Optimizing Suicide Prevention in Adolescents: A Longitudinal Approach to Risk Modeling and Resource Allocation

Abstract

Adolescent suicide remains a top public health concern in the United States. From 2018 to 2023, it was the third leading cause of death among youth aged 15–19. During the COVID-19 pandemic, suicide attempts increased significantly among adolescents, especially girls. While clinical screening tools like the ASQ are widely used and effective, many at-risk students do not interact with healthcare systems, leaving their risk undetected. This gap highlights the need for proactive, data-driven strategies that can identify risk in broader populations, such as school-based settings. Existing suicide prediction models often rely on electronic health records and clinical samples, limiting their generalizability. Few models incorporate dynamic behavioral factors or account for missing suicidality data. Moreover, most studies have focused on cross-sectional prediction rather than longitudinal modeling. In contrast, this thesis applies machine learning to four years of school-based survey data (the SURF dataset) to build interpretable, longitudinal risk models for adolescent suicide prevention. We introduce a multi-stage modeling framework that includes (1) imputation of missing suicidality data in earlier years using non-suicide predictors, (2) one-year prediction models using logistic regression and XGBoost, (3) a multi-year compounding ensemble that links predictions across time, and (4) a simulation of resource allocation under budget constraints. We find that suicidality can be accurately imputed and predicted using behavioral and mental health variables, and that risk-based intervention strategies significantly outperform random allocation. These results offer a scalable framework for early identification and support of high-risk students in school-based prevention efforts.