Related Experiment Videos
Suicide risk prediction and analysis of associated factors among high school students in four provinces of China
1Guangxi Zhuang Autonomous Region Center for Disease Control and Prevention (Guangxi Zhuang Autonomous Region Academy of Preventive Medicine), Guangxi Key Laboratory for the Prevention and Control of Viral Hepatitis, Nanning, Guangxi, China.
Background:
Suicidal ideation (SI) among adolescents is a severe global public health issue; however, accurate identification and mechanistic explanation of its multi-dimensional associated factors remain insufficient. This study aimed to construct a machine learning framework that balances predictive capability with transparent interpretability to identify the key factors associated with SI among high school students in four Chinese provinces and to evaluate its potential utility as a screening tool.
Methods:
A cross-sectional survey design was employed, including 6222 high school students from Hainan, Hubei, Jiangxi, and Chongqing in China. Feature selection was performed using LASSO regression and the SVM-RFE algorithm; the intersection of features selected by both methods yielded 11 core predictive factors. Subsequently, the predictive performance of five classification models-logistic regression (LR), decision tree (DT), random forest (RF), support vector machine (SVM), and artificial neural network (ANN)-was compared using DeLong's test. Finally, the SHAP framework was applied to conduct global and local explanation analyses of the best-performing model.
Results:
The prevalence rate of SI was 30.6%, with significantly higher rates among females (34.2%) and students from rural areas (35.9%) (P < 0.001). The RF model achieved an accuracy of 0.716, an AUC of 0.658 (95% CI: 0.629-0.687), a sensitivity of 0.186, and a specificity of 0.950. DeLong's test indicated no statistically significant difference in AUC between RF and LR (P > 0.05). SHAP analysis revealed that bullying experience was the most important predictor associated with SI, followed by frequency of parental conflict and parental relationship quality.
Conclusion:
The interpretable machine learning model constructed in this study, with a specificity of 0.950, may serve as a potentially useful tool for screening out low-risk individuals in school settings, although its low sensitivity (0.186) limits its utility for identifying high-risk students. The results suggest that public health interventions should prioritize integrated traditional and online anti-bullying mechanisms, alongside restoration of family functioning, to reduce SI risk among high school students.