Explainable Machine Learning Framework for Depression and Suicide Risk Screening in Thai University Students: A Validated Extra Trees-Based Model

Depression poses a public health challenge in Thailand, affecting approximately 1.5 million individuals, yet only 28% receive adequate clinical care. Screening practices within university environments remain inadequate due to three persistent limitations: (1) low screening compliance associated with lengthy 19-item questionnaires, (2) suboptimal predictive performance and only moderate diagnostic accuracy of conventional tools (AUC = 0.897), and (3) the absence of interpretable, AI-enabled decision-support systems. To address these gaps, compact machine-learning models were developed for depression risk, suicide risk, and clinical depression diagnosis using demographic, academic, behavioral, and psychometric data derived from validated Thai instruments (2Q, 9Q, and 8Q). Extra Trees classifiers consistently achieved strong performance, with accuracies of 0.8778, 0.9184, and 0.9633, and AUC values ranging from 0.9599 to 0.9792. In addition, the models demonstrated precision values of 0.7932-0.9206, recall values of 0.7312-0.8628, and F1-scores of 0.7768-0.8526, representing a 7-9% improvement over the 9Q baseline (p < 0.05). SHAP-based interpretability analyses revealed that emotional symptoms, suicidal ideation, fatigue, and negative self-perception-captured directly from the screening questionnaires-were the most influential predictors, whereas demographic variables obtained from the university's online registration system provided secondary contextual information. Importantly, the proposed models outperformed reported benchmarks. Additionally, the compact 8-12-item model structure reduced respondent burden by up to 58%, enabling practical integration into university information systems for automated risk stratification and continuous mental-health surveillance. Overall, these findings demonstrate that culturally adapted, interpretable, and resourceefficient machine-learning models offer a practical solution for early mental-health detection in higher-education settings.