KMAN Publication Inc. (KMANPUB) Journal of Psychosociological Research in Family and Culture 3041-8550 In Press 2026 07 01 Predicting Suicidal Ideation Among Youth Using Family, Cultural, School, and Digital-Ecology Variables: A Stacked Ensemble Model with Calibration, Explainability, and Subgroup Fairness Assessment Predicting Suicidal Ideation Among Youth Using Family, Cultural, School, and Digital-Ecology Variables: A Stacked Ensemble Model with Calibration, Explainability, and Subgroup Fairness Assessment 1 23 EN Journal Article 2025 12 29 <table> <tbody> <tr> <td> <p><strong>Objective: </strong>This study aimed to develop and independently evaluate a calibrated and explainable stacked ensemble model for predicting clinically significant suicidal ideation among youth in Taiwan using family, cultural, school, and digital-ecology variables.</p> <p><strong>Methods and Materials: </strong>A cross-sectional predictive modeling study was conducted with 4,812 Taiwanese students aged 15–24 years. Participants were recruited from secondary schools, vocational schools, junior colleges, and universities through multistage stratified cluster sampling. Suicidal ideation was assessed using the Suicidal Ideation Attributes Scale. Candidate predictors included family functioning, family support, family conflict, filial-piety orientations, loss-of-face concerns, mental-health stigma, school belonging, academic stress, bullying, social-media disorder symptoms, cyberbullying, nighttime digital use, exposure to self-harm content, and digital resilience. Data were divided into development, calibration, and independent holdout samples. Elastic-net regression, support vector machines, random forests, gradient boosting, and extreme gradient boosting were combined using a regularized logistic meta-learner. Model performance was examined through discrimination, calibration, explainability, decision-curve analysis, and subgroup fairness assessment.</p> <p><strong>Findings: </strong>The calibrated stacked ensemble outperformed the conventional logistic-regression benchmark and all individual learners. In the independent holdout sample, the model achieved an area under the receiver operating characteristic curve of .889, an area under the precision–recall curve of .682, sensitivity of .861, specificity of .812, positive predictive value of .549, negative predictive value of .957, and Brier score of .115. Calibration was strong, with an intercept of 0.02 and slope of .98. School belonging, family functioning, cyberbullying, academic stress, family conflict, and exposure to self-harm-related digital content were the most influential predictors. Subgroup analyses showed generally comparable sensitivity, although specificity and calibration were weaker among several disadvantaged and smaller subgroups.</p> <p><strong>Conclusion: </strong>A calibrated stacked ensemble integrating multiple ecological domains can provide accurate and interpretable estimates of youth suicidal ideation, although its predictions should support rather than replace direct professional assessment and must be accompanied by ongoing fairness monitoring.</p> </td> </tr> </tbody> </table> https://www.journals.kmanpub.com/index.php/jprfc/article/download/6268/11736