Development of a Machine-Learning-Based Predictive Model for Burnout Among Senior High School Teachers

Teacher burnout is a persistent problem characterized by psychological exhaustion and occupational stress, posing serious constraints on teachers’ professional development, the improvement of teaching quality, and the stable functioning of the educational ecosystem. Conventional questionnaire-based surveys, which dominate teacher burnout research, primarily assess prevalence and bivariate correlations, often at the expense of predictive precision. Such approaches often have limited predictive accuracy and generalizability, making it difficult to identify teachers at risk of burnout with precision. To develop a more effective predictive model for burnout among senior high school teachers and clarify its key underlying mechanisms, this study recruited 2441 in-service teachers from 42 senior high schools in Guangdong Province and the Macao Special Administrative Region of China. The performance of logistic regression, random forest, XGBoost, LightGBM, and CatBoost was compared to identify the optimal predictive framework. Interpretable machine learning methods were then employed to investigate the key factors associated with teacher burnout and the mechanisms through which they operate. The results showed that CatBoost achieved the best overall performance, with an accuracy of 0.815, a specificity of 0.836, and an area under the curve (AUC) of 0.817, the highest among the five models. This superior discriminatory capacity (AUC = 0.817) suggests that gradient-boosted decision trees are more adept than linear classifiers at capturing the multivariate complexity inherent in teacher burnout. Using SHAP (SHapley Additive exPlanations), an interpretable machine-learning technique, we analyzed factors associated with teacher burnout and identified several variables showing significant correlations with burnout among senior-high-school teachers. Depression, neuroticism, surface acting, family–work conflict, and high perceived stress emerged as positive correlates of teacher burnout. By contrast, work engagement, deep acting, psychological need satisfaction, social support, psychological resilience, teacher self-efficacy, high conscientiousness, and openness to experience were negative correlates of teacher burnout. Based on the machine-learning framework, this study constructed an interpretable prediction model and a visual predictive tool for teacher burnout, and unpacked the association patterns between each variable and teacher burnout. The findings of this study can provide robust theoretical evidence and practical support for educational administrative authorities to conduct cross-sectional association screening for teachers with high-burnout profiles and formulate tiered and differentiated psychological intervention strategies.