Modeling suicidal thoughts and behaviors in university students: integrating suicide theory with interpretable machine learning.

Abstract

University students are at risk of suicidal thoughts and behaviors (STB), with prevalence rates higher than those in the general population. Machine learning (ML) enhances the accuracy of STB prediction and provides new possibilities and explanatory frameworks for this issue. This study extends prior research in two ways: (1) by employing, for the first time, a sample of Taiwanese university students and (2) by synthesizing suicide-related theoretical knowledge and empirical findings to identify individual, risk, and protective factors. Based on these elements, a comprehensive ML-based prediction model of STB among university students was developed with four analytic modes to evaluate the relative importance and contributions of three predictor categories. Using data from the NTU Students' Physical and Mental Health Assessment System, a total of 3,901 students were retained in the current study, with 13 self-report assessments administered. Classic logistic regression and five ML classifiers (e.g., bagged LR, RF, LightGBM, linear SVM, and RBF SVM) were utilized. Compared with traditional statistical methods, ML algorithms improved prediction of STB among high‑risk students. RF and LightGBM performed best, with similar results in Mode 2 (risk factor only) and Mode 4 (integrated individual, risk, and protective factors), both outperforming Mode 3 (protective factor only) and Mode 1 (individual factor only). Ultimately, this study enhances the interpretability of ML-based models for predicting STB. Moreover, it provides researchers with deeper insights for developing systematic prediction models and offers a screening tool for identifying university students at high risk of STB.