2.2.1 Outcome variable.

The primary outcome was adherence to the MSE guideline. Students reported the frequency of activities designed to build muscular strength (e.g., push-ups, sit-ups, weight training) over the previous 7 days. Consistent with international guidelines, responses were binary-coded as meeting the recommendation (≥ 3 days/week) or not meeting the recommendation (0–2 days/week) [2].

2.2.2 Predictor variables.

Candidate predictors were defined a priori to reflect key domains of adolescent characteristics and context (see S1 Table for the full list of variables). Generally, variables were grouped into four domains: (1) sociodemographic factors (age, sex, and race/ethnicity); (2) weight status, derived from BMI-for-age percentiles; (3) lifestyle and health behaviours (e.g., dietary intake, physical activity, sleep duration, screen time); and (4) psychosocial and social-context indicators (e.g., depressive symptoms, bullying victimisation, school belonging).

Race/ethnicity (source variable: raceeth) was harmonised into four categories: White; Black or African American; Hispanic/Latino (including multiple Hispanic/Latino); and All Other Races (American Indian/Alaskan Native, Asian, Native Hawaiian/Other Pacific Islander, and multiple non-Hispanic/Latino). Weight status was derived from BMI-for-age percentiles (BMIPCT) using CDC cut-points and categorised as underweight (<5th percentile), healthy weight (5th to <85th), overweight (85th to <95th), and obesity (≥95th) [28]. Predictors were otherwise modelled using the survey’s original categorical or ordinal response options, with limited recoding applied to align response categories and reduce sparse cells where necessary.

In this study, dietary indicators (e.g., fruit and salad intake) were included as candidate predictors within the machine-learning model to capture behavioural patterns associated with muscle-strengthening exercise. In predictive modelling, variables are selected based on their ability to improve classification performance rather than on an assumed causal direction [25]. Previous research shows that dietary behaviours, physical activity, and weight-control practices commonly cluster among adolescents and often reflect a broader health-oriented lifestyle [23,24]. Therefore, dietary indicators were incorporated to represent this behavioural context. However, because the data are cross-sectional, the temporal ordering between diet and muscle-strengthening exercise cannot be determined. Reverse or bidirectional relationships remain possible.

2.3.1 Data preprocessing.

Prior to model training, all nominal categorical variables were converted into dummy variables using one-hot encoding, while ordinal variables were encoded as integer values to preserve their natural ranking structure. To prevent information leakage, the raw dataset was first randomly divided into a training set (70%) and an independent testing set (30%) before any imputation procedures were performed. Missing data were handled using multiple imputation by chained equations (MICE) [29]. The imputation procedure was conducted independently within the training and testing datasets. This approach ensured that information from the testing set did not influence the model development process. MICE was selected because it performs well with datasets containing mixed variable types (categorical and ordinal), which are common in the YRBS survey. Alternative approaches such as KNNImpute and MissForest were considered; however, MissForest can be computationally intensive for large datasets, while KNNImpute may be less suitable for complex categorical structures. To support consistent model interpretation, we used the 20th imputed dataset from the MICE procedure, after confirming stable Markov chain convergence through trace plots [30]. Currently, no standardized approach exists for combining SHAP explainability outputs across multiple imputed machine learning models. Therefore, using a single converged imputation dataset provides a practical solution for generating interpretable SHAP results. All candidate predictors defined a priori were included in the model without additional feature selection. The XGBoost algorithm applies internal regularization, which helps control model complexity and automatically adjusts the contribution of correlated predictors.

2.3.2 Model development.

An eXtreme Gradient Boosting (XGBoost) classifier was used to estimate the probability of meeting the MSE recommendation (≥3 days per week) [31]. Hyperparameters, including the learning rate (eta), maximum tree depth, and subsampling ratios, were optimized using a step-wise, granular grid search approach combined with 5-fold cross-validation within the training dataset. This iterative process involved an initial coarse-grained search over a broad parameter space to identify the optimal region, followed by a second stage of fine-grained refinement around the identified optima. This granular strategy was employed to ensure that the global optimum between initial grid lines was not missed, thereby improving model stability and reducing the risk of overfitting. The optimal hyperparameters were selected based on the highest mean validation area under the receiver operating characteristic curve (AUC). AUC was chosen because it measures classification performance across all possible probability thresholds and is less sensitive to class imbalance than simple accuracy. To reduce the risk of overfitting, early stopping was applied during model training. The boosting process was terminated when the validation AUC failed to improve for 20 consecutive iterations. This procedure, together with tree-depth constraints and subsampling, improves the generalizability of the final model.

2.3.3 Model evaluation.

Predictive performance was evaluated using the independent testing set. Performance metrics included accuracy, sensitivity, specificity, and AUC, with 95% confidence intervals (CIs) estimated using bootstrap resampling (1,000 iterations).

2.3.4 Model explainability.

To interpret the trained model, we applied SHapley Additive exPlanations (SHAP), a model-agnostic method derived from cooperative game theory [32]. SHAP quantifies how much each predictor contributes to the predicted probability for a given observation. In simple terms, SHAP values represent the change in the model output associated with a specific feature compared with the average prediction. In this study, mean absolute SHAP values (Mean |SHAP|) were used to quantify the overall importance of each predictor across all observations. Larger values indicate stronger influence on model predictions. We also generated SHAP summary plots, which display both the magnitude and direction of each variable’s contribution across the dataset. In these plots, each point represents an individual observation, and the color scale indicates the relative level of the predictor variable. To further examine non-linear relationships between predictors and the probability of meeting the MSE guideline, partial dependence plots (PDPs) were produced. PDPs show the average predicted probability of the outcome across the range of a predictor while holding other variables constant. Data management was conducted using SPSS version 27 and Stata version 17, and all machine learning analyses were performed in R version 4.4.0.

4.2.1 MVPA.

MVPA emerged as the strongest correlate of meeting the MSE guideline in our explainable ML models. This pattern is consistent with youth movement guidelines that position aerobic activity and muscle-strengthening as complementary targets rather than independent behaviours [2,14]. Evidence from U.S. surveillance data also shows substantial overlap between aerobic and muscle-strengthening behaviours among adolescents [19].

One plausible interpretation is behavioural clustering. Adolescents who accumulate higher levels of MVPA may also participate in activity settings where resistance-type exercises are available or encouraged [2]. Previous YRBS-based analyses show that muscle-strengthening exercise frequently co-occurs with vigorous activity within broader activity profiles [24]. Large school-based studies also report that associations between MSE and weight status differ depending on whether adolescents meet general physical activity guidelines, suggesting that these behaviours often co-exist within an overall activity pattern [16]. Our partial dependence results similarly indicate that higher MVPA levels are associated with a higher predicted probability of meeting the MSE guideline. Taken together, MVPA in this model likely reflects a broader active lifestyle context rather than an independent behavioural driver of MSE participation [19].

4.2.2 Sex.

Sex was a dominant correlate in our models, with boys showing a consistently higher probability of meeting the MSE guideline. This pattern aligns with prior evidence that male adolescents report more frequent MSE and related muscle-oriented behaviours than females [18,19]. Analyses using U.S. YRBS data also indicate that boys are more likely than girls to meet MSE recommendations and to show joint adherence with aerobic activity [19].

Several mechanisms may explain this gap. Strength and muscularity are often framed as masculine ideals, which can increase encouragement and social acceptance for boys to engage in resistance-type activity [18]. These norms may be reinforced in sport and school settings where strength training is more visible or more strongly promoted for male participants [18,19]. In contrast, girls may face fewer invitations, fewer role models, and less perceived belonging in strength-focused spaces, which can reduce uptake even when opportunities exist [18]. Psychological barriers may also differ by sex. Boys may enter adolescence with greater familiarity and confidence in strength-oriented activities, while girls may be more likely to anticipate negative evaluation, feel discomfort in mixed-gender environments, or worry about appearance-related consequences of training [18]. Taken together, the sex gradient in MSE participation is more plausibly linked to modifiable social and contextual factors than to biological constraints [18].

4.2.3 Dietary indicators (fruit and salad intake).

Dietary indicators, particularly fruit and salad intake, were prominent predictors in the model. Higher intake was associated with a higher predicted probability of meeting the MSE guideline. This pattern may reflect health behaviour clustering, rather than a direct dietary effect. Adolescents who report exercise-based weight control often report higher fruit and vegetable intake, consistent with a broader health-oriented behavioural profile [24]. However, this behavioral clustering may also be influenced by unmeasured or residual confounding, such as household income or neighborhood environments, which jointly shape both diet and physical activity.

A nutritional pathway is also possible but should be interpreted cautiously. Diet and physical activity are jointly related to muscle-related outcomes in youth [23], and experimental evidence suggests that fruit-derived compounds may help reduce exercise-induced muscle damage and inflammation [33]. Taken together, higher fruit and salad intake may act as a marker of better overall diet quality that supports sustained engagement in strengthening activity [23,33]. However, given the cross-sectional nature of the present data, these dietary variables are better interpreted as markers of a healthier lifestyle pattern that co-occurs with strengthening activity rather than as causal determinants of MSE participation.

4.2.4 Attempts to lose weight.

Attempts to lose weight emerged as a key predictor of meeting the MSE guideline, indicating that weight-control motivation is associated with adolescents’ strengthening behaviour. Previous YRBS analyses have similarly reported that adolescents trying to lose weight are more likely to report muscle-strengthening activity [24].

Several interpretations are possible. First, MSE can be perceived as a practical strategy for changing body shape (e.g., “toning”) and increasing muscle definition, which may appeal to adolescents who are actively trying to manage weight [24]. Second, weight-related goals often co-occur with broader body-change motives, including efforts to increase muscle size or tone, and such motives are common in both boys and girls [18]. Third, there is a plausible physiological pathway: resistance training can improve muscular strength and lean mass and may support favourable body-composition and cardiometabolic profiles among youth with overweight or obesity [15,34]. This association may also reflect reverse causality and shared confounding. Adolescents with higher adiposity or muscularity-motivated may be more likely to both attempt weight loss and engage in muscle-focused training, and weight status has been linked with higher use of muscle-enhancing behaviours [18]. However, this association may also reflect reverse causation or shared underlying factors, such as body image concerns or higher adiposity, which may increase both weight-loss attempts and engagement in strengthening activity [18]. Therefore, the observed relationship should be interpreted as an association rather than evidence of a directional behavioural pathway.

4.2.5 Mental health indicators.

In our model, indicators of psychological distress (e.g., low mood and suicidality) contributed meaningfully to predicting MSE guideline adherence. This aligns with evidence that adolescents who engage in MSE, especially when combined with aerobic activity, tend to report better mental health profiles.

The direction is plausibly bidirectional. On one hand, poorer mental health may reduce motivation, self-efficacy, and social participation, which can lower engagement in structured exercise contexts where MSE occurs. On the other hand, higher MSE participation is consistently associated with fewer depressive symptoms and better positive mental health outcomes in adolescent studies, suggesting potential mental-health benefits of MSE [8,17]. Taken together, the model output likely reflects that MSE sits within a broader “functioning and coping” profile, where psychological distress co-occurs with lower activity participation, while better mental health co-occurs with higher participation [7]. Given the cross-sectional design, the present findings should be interpreted as reflecting co-occurring behavioural and psychosocial patterns, rather than evidence that MSE directly improves mental health.

4.2.6 Other contributing correlates.

Beyond MVPA, sex, weight-control attempts, dietary indicators, and mental health, the ML outputs suggested additional, smaller contributors: age, sleep, breakfast, sugar-sweetened beverages (soda/fruit juice), alcohol use, potato intake, family support, and school belonging.

Age showed a modest contribution, consistent with evidence that MSE participation tends to decline across adolescence and that older youth can be an at-risk group for low engagement [10]. Sleep and daily routines (sleep, breakfast) likely reflect broader lifestyle patterning rather than isolated drivers. In a large adolescent survey, longer sleep duration was associated with a higher likelihood of muscle-strengthening physical activity and a lower likelihood of breakfast skipping, alongside healthier dietary behaviours [22].

Unhealthy dietary markers (soda/fruit juice, potato) and alcohol use may capture clustering of behaviours that co-occur with lower (or less sustained) structured exercise. For example, sleep duration in adolescents has been linked with multiple health-related behaviours, including alcohol use, suggesting these behaviours form interrelated profiles rather than independent exposures [22]. Family support and school belonging likely represent access, norms, and encouragement that make MSE more feasible. Reviews highlight parental support as a key predictor of youth compliance with muscle-strengthening activities through encouragement, logistical support, and co-participation [12]. Schools are also a central setting for MSE opportunities via PE curriculum, staff, and infrastructure, and gendered access to spaces may matter [11].

Race and ethnicity also contributed modestly to the prediction of meeting the MSE guideline. Descriptive results indicated differences across racial and ethnic groups. In population health research, race and ethnicity are typically interpreted as social indicators rather than biological determinants. Differences in physical activity behaviors may reflect broader contextual conditions, including socioeconomic resources, access to sports facilities, school physical education opportunities, and neighborhood environments [10,12]. These contextual factors may shape adolescents’ opportunities and support for participating in muscle-strengthening activities. Therefore, the observed differences in this study are more likely to reflect social and environmental contexts than inherent individual characteristics.

4.4.1 Strengths.

This study has several strengths. It used a large sample of U.S. high-school students, providing broad coverage of U.S. high school students within the YRBS sampling frame. The availability of demographic, behavioural, dietary, and psychosocial variables enabled a broad assessment of correlates of meeting the MSE guideline. The machine-learning approach, paired with explainability tools (SHAP and partial dependence plots), supported modelling of non-linear patterns and interactions and provided interpretable summaries of the predictors most relevant to classification. Together, these features strengthen confidence in the robustness of the observed correlate patterns.

4.4.2 Limitations.

Several limitations should be acknowledged. First, YRBS measures are self-reported, which may introduce recall and social desirability bias, particularly for physical activity, dietary behaviours, and MSE. Second, a significant proportion of participants had missing data for the outcome variable. Although we employed multiple imputation to mitigate bias, our reliance on a single imputed dataset to facilitate SHAP value calculation means that the variance between imputations was not accounted for, which may underestimate the standard errors. Third, the cross-sectional design precludes causal inference and limits interpretation of temporal ordering. Fourth, complex survey weights were not incorporated into the machine learning pipeline; thus, while the findings robustly rank correlates at the individual level, they may lack absolute representativeness for estimating exact national population parameters. Finally, while machine learning excels at capturing non-linear patterns, it remains an observational technique. SHAP values represent predictive contributions rather than causal effects, and the risk of residual confounding cannot be eliminated. Therefore, ML-derived importance and explainability outputs describe how predictors contribute to classification within this dataset; contributions may be shared across correlated variables, and external validation is needed to assess transportability.