Drawing on the theoretical framework of psychological resilience, this study uses a “theory-first, variable-grounded” modeling method. It considers the mental toughness of vocational school students to be a form of adaptation that develops naturally through a developmental setting and stress. This level of resilience is not determined solely by internal personal factors but is shaped together by the school environment, academic demands, and external support systems of school. The selection predictive variables follows two principles: first, prioritizing the core psychological mechanisms that have been consistently validated in the resilience literature as closely associated with adaptive outcomes; and second, emphasizing the unique context of vocational students, including occupational learning workload, internship pressure, and employment transition stress, while also considering theoretical explanatory ability and relevance to the context. Based on this rationale, this study will combine the candidate predictors into four theoretical dimensions and form an operational hypothesis model (Figure 1) to explain and predict the level of resilience among students.

Campus belonging and participation dimensions express the quality of connections between people and their schools (Wright et al., 2022). Being part of a social group and forming peer bonds is central to socio-emotional development, providing a feeling of security and self-worth, reducing perceived threats, and thus promoting active engagement. On the contrary, a lack of belonging often leads to social withdrawal and limited resource consumption, making it easier for stress to turn into psychological exhaustion. Self-system resources focus on the internal psychological capital for resilience, such as self-esteem, self-efficacy, and emotional regulation (Xie et al., 2025). Together, these elements determine how individuals understand their skills, control their feelings, and keep trying when faced with challenges, serving as important “cushions” between tension and adaptation. Academic and occupational stressors highlight the significant risks that vocational college students encounter during their learning and professional training, including academic burnout, high-pressure internships, procrastination, and problematic internet usage (Nan, 2025). Such factors undermine individual resilience by maintaining stress and depleting self-regulation resources, increasing the risk of low resilience. Fourth, the support and coping dimension concerns protective resources such as external support systems and coping flexibility. Social support provides comfort, while coping flexibility involves the selection and adjustment of coping methods based on different types of stress (Dong et al., 2024). All these factors influence the routes taken in stress processing and how our final mental health outcomes impact our resilience.

This study aims to develop and validate a theory-driven psychological resilience prediction framework for vocational college students. Guided by the multidimensional psychological predictor system illustrated in Figure 1, it will systematically identify risk and protective factors associated with low psychological resilience. On this basis, multiple prediction models will be constructed and compared using a 70% training set and a 30% test set split strategy. The optimal and most interpretable model will be selected through a comprehensive evaluation of discriminative power, fit, and clinical net benefit. Furthermore, feature importance analysis will be conducted to clarify the core psychological mechanisms and directional effects, providing empirical evidence and transferable tools for the early identification, stratified management, and targeted psychological intervention of low-resilience risks among vocational college students.

This study conducted a cross-sectional questionnaire survey, combined with a “dual-track validation” method to establish and validate the psychological resilience evaluation model for vocational college students. The research process, as seen in Figure 2, first involved sample recruitment and scale administration, which was done in “Recruitment and Survey”. Data were entered into the database, and preprocessing was carried out (missing values and outliers were detected, variables were coded, and normalized). A total of 1,588 valid samples were included at the end. After that, the data were stratified and randomly divided (by gender, grade, and major) into a training set (70%, n = 1,112) and a test set (30%, n = 476). The training set served as the model for developing and validating it, while the testing set was used to validate the model’s ability to extend the range of application and stability.

All preprocessing steps were implemented strictly to prevent data leakage. Variables with missingness < 5% were imputed using the median (for continuous features) or mode (for categorical features) computed exclusively from the training set; variables with ≥5% missing values were excluded from analysis. For models requiring feature scaling (e.g., SVM and MLP), standardization (z-score normalization) was applied using the mean and standard deviation estimated only from the training set, which were then used to transform both the training and test sets. This ensured that no information from the test set influenced any aspect of preprocessing or model development.

Under “dual-track verification,” Track A is aimed at psychometric validation through a series of content validation by experts, cognitive interviews, item analysis, exploratory factor analysis (EFA), confirmatory factor analysis (CFA), and reliability/validity tests (Cronbach’s α, convergent validity, and discriminant validity) to ensure structural validity and measurement reliability. Track B emphasizes predictive validity by constructing and comparing multiple models (logistic regression, XGBoost, random forest, support vector machines, K-nearest neighbors, decision tree) using training data. A comprehensive evaluation is conducted on test data across discriminant power, calibration, and decision benefit metrics, with key predictors identified through SHapley Additive exPlanations (SHAP) analysis. The final output provides risk stratification results for “high/low psychological resilience” (see Figure 2).

The baseline characteristics of the samples are presented in Table 1. No significant differences were observed between the training and test sets in terms of major demographic variables (age, gender, grade, major category, place of origin, whether the child was an only child, family monthly income), academic and internship status (GPA, ranking, internship experience, part-time employment), or core psychological constructs and mental health indicators [psychological resilience, self-efficacy, perceived social support, emotion regulation, coping style, PHQ-9 depression level (Kroenke et al., 2001)] (all p > 0.05). This indicates that the stratified randomization achieved good intergroup comparability, providing a reliable sample foundation for subsequent model training and testing.

This study constructs a candidate predictor variable system according to the “theory-driven psychological resilience prediction framework” in the following aspects: demographics, academic performance, career and internship, psychological resources, mental health indicators, and health and social behavior, to describe the psychological resilience status of vocational college students from different dimensions. The operational definitions, measurement tools, coding methods, and descriptive statistics for each variable are summarized in Table 2 and serve as the unified input specification for subsequent modeling.

Demographic and academic baseline information (e.g., age, gender, grade level, place of origin, family income, and GPA) were primarily obtained through self-report or academic records and coded in continuous, ordered, or binary forms. Variables related to academic and occupational stress (e.g., academic pressure, learning motivation, employment anxiety, and career clarity) were scored using multi-item scales, with total scores reflecting individual levels (Jayaraj et al., 2025). Psychological resource variables (e.g., self-efficacy, perceived social support, optimism, emotion regulation, problem-focused coping, and self-compassion) were measured using established scales with proven reliability and validity (Lei et al., 2025). Mental health indicators (e.g., depression, anxiety, perceived stress scale, academic burnout, and life satisfaction) were used to characterize the coexistence of negative and positive psychological states. Health and social behavior variables (e.g., sleep quality, physical activity, peer relationship quality, family cohesion, romantic relationships) were employed to supplement risk and protective factors at the lifestyle and social connection levels. Overall, Table 2 demonstrates clear value ranges and interpretable distribution characteristics (providing skewness and kurtosis information), laying a reproducible foundation for variable construction in subsequent model training, comparison, and interpretive analysis.

The initial item pool for the psychological resilience scale was developed through an extensive review of existing literature and aligned with the theoretical framework described in Section 2.1. To ensure content validity, a panel of five experts—including two professors specializing in psychology (with expertise in positive psychology and psychometrics), two frontline mental health practitioners from vocational colleges, and one associate professor in vocational education—reviewed the draft containing 25 initial items. Experts independently evaluated each item’s relevance, clarity, and conciseness using a 4-point scale (1 = not relevant, 4 = highly relevant). The Scale-Level Content Validity Index (S-CVI/Ave) was calculated as 0.92, indicating strong overall content validity. Items with an Item-Level CVI (I-CVI) below the recommended threshold of 0.78 were either revised or removed based on both quantitative ratings and qualitative feedback from the expert panel.

Following expert review, cognitive interviews were conducted with a purposive sample of 12 vocational college students (6 male, 6 female; representing different academic years and majors) to examine how respondents interpreted and processed the items. Using a retrospective verbal probing approach, participants completed the draft scale and were then asked to explain their understanding of each item and the rationale for their responses. Audio-recorded interviews were transcribed and analyzed thematically (Nowlin et al., 2025). Findings revealed that three items were consistently misinterpreted or perceived as ambiguous (e.g., Item 7 was frequently confused with emotional suppression rather than adaptive coping). These items were reworded to improve clarity, reduce jargon, and better reflect the intended construct. After two rounds of refinement based on cognitive interview feedback, the final 18-item scale was completed for subsequent psychometric validation.

The primary outcome variable for the predictive modeling in Track B was psychological resilience, operationalized as a continuous total score derived from the newly developed 18-item Psychological Resilience Scale (see Section 3.3). This total score was subsequently dichotomized for classification purposes (see Section 5.1).

Furthermore, the three core dimensions of psychological resilience—Tenacity, Strength, and Optimism—were constructed as subscale scores. Each subscale score was calculated as the sum of its corresponding six items, as defined by the final factor structure presented in Table 3.

The other psychological constructs mentioned in the analysis (i.e., self-efficacy, social support, emotion regulation, depression, anxiety, stress, and academic burnout) were not constructed by the authors but were measured directly using established, validated scales. Their total scores, as provided by the respective instruments listed in Table 2, were used as predictor variables in the machine learning models and for correlation analyses in the psychometric validation (Track A). Specifically:

Self-efficacy: Total score from the General Self-Efficacy Scale (GSES).

Social support: Total score from the Multidimensional Scale of Perceived Social Support (MSPSS).

Emotion regulation: Total score from the Emotion Regulation Questionnaire (ERQ).

Depression: Total score from the Patient Health Questionnaire-9 (PHQ-9).

Anxiety: Total score from the Generalized Anxiety Disorder-7 (GAD-7).

Stress: Total score from the Perceived Stress Scale-10 (PSS-10).

Academic burnout: Total score from the Maslach Burnout Inventory–Student Survey (MBI-SS).

In summary, the resilience-related dimensions (Tenacity, Strength, Optimism, and Total Resilience) are composite scores from our new scale, while the other listed constructs are direct total scores from pre-existing, widely used measurement tools.

To validate the structural validity and item quality of the psychological resilience scale in vocational college students, this study first conducted item analysis, followed by exploratory factor analysis (EFA). Prior to EFA, the suitability of the data for factor analysis was examined. The Kaiser–Meyer–Olkin (KMO) measure of sampling adequacy was 0.932, indicating excellent sample adequacy (well above the recommended threshold of 0.80). Bartlett’s test of sphericity was highly significant (χ² = 3876.42, df = 153, p < 0.001), confirming that the correlation matrix was appropriate for factor analysis.

An EFA was performed using principal axis factoring (PAF) as the extraction method. Given the theoretical expectation that the latent factors of psychological resilience are intercorrelated, an oblique rotation (Promax) was applied to allow for correlated factors and enhance interpretability. Initial eigenvalues and the scree plot supported a three-factor solution, which together explained 62.8% of the total variance.

The final factor structure demonstrated strong factorial clarity. Items were retained only if they met the following criteria: (1) primary factor loading ≥ 0.40, and (2) no substantial cross-loadings—defined as a secondary loading < 0.30 and a difference between the highest and second-highest loadings ≥ 0.20. All 18 items satisfied these conditions, with each loading predominantly on one of the three factors and showing negligible cross-loadings on others.

As shown in Table 3, the corrected item-total correlation (CITC) values ranged from 0.52 to 0.78, and all critical ratio (CR) values were statistically significant (p < 0.001). These results indicate that the items possess high discriminative power and internal consistency, effectively differentiating among students with varying levels of psychological resilience.

To further validate the scale’s structural validity, Confirmatory Factor Analysis (CFA) was conducted using the test set derived from the machine learning analysis described in the preceding section. As shown in Table 4, the three-factor model achieved a satisfactory fit (χ²/df = 3.126, RMSEA = 0.041, SRMR = 0.033, CFI = 0.968, TLI = 0.962), demonstrating that the’ resilience–power–optimism’ framework effectively explains the psychological resilience data of vocational college students and supports the structural framework derived from the earlier exploratory factor analysis (EFA).

Building on this foundation, the study further examined the measurement invariance of the scale across gender groups (n = 778 males, n = 810 females). The results demonstrated that the configural invariance model (M1) showed a good fit. In the measurement invariance test (M2) with equal factor loadings, the changes in ΔCFI and ΔRMSEA were minimal (see Table 4), indicating comparable measurement units and structural meanings between genders. The scalar invariance test (M3) showed large χ² differences but small incremental changes in fit indices (ΔCFI ≈ − 0.001, ΔRMSEA ≈ 0.000; see Table 4). These findings together imply that there is high cross-group comparability of the scale, making it appropriate to conduct subsequent comparisons and predictions of students’ psychological resilience levels across different genders.

This study provides evidence for the quality of the scale through two dimensions: internal consistency and construct correlation. First, the three dimensions and total scores all demonstrate good internal consistency. Cronbach’s α and McDonald’s ω for resilience, strength, optimism, and total resilience are all at a high level, as shown in Table 5, indicating that the psychological resilience scale is stable and reliable in measuring vocational college students.

Second, the correlation analysis results support its construct validity. As shown in Table 5, the three dimensions of psychological resilience exhibit significant positive correlations with each other and with the total score, indicating that the three-dimensional structure is both interrelated and collectively reflects overall psychological resilience. Meanwhile, total resilience and its dimensions show significant positive correlations with self-efficacy, social support, and emotional regulation, aligning with the theoretical expectation of “psychological resources promoting resilience.” Conversely, they demonstrate significant negative correlations with depression, anxiety, stress, and academic burnout, reflecting the protective role of psychological resilience against negative mental health outcomes and learning exhaustion. Overall, these systematic correlations in both direction and strength provide robust support for the scale’s convergent validity and criterion-related validity (see Table 5).

Prior to model development, the continuous total score of psychological resilience was dichotomized to define the binary outcome of “low psychological resilience.” To identify students at the greatest risk and most in need of targeted support within our sample, we employed a percentile-based cutoff derived from the empirical distribution of resilience scores in the full cohort. Specifically, participants scoring at or below the 25th percentile (P25 = 59.0) were classified as having “low psychological resilience,” while those above this threshold were categorized as having “normal/high psychological resilience.” This approach aligns with common practice in educational and psychological screening studies that aim to flag the most vulnerable subgroup (typically the lowest quartile) for early intervention.

In the total sample (N = 1,588), the prevalence of the low-resilience group was 25.0% (n = 397). The dataset was then partitioned into training and testing sets using stratified random sampling based on gender, academic year, and major discipline to preserve the original distribution of key covariates. Consequently, the prevalence of low psychological resilience remained stable across subsets: 25.1% (n = 279) in the training set (n = 1,112) and 24.8% (n = 118) in the hold-out test set (n = 476), with no statistically significant difference between groups (χ² = 0.04, p = 0.841). This balanced and stable class distribution ensures robust model training and valid performance evaluation, minimizing bias due to data splitting.

Using this well-defined binary outcome, we developed and compared multiple machine learning models—including logistic regression, random forest, support vector machine, and XGBoost—on the training set (70% of the total sample). Hyperparameter tuning was performed via five-fold cross-validation within the training set to optimize predictive performance while mitigating overfitting.

To avoid optimistic bias in performance estimation, all model hyperparameters were optimized using a nested cross-validation framework within the training set. Specifically, an outer five-fold cross-validation loop was used for unbiased performance evaluation, while an inner stratified five-fold cross-validation loop was employed for hyperparameter selection. We adopted Bayesian optimization to efficiently search over predefined hyperparameter spaces. For example, the search space for XGBoost included: n_estimators ∈ [100, 500], max_depth ∈ {3, 4,…, 10}, learning_rate ∈ [0.01, 0.3], and subsample ∈ [0.6, 1.0]. The final model was retrained on the full training set using the best hyperparameters identified by the inner loop and evaluated on the held-out test set.

In the independent test set (30%), all candidate models were subjected to consistent evaluation and cross-reference (Table 6). XGBoost had the best overall performance, achieving an AUC of 0.883 (95% CI: 0.854–0.908), an accuracy of 0.826, an F1-score of 0.800, and a balanced accuracy of 0.825. It also had the lowest Brier score of 0.116, demonstrating the best balance between distinguishing well and maintaining good probability calibration (Table 6). Random forest and neural network followed with AUCs of 0.871 and 0.864, respectively, while support vector machine and logistic regression performed slightly lower. Decision tree and k-nearest neighbors exhibited relatively weaker performance (Table 6).

In terms of robustness, AUC distributions from the 100 bootstrap resampling trials show that both XGBoost and Random Forest have consistently high and stable AUCs across multiple samplings (Figure 3), indicating they are better at generalizing.

Calibration level refers to the discrepancies between predicted probabilities and actual incidence rates from the models based on calibration curves. Overall, the calibration of the integrated model is closer to the ideal diagonal than some models in the medium-to-high risk range (Figure 4).

The multi-criteria heat map of the test set clearly summarizes the overall rankings of all models in terms of various AUC, accuracy, sensitivity, specificity, and F1 scores, allowing for direct identification of trade-offs such as high discrimination but mediocre calibration, or good calibration but insufficient discrimination (Figure 5). Combining all three: Discrimination, Calibration, and overall Classification Quality, XGBoost emerges as the primary model for further interpretation and application.

To improve the interpretability and auditability of the model’s decision-making process, this study reports the final selected model’s classification results and the distribution of error types in the test set (Table 7). From the confusion matrix, we can see that the model correctly identifies the majority of the TP (97) as low resilience students, with stable performance in differentiating normal/high-resilience students, resulting in TN = 299, indicating a balanced discriminatory ability for all samples (Table 7). On a metric level, the model achieved a sensitivity of 0.815 and a specificity of 0.837, indicating it can “identify high-risk students” without misclassifying normal students as low resilience (Table 7).

The model shows a high NPV (0.931), making the classification of normal/high resilience more trustworthy. On the contrary, the positive predictive value (PPV = 0.626) indicates that some “low resilience” samples have been misclassified as “low resilience,” necessitating secondary screening or stratified intervention in practice to avoid wasting resources (Table 7). The model attained an overall accuracy of 0.832, an F1-score of 0.708, and an MCC of 0.614, performing well in risk screening while remaining robust to class imbalance (Table 7).

To explain why the model made certain decisions, this study used SHAP to decompose the contributions of key predictors and their magnitude of impact (Figure 6), and stability checks were conducted separately on the training and test data. Overall, we observed extremely high agreement regarding the importances of the two datasets, suggesting that our model is not picking up on random patterns in the data and is instead identifying particular psychological and behavioral risk signals.

From the direction of feature contributions, sleep quality (PSQI), chronic disease, perceived stress, and depressive symptoms all contribute more to predictions of “low psychological resilience” when at higher levels, indicating the ongoing detrimental effects of physical and mental loads and negative emotions on resilience (Rodríguez-Planas et al., 2025). Conversely, resource-related variables such as social support and adherence to mindfulness training show a protective role in reducing the low-resilience risk at higher levels, reflecting the protective effect under the stress–resource perspective. Factors including online gaming addiction, academic procrastination, and self-esteem demonstrate both behavioral moderation and self-system resource explanations, indicating conceptual compatibility between the model’s outputs and the theoretical framework of the study. This facilitates the subsequent translation of predictive results into actionable, stratified intervention targets.

To evaluate the model’s effectiveness in real-world student support decisions, this study employed Decision Curve Analysis (DCA) to compare net benefits across different risk thresholds (Figure 7). Overall, within a broad threshold range, the ensemble learning models (particularly XGBoost and Random Forest) consistently outperformed the ‘treat none’ and ‘treat all’ strategies, demonstrating higher net benefits than other models. This indicates their superior practical utility in the task of identifying students with low psychological resilience at high risk and initiating support services.

From an implementation perspective, DCA results indicate that schools can select models and trigger criteria with greater net benefits based on different risk thresholds when implementing tiered interventions (e.g., psychological screening reviews, counseling appointment prioritization, targeted sleep and stress management courses) under limited resources. This approach reduces ineffective interventions while improving coverage efficiency for genuinely high-risk students, thereby transforming predictive outputs into actionable student support strategies.