A total of 2115 students from six Chinese universities in central and south China participated in the study. All data were collected offline using self-report questionnaires. The questionnaires were distributed in university classrooms by graduate students and psychology professors, accompanied by identical verbal and written instructions. After excluding questionnaires with non-standard responses (e.g., implausible age values and missing critical data such as demographics), 1,635 valid questionnaires were retained for analysis. The sample consisted of 558 males (34.12%) and 1,077 females (65.88%) aged 17–24 years (M_age = 18.93 years, SD = 1.23 years).

Descriptive statistical analyses were first conducted using SPSS 26.0. ML modeling was then performed using the Scikit-learn library in Python (53). To enhance the generalization ability of the models, data preprocessing was conducted using the StandardScaler method, which normalized all features to have a mean of zero and a standard deviation of one. This approach reduced data bias arising from differences in measurement scales. The outcome variable was whether participants reported mild to severe depressive symptoms. We used an independent-samples t-test and a chi-square test for feature selection. Thirty-eight variables associated with depressive symptoms were selected. These thirty-eight variables were used as predictors, including sociodemographic characteristics, emotional disorder states, and coping styles.

Four ML algorithms were employed to construct separate risk prediction models: Random Forest, eXtreme Gradient Boosting (XGBoost), the Light Gradient Boosting Machine (LightGBM), and Support Vector Machine (SVM) (54). Each algorithm was selected for its ability to handle imbalanced data and enhance model generalizability. Random Forest improves model stability and predictive accuracy by aggregating multiple decision trees and randomly selecting feature subsets at each split, thereby reducing the risk of overfitting (55). XGBoost, based on gradient-boosted decision trees, optimizes its objective function with regularization terms, demonstrating high efficiency and accuracy in handling complex datasets (56). LightGBM, similar to XGBoost, adopts a leaf-wise splitting strategy to improve computational efficiency, although it may be more prone to overfitting in certain scenarios (57). SVM identifies the optimal hyperplane for classification and demonstrates strong generalization capabilities, particularly in high-dimensional and small-sample datasets (58). To evaluate model stability and reliability, this study employed 10-fold cross-validation to minimize sample bias (59, 60). For each fold following data partitioning into training and test subsets, SMOTE was employed on the training data. The SMOTE object, initialized with a random state, used the “fit_resample” method to synthesize minority class instances, creating a balanced training set. This augmented dataset was subsequently used to train the Random Forest classifier (61). By generating synthetic samples for underrepresented classes, SMOTE enhances classifier efficacy, especially in classification tasks (62). The models were then trained on each cross-validation training set and evaluated on the corresponding validation set. Multiple performance metrics were calculated (63), including precision, F1-score, accuracy, recall, area under the receiver operating characteristic (ROC) curve (AUC), and the average of each metric across all folds.

To identify the most critical features for depression prediction, this study analyzed feature importance within the ML models and visualized the results using horizontal bar charts. In addition, SHapley Additive exPlanations (SHAP) values were used as a feature importance metric to interpret model predictions (64). As model complexity increases, particularly in ensemble and deep learning models, prediction accuracy tends to improve while interpretability declines. SHAP values, derived from Shapley’s value theory in game theory, quantify feature importance by computing the average marginal contribution of each feature to a given prediction. These values satisfy key properties such as fairness, uniqueness, and efficiency and provide both baseline and individual feature contributions for each prediction. The sum of the feature contributions equals the difference between the model output and the baseline, allowing users to understand the model’s logic and the actual impact of each input variable (65). Accordingly, this study calculated and visualized SHAP values for all models.

The study analyzed 1,635 valid responses, with 827 participants (50.5%) meeting criteria for depressive symptoms. Table 1 presents descriptive statistics and general characteristics of the study variables. Statistical analyses revealed significant differences in nearly all measured variables between participants exhibiting depressive symptoms and those without, except for gender, only child status, and family structure (see Tables 1 , 2 ).

The ROC curve is a graphical tool used to evaluate the performance of binary classifiers. It illustrates classifier performance across all possible thresholds, allowing assessment of the model’s ability to distinguish between classes regardless of a specific cutoff point. A curve closer to the top-left of the chart indicates better classification performance. A higher AUC value reflects a stronger separation between positive and negative classes. Figure 1 presents the ROC curves for all four models, each achieving an AUC above 85%, demonstrating strong predictive performance. Figure 2 illustrates the performance comparison of the four models across each fold, providing a comprehensive view of their respective metrics.

Table 3 presents the AUC score, accuracy, precision, and specificity of each ML model across all folds of cross-validation. The Random Forest model outperformed the others across all metrics.

To visualize the impact of each variable on depressive symptoms more intuitively, we employed the built-in feature importance method to interpret the variable importance in the best-performing Random Forest model. The resulting feature importance plot ( Figure 3 ) displays the top 20 most important variables. The most important predictor was sleep disturbance (importance score = 0.17), followed by perceived stress (importance score = 0.11) and experiential avoidance (importance score = 0.09).

To further analyze the contribution of each feature to depressive symptom prediction, SHAP analysis was conducted on the best-performing Random Forest model, with the results shown in Figure 4 . In this visualization, higher SHAP values indicate greater impact on the model’s predictions. Each point represents a subject’s feature value, with color indicating value level (red: higher, bule: lower). Points are stacked vertically to indicate data density. Features with high positive SHAP values substantially increase the predicted probability of depressive symptoms, while features with negative SHAP values may reduce it. This visualization facilitates the interpretation of both the individual contribution and interactive role of each predictor within the model. Sleep disturbance had the greatest influence, followed by perceived stress, experiential avoidance, self-criticism, and frustration of basic psychological needs, with predictive influence progressively declining thereafter.

Given the high predictive value of sleep disturbance shown in both the feature importance and SHAP value plots, we conducted a SHAP scatter plot analysis to further examine this relationship. Figure 5 illustrates the association between the actual values of the sleep disturbance feature and their corresponding SHAP values. The plot shows that as sleep disturbance scores increase, SHAP values rise steadily within the range of 0 to 6. This suggests that greater sleep disturbance contributes positively to the model’s prediction of depressive symptoms. Between values of 6 and 12, the upward trend continues but becomes more variable, indicating possible instability or the influence of additional interacting factors at higher levels of sleep disturbance. Overall, sleep disturbance exerts a consistently positive effect on model predictions across its range, reinforcing its role as a key predictor of depressive symptoms.

In the current research, we developed and compared multiple ML models incorporating 38 psychosocial and demographic predictors to predict depressive symptoms among higher education students in China. Among the tested algorithms, the Random Forest model demonstrated superior performance, outperforming XGBoost, LightGBM, and SVM in terms of accuracy (0.7908), precision (0.7878), recall (0.8041), F1-score (0.7956), and AUC (0.8704). These results align with previous research (22), indicating that the model achieves high predictive accuracy and effectively identifies college students at elevated risk of depressive symptoms.

The superiority of the Random Forest model can be attributed to its unique learning mechanisms, which are well-suited to the complexity of the data used in this study. As an ensemble learning method, Random Forest enhances model stability and predictive accuracy by constructing multiple decision trees and aggregating their predictions through majority voting or averaging. This ensemble approach substantially reduces the risk of overfitting, a common limitation in traditional single-tree models. The dataset in this study included 38 predictive variables. Random Forest’s capacity to handle high-dimensional data and model complex feature interactions makes it particularly suitable for predicting depressive symptoms, which are influenced by numerous interrelated factors (66). Additionally, the algorithm introduces randomness during model construction by selecting a random subset of features at each split. This embedded feature selection not only improves computational efficiency but also helps reduce the impact of noisy variables (67).

In this study, although XGBoost and LightGBM are known for their effectiveness in handling imbalanced data and complex datasets, they underperformed compared to the Random Forest model. XGBoost, while efficient in large datasets, is sensitive to hyperparameter tuning and prone to overfitting in noisy data (56, 68). Similarly, LightGBM’s leaf-wise growth strategy improves computational efficiency but may lead to overfitting in smaller or imbalanced datasets (57). Although SVM is recognized for its generalization capability in high-dimensional, small-sample scenarios, it lagged behind Random Forest in terms of accuracy and AUC, likely due to its limited capacity to capture nonlinear relationships among predictors (58, 69).

Previous studies predicting depressive symptoms have rarely used cognitive and emotion-related variables (70, 71), but this study has used these variables, such as self-control, emotion regulation, meaning in life, growth mindset, Self-compassion, experience avoidance, and more. This gives us a glimpse into the mechanisms by which cognitive and emotional factors contribute to depressive symptoms. Moreover, while previous research has often focused on predicting depressive symptoms in children and adolescents (72, 73), this article focuses specifically on the university student population – a group highly susceptible to depressive symptoms – yielding valuable insights.

Model interpretation was performed using the SHAP approach. SHAP is a personalized feature attribution method that strictly satisfies the Consistency axiom. This property is a core strength of its mathematical foundation (Shapley values from game theory). When a feature’s contribution to model prediction increases, its SHAP value must increase (or remain unchanged), ensuring no contradictory attributions (74). Therefore, the interpretability of our machine learning models meets the requirements for consistency. SHAP analysis reveals that predictors spanning multiple domains significantly contribute to forecasting depressive symptoms in Chinese university students, aligning with ecological systems theory (8).

First, the results suggest sleep disturbance emerged as the most influential contributor. This finding corroborates earlier studies that have found a close connection between sleep disturbances and depressive symptoms in university populations (75, 76). Shorter sleep duration exhibits a relationship with a higher risk of developing depressive disorders (77), and inadequate sleep hygiene may contribute to the development of depressive symptoms (78). Underlying these associations is the potential for neural dysregulation. Specifically, inadequate sleep disrupts dopamine activity in the limbic system and striatum, compromising the brain’s reward circuitry and increasing susceptibility to mental illness (79, 80). Additionally, substantial evidence suggests that sleep problems during adolescence can elevate depressive symptoms by impairing both cognitive performance and emotion regulation (81–83).

In addition, this study found that perceived stress was among the top four predictors of depressive symptoms. The present findings resonate with established research. For example, Leung et al. (84) identified perceived stress as a robust predictor of adolescent depression. Parallel results emerged in a Chinese student sample, where greater perceived stress predicted more severe depressive symptoms (85). We can explain it theoretically. Firstly, according to the stress–reward–mentalizing model of depression, the disorder can be conceptualized as a developmental, stress-related condition. When combined with heightened stress sensitivity, elevated stress significantly increases vulnerability to depressive symptoms (86). Additionally, evidence suggests that recent life stress may exacerbate underlying vulnerabilities in HPA axis functioning, with these risk factors potentially interacting synergistically to increase the likelihood of adolescent depression (87).

Experiential avoidance also emerges as a risk factor for depressive symptoms among college students in this study, consistent with findings from Núñez et al. (88). Experiential avoidance manifests as inflexible cognitive-behavioral patterns aimed at suppressing distressing intrapsychic content (89). Research has shown that experiential avoidance is a core transdiagnostic process underlying a range of psychological disorders, including depression, anxiety, and post-traumatic stress disorder (90). College students who engage in experiential avoidance are often unwilling to experience unpleasant emotions or thoughts and may actively suppress or avoid them. Paradoxically, attempts to avoid depressive thoughts can heighten their salience, ultimately intensifying depressive emotional experiences (91).

Self-criticism emerged as a key predictor of depressive symptoms in the current research. Cumulative research findings establish that self-criticism contributes to increased levels of depression and functions as both a risk factor and a perpetuating factor for depressive disorders (92, 93). Blatt (94) theorized that individuals with pronounced self-critical tendencies exhibit “intense feelings of inferiority, guilt, and worthlessness and by a sense that one has failed to live up to expectations and standards”. Research further suggests that individuals who frequently engage in self-critical thought patterns show greater susceptibility to fall into cycles of rumination and self-doubt, perpetuating ongoing emotional distress (95). These cognitive-emotional patterns are strongly associated with heightened vulnerability to depressive symptoms (94). Among university students, self-criticism often arises from the internalized pressure to meet excessively high personal expectations, thereby increasing their susceptibility to depression and related mental health concerns (96).

This study has several limitations. First, the sample size was relatively small; expanding the sample in future research would improve the model’s generalizability. Second, participants were drawn from only six universities located in central and southern China, which may not fully capture regional differences, particularly those between urban and rural areas or between eastern and western regions of the country. For example, disparities in access to mental health resources among rural students may influence the prevalence or severity of depressive symptoms. Third, the current models included only sociodemographic and psychosocial variables, lacking objective physiological indicators such as cortisol concentration or EEG alpha wave power. As demonstrated by Wollenhaupt-Aguiar et al. (97), combining ML with peripheral biomarker measurements can produce more objective and robust predictive features. Future research should consider integrating physiological indicators to further enrich the comprehensiveness and accuracy of depressive symptom prediction among college students.

Although limited in some aspects, this study demonstrates meaningful strengths with important implications. First, it constructed a predictive model based on Bronfenbrenner’s ecosystem theory using ML techniques, incorporating 38 predictors across multiple levels simultaneously, which has stronger explanatory power than traditional regression models. Second, multiple models were used to systematically compare the predictive performance of Random Forest, XGBoost, LightGBM, and SVM for depression symptoms among Chinese university students, overcoming the limitations of single-algorithm research. Finally, through feature importance and SHAP analyses, the factors influencing depressive symptoms among college students were ranked, and several key influencing factors were identified. Analyzing the conditions of college students in these areas can help to promptly identify those at risk of developing depressive symptoms and support more targeted, data-informed interventions, potentially offering a more accurate and efficient approach than standard PHQ-9 screening.