Machine Learning–Based Prediction of Suicidal Behavior among Unemployed Graduates in Bangladesh: A Data-Driven Public Health Study

Unemployment saliently constitutes severe psychosocial strains that potentially trigger suicidal vulnerability among graduate youths. The sample was recruited using a convenience sampling approach. The current study developed an empirical framework based on 416 jobless graduates in Bangladesh, integrating psychological, demographic and job-related variables for predicting suicidality. Emotional states, insomnia and suicidal ideation were assessed using standardized scales, including DASS-21, ISI-7 and SBQR-4, respectively. Afterwards, seven supervised classifiers, like XGBoost, Random Forest, Decision Tree, Logistic Regression, K-Nearest Neighbours, Naive Bayes, and Support Vector Machine, were implemented with ADASYN-balanced data using stratified 5-fold cross-validation for enhanced generalizability. Furthermore, recursive feature elimination (RFE) was utilized to select the most important features. The XGBoost algorithm outperformed other classifiers with a decent accuracy (69.88%), kappa score (34.97%), ROC-AUC of 80.33% and a balanced sensitivity (72.72%) and specificity (68.85%). The most influential factor was identified as insomnia severity, while depression, anxiety, stress, job loss and job-seeking attempts also attained predictive significance. The findings underline the heightened risk of suicidal attempts driven by multidimensional psychological stress. Moreover, implementing machine learning in a psychosocial monitoring system offers potentiality in early detection and targeted intervention. This approach informed adaptive strategies supporting precision and prevention strategies for suicidal risk in real settings.