified national clinical standard for adolescent health-related quality of life (HRQoL) in China, the median split in this study was used solely as an operational threshold for the classification task rather than as a clinically meaningful or normative cutoff. Therefore, for the purposes of the present analysis, HRQoL was dichotomized using the median score (142 points): scores > 142 were coded as 1 (high HRQoL), and scores ≤ 142 were coded as 0 (low HRQoL).

Covariates

Covariates included sex (boys/girls), age, grade level (middle school/high school), ethnicity (Han/other), household registration type (agricultural/urban/resident/none), and place of residence (urban or county center, suburban or peri-urban area, township, other city or county, rural).

Statistical analysis

All analyses were performed in Python (version 3.12.3). Continuous variables are summarized as means and standard deviations, and categorical variables as frequencies and percentages. Before model development, Pearson correlation analysis, multivariable-adjusted least absolute shrinkage and selection operator (LASSO) regression, and variance inflation factor (VIF) diagnostics were used to mitigate overfitting. The dataset (N = 20,236) was randomly split into a training set (n = 16,189; 80%) and a test set (n = 4,047; 20%). We developed and evaluated nine supervised machine learning models—random forest (RF), ridge regression (RR), decision tree (DT), gradient boosting decision tree (GBDT), support vector machine (SVM), voting classifier (VC), multilayer perceptron (MLP), AdaBoost (AB), and k-nearest neighbors (KNN)—with logistic regression (LR) as the baseline [27]. All models underwent hyperparameter tuning; detailed configurations are provided in Table S2 [see Additional file 1]. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity (recall), specificity, false positive rate (FPR), false negative rate (FNR), positive predictive value (PPV), negative predictive value (NPV), and F1 score. Based on the best-performing model, permutation feature importance, partial dependence plots (PDPs) [28], and SHAP analyses were conducted [29]. To examine the robustness of the findings, multiple machine learning regression models were constructed to analyze continuous HRQoL, followed by model comparison, permutation feature importance, SHAP, and PDP analysis. Statistical significance was set at p < 0.05. The code is publicly available at: https://github.com/LLYYZZ666/Machine-Learning-Research/tree/main.

Results

Descriptive statistics, correlation analysis, and variable selection

A total of 20,236 adolescents were included in the final analysis (mean age = 15.97 years, SD = 1.82). Of these, 10,465 were female (51.71%), 14,739 were junior high school students (72.84%), and 5,272 lived in rural areas (26.05%). A total of 9,972 participants (49.28%) were classified into the high-HRQoL group, whereas 10,264 (50.72%) were classified into the low-HRQoL group. Other related characteristics are presented in Table S3 [see Additional file 1]. Pearson correlation coefficients for all study variables are presented in Figure S1 [see Additional file 1]; with the exception of ethnicity, all variables were significantly correlated with HRQoL (p < 0.001). LASSO regression showed that, at the optimal λ value (0.0029), the coefficients for ethnicity and household registration type were shrunk to zero (Figure S2 and Figure S3 [see Additional file 1]); these variables were therefore excluded from subsequent analyses. VIF diagnostics indicated that all predictors had VIF values < 10, suggesting no evidence of problematic multicollinearity in the final models.

Model selection, permutation importance, and partial dependence

Receiver operating characteristic (ROC) curves for the 10 predictive models are presented in Figure S4 [see Additional file 1]. The RF model yielded the highest area under the curve (AUC = 0.8752). Performance metrics on the test set are summarized in Table 1. Overall, the RF model showed the best performance, with an accuracy of 79.55% and an F1 score of 0.7869, together with balanced sensitivity (0.7821) and specificity (0.8079). Confusion matrices for all models are provided in Figure S5 [see Additional file 1]. Based on these results, the RF model was selected for subsequent interpretability analyses. In the regression analysis of continuous HRQoL, the Random Forest Regressor also emerged as the best-performing model (Table S4 [see Additional file 1]).

Table 1 Model evaluation results
Full size table

Using the RF model, we conducted permutation feature importance analyses for 24 variables to quantify the model’s reliance on each predictor. The 10 predictors with the greatest contributions to model performance were, in descending order, self-management, peer relations, diet, frequency of exercise with parents, sleep, hostile-irritable behavior, defiant-disruptive behavior, academic behavior, father’s education level, and screen time (Table 2). PDPs for the top six predictors further characterized their associations with the predicted probability of high HRQoL (Fig. 1). Higher levels of self-management, peer relations, healthier diet, and more frequent exercise with parents were associated with a higher predicted probability of high HRQoL and showed approximately linear patterns. Sleep duration displayed an inverted U-shaped pattern, with the predicted probability of high HRQoL increasing up to approximately 8–10 h per night and then slightly decreasing. Hostile-irritable behavior (lower scores indicating poorer behavior) was associated with a lower predicted probability of high HRQoL.

Fig. 1
Fig. 1
Full size image

Actual prediction plots and partial dependence plots for the top six important variables. (a) self-management/compliance; (b) peer relations; (c) diet; (d) frequency of exercise with parents; (e) sleep; (f) hostile/irritable

PDP interaction plots of the key predictors suggested clear joint effect patterns (Fig. 2). Specifically, when both self-management and peer relations were at higher levels, the predicted probability of high HRQoL was most favorable, highlighting the central importance of behavioral self-management and positive peer relations (Fig. 2a). Peer relations and healthy diet also showed a positive interaction, such that a healthier dietary pattern was associated with a more favorable prediction of HRQoL, particularly among adolescents with better peer relations (Fig. 2f). In addition, diet and the frequency of exercise with parents exhibited a positive joint pattern, with the most favorable predicted HRQoL observed when both healthy diet and frequency of exercise with parents were high (Fig. 2j). These findings suggest that school adjustment, family characteristics, and health behaviors may interact in shaping adolescent HRQoL, rather than operating independently.

Fig. 2
Fig. 2
Full size image

Interaction effects of key predictors on adolescent HRQoL. (a) self-management and peer relations; (b) self-management and diet; (c) self-management and frequency of exercise with parents; (d) self-management and sleep; (e) self-management and hostile/irritable; (f) peer relations and diet; (g) peer relations and frequency of exercise with parents; (h) peer relations and sleep; (i) peer relations and hostile/irritable; (j) diet and frequency of exercise with parents; (k) diet and sleep; (l) diet and hostile/irritable; (m) frequency of exercise with parents and sleep; (n) frequency of exercise with parents and hostile/irritable; (o) sleep and hostile/irritable

In the analysis of continuous HRQoL, permutation feature importance indicated that self-management was the most influential predictor, followed by defiant-disruptive behavior, peer relations, frequency of exercise with parents, hostile-irritable behavior, and sleep (Table S5 [see Additional file 1]). The PDP results were consistent with those observed for the categorical HRQoL analysis (Figure S6 [see Additional file 1]). Overall, the findings from the continuous regression analysis were broadly in line with those from the classification analysis.

Table 2 Permutation feature importance
Full size table

SHAP-based interpretability analysis

We used SHAP to quantify and interpret predictor contributions to HRQoL predictions at both the global and individual levels (Fig. 3). The local explanation plot showed that self-management, peer relations, academic behavior, and hostile-irritable behavior had the largest contributions to the model’s HRQoL predictions. The global SHAP summary plot ranked predictors by overall contribution and showed directionality: higher values of self-management, peer relations, and academic behavior were generally associated with positive SHAP values, whereas higher values of screen time, smoking, and alcohol use were associated with negative SHAP values. The decision plot visualized prediction pathways for the first 500 samples. In the waterfall plot for the individual with the highest predicted probability of high HRQoL, the predicted probability increased from the baseline value of 0.495 to 0.994. Self-management, peer relations, and academic behavior contributed most, with approximate SHAP values of 0.10, 0.08, and 0.07, respectively. The SHAP analysis based on continuous HRQoL further supported these findings, showing that variables from the school adjustment domain remained the most important contributors in the ranking of feature importance (Figure S7 [see Additional file 1]). Therefore, the results of the continuous HRQoL regression analysis were broadly consistent with the classification-based findings, supporting the reliability of the study conclusions.

Fig. 3
Fig. 3
Full size image

SHAP interpretability analysis of RF model predicting adolescent HRQoL. (a) SHAP force plot; (b) SHAP summary plot; (c) SHAP decision plot; (d) SHAP waterfall plot

Subgroup analysis

To examine domain-specific prediction patterns, we built separate models for three domains: school adjustment, family characteristics, and healthy lifestyle. The RF model showed the best performance across domains (Table S6-S8 [see Additional file 1]). SHAP analyses identified the most influential predictors within each domain: self-management and peer relations (school adjustment), frequency of exercise with parents and interparental relationship quality (family), and physical activity and screen time (lifestyle) (Fig. 4). With the exception of screen time, which was associated with negative SHAP values, the remaining key predictors were generally associated with positive SHAP values.

Fig. 4
Fig. 4
Full size image

Domain-specific SHAP interpretation of ML models for adolescent HRQoL. (a) school adjustment dimension; (b) family characteristics dimension; (c) healthy lifestyle dimension

One-way analysis of variance (ANOVA) and chi-square tests indicated significant sex differences in all predictors and the outcome (p < 0.05). Significant age differences were observed for all variables except only-child status (p = 0.107) and smoking (p = 0.065). We subsequently conducted machine learning modeling and SHAP analyses within sex- and age-stratified subgroups. In sex-stratified analyses, peer relations and physical activity showed larger contributions to HRQoL predictions in boys, whereas defiant-disruptive behavior, antisocial–aggressive behavior, and smoking were associated with more negative SHAP values. In contrast, self-management, frequency of exercising with parents, and sufficient sleep showed higher predictive importance in girls. In both sex groups, frequency of exercising with parents remained the most robust family-domain predictor associated with positive SHAP values (Figure S8 [see Additional file 1]).

Grade-stratified analyses indicated that, among high school students, frequency of exercise with parents, healthier diet, and physical activity showed larger positive contributions to HRQoL predictions, whereas among middle school students, defiant-disruptive behavior, antisocial–aggressive behavior, and screen time were associated with more negative SHAP values. Sufficient sleep also showed a larger positive contribution among middle school students. In both grade groups, the top four predictors were all from the school adjustment domain, and frequency of exercise with parents ranked as the second most important predictor. Within the healthy lifestyle domain, predictor importance was more evenly distributed among middle school students, whereas in high school students, smoking and alcohol use showed comparatively weaker associations with HRQoL (Figure S9 [see Additional file 1]).

Discussion

This study is the first to integrate machine learning with SHAP-based interpretability to examine the relative contributions, interactions, and sex- and grade-level differences of school adjustment, family characteristics, and healthy lifestyle in adolescents’ HRQoL. The results indicated that the five predictors with the strongest contributions to HRQoL predictions were all drawn from the school adjustment domain, highlighting the central role of school adjustment in this modeling framework. The analyses also identified specific, potentially modifiable predictors within the family and healthy lifestyle domains and suggested key differences by sex and grade. Overall, these findings provide an interpretable predictive framework that extends beyond single-domain approaches, offering additional insight into the relative importance and developmental variability of school, family, and lifestyle factors and informing more targeted, developmentally appropriate strategies.

Within the school adjustment domain, self-management and peer relations consistently emerged as the two most influential predictors in the model, with higher levels associated with a greater predicted probability of better HRQoL, consistent with prior research [30]. Previous evidence indicates that better school adjustment is associated with a stronger sense of competence, higher academic achievement, better mental health, and greater overall life satisfaction [31]. Conversely, poorer school adjustment—such as problematic peer relationships—is associated with higher levels of aggression and poorer mental health, and these associations appear more pronounced among East Asian adolescents than among their Western counterparts [11, 32]. Intrinsic self-management and adherence to academic norms have also been associated with academic success and psychological resilience. In this study, adolescents with higher self-management and better peer relations showed more favorable predicted HRQoL outcomes, which is of particular importance during adolescence. Self-management reflects not only behavioral compliance, but also emotional regulation, delayed gratification, stress management, and goal-directed functioning. Therefore, adolescents with stronger self-management abilities may be better able to cope with academic stress and everyday challenges, thereby reducing psychological burden and the likelihood of mental health problems. Peer relations are likewise critically important during adolescence. Sensitivity to peer acceptance, sense of belonging, and social comparison increases markedly during this developmental stage. Positive peer relations can provide emotional validation, interpersonal security, and subjective well-being, and may also promote learning motivation and academic achievement [33]. Previous studies have shown that peer relations exert a cumulative protective effect on adolescent mental health during both junior and senior high school, reducing the risk of depression. In contrast, peer rejection or strained peer relations may exacerbate loneliness, social stress, and the risk of mental health problems [34].

Notably, even after adjusting for family- and lifestyle-related factors, school adjustment remained a substantial contributor to HRQoL. This finding suggests that school is not merely an external setting in adolescent life, but rather a proximal and structurally central developmental context in shaping HRQoL. Compared with family factors or lifestyle behaviors, school adjustment encompasses multiple routine experiences, including academic performance, rule compliance, teacher expectations, peer interactions, and behavioral regulation. Because these school-based experiences occupy a large proportion of adolescents’ daily lives, they may more directly influence adolescents’ overall perceptions of their functional status and quality of life. In the Chinese context, this pattern may be further reinforced by a highly competitive and examination-oriented educational environment. Academic performance, class ranking, classroom discipline, and external evaluations are highly central to school life, making success or failure in school adjustment more likely to be internalized into adolescents’ judgments of their own competence, self-worth, and developmental prospects [35]. Accordingly, the persistent and prominent role of school adjustment observed in this study may reflect not only its statistical importance but also its contextually salient position in the developmental trajectories of Chinese adolescents. Overall, this finding is consistent with ecological systems theory [36].

Within the family domain, frequency of exercise with parents emerged as the most important predictor in the model, with higher frequency associated with a greater predicted probability of better HRQoL. Prior research has largely emphasized the adverse associations between negative family factors (e.g., poor parental relationship quality, frequent family conflict, and divorce) and adolescents’ HRQoL, with comparatively less attention to the potential relevance of positive family behaviors [37]. In this study, frequency of exercise with parents showed consistently positive contributions in both the overall model and the family-domain analyses. Previous evidence indicates that regular physical activity with parents is associated with lower cortisol levels [38]. Frequency of exercise with parents has also been linked to lower stress and to fewer anxiety and depressive symptoms, which may relate to better subjective well-being and overall quality of life [39]. Interparental relationship quality was additionally identified as a key family-domain predictor, with higher quality associated with more favorable HRQoL predictions, suggesting potential foundational and moderating roles within family functioning that are linked to emotional adaptation, prosocial behavior, and overall well-being. Accordingly, for middle and high school students in China, efforts to promote parent–child physical activity and strengthen parental relationship quality may inform targeted, developmentally appropriate strategies to support adolescent HRQoL.

Within the healthy lifestyle domain, regular physical activity and healthier dietary patterns were associated with better HRQoL, whereas insufficient or excessive sleep and longer screen time were associated with poorer HRQoL. In the context of increasing academic pressure and greater day-to-day autonomy, middle and high school students often report longer screen time, more sedentary behavior, and lower levels of physical activity, which are associated with poorer physical and mental health [40]. Conversely, regular physical activity and a balanced diet are associated with lower obesity risk, higher self-esteem, and greater psychological resilience [41], which may relate to better HRQoL. Overall, greater engagement in health-promoting behaviors is associated with higher well-being [42]. These findings support a cumulative perspective, suggesting that more favorable HRQoL is associated with the co-occurrence of multiple interrelated lifestyle factors.

In the analyses stratified by sex and grade level, peer relations, physical activity, and externalizing behaviors (such as defiance and smoking) showed relatively stronger associations with HRQoL among boys [43], whereas self-management, frequency of exercise with parents, and adequate sleep were more strongly associated with HRQoL among girls [5]. This pattern may reflect that boys are more sensitive to peer group dynamics and norms surrounding externalizing behaviors, whereas girls may be more influenced by self-regulatory demands, relational support, and restorative behaviors such as sleep. The associations of HRQoL with the frequency of exercise with parents, healthier diet, and regular physical activity were more pronounced among senior high school students [13], whereas junior high school students showed relatively stronger associations with externalizing behaviors, excessive screen use, and sleep patterns [44]. This may reflect differences in developmental tasks and socialization processes across stages of adolescence [45]. Early adolescence may be characterized by greater emotional reactivity, less stable behavioral regulation, and heightened susceptibility to environmental disturbances. In contrast, during later adolescence, against a background of increasing academic pressure and tightly structured daily routines, supportive family routines and self-maintained healthy behaviors may become increasingly important.

Implications and limitations

This study is the first to apply multiple machine learning models with SHAP-based interpretability in a large, representative sample of Chinese adolescents. By examining more than 20 variables spanning the school, family, and lifestyle domains, we identified individual predictors and interaction patterns relevant to HRQoL predictions. The findings of this study also have important practical implications. Given the prominent role of school adjustment, school-based strategies may be particularly effective in promoting HRQoL among adolescents. Interventions designed to strengthen self-management, emotion regulation, adaptive coping with academic stress, and supportive peer relationships may be especially beneficial. The family-related findings further suggest that interventions should focus not only on reducing conflict but also on fostering positive family routines, including regular parent–child physical activity. In addition, the lifestyle-related results support an integrated intervention approach targeting physical activity, diet, sleep, and screen time concurrently. Because the patterns of predictive factors differed across sex and grade levels, intervention priorities may need to be tailored to adolescents at different developmental stages rather than adopting a one-size-fits-all approach for all adolescent populations. Several limitations should be noted. First, the cross-sectional design precludes causal inference, and the directionality of the observed associations should be evaluated in longitudinal or experimental studies. Second, the use of self-reported measures may introduce reporting bias; future work should incorporate objective indicators to improve measurement accuracy and reliability. Third, because the sample was drawn from Shandong Province, China, the generalizability of these findings may be limited; cross-regional validation is needed to assess broader applicability.

Conclusions

Using multiple machine learning algorithms with SHAP-based explainability, this study identified school adjustment as the domain with the largest contributions to adolescents’ HRQoL predictions, with self-management and peer relations ranking as the most influential predictors. Within the family domain, frequency of exercise with parents emerged as the most prominent factor. Regular physical activity, sufficient sleep, and healthier diet were consistently associated with more favorable HRQoL predictions. Differences by sex and educational stage suggest that priorities for supportive strategies may vary across groups. Overall, these findings provide a more comprehensive and actionable framework to inform targeted, developmentally appropriate approaches to support adolescents’ HRQoL in educational, family, and public health settings.