Suicide among young people is a pressing public health crisis, with rates rising significantly over the past two decades. Suicide is now the second leading cause of death for this age group, responsible for 6,533 deaths in 2022. Nonfatal suicidal behavior is even more prevalent, leading to significant morbidity and increased risk of suicide. To effectively intervene and ultimately prevent suicide in youth, we need to be able to identify and predict which youths are at greatest risk. While many studies have employed machine learning approaches to predict suicide risk in adults, there has been limited work focused on youth populations. To our knowledge, no studies have specifically applied machine learning for suicide risk prediction in vulnerable populations, such as Medicaid beneficiaries. Furthermore, existing models tend to emphasize clinical factors, often neglecting the critical role of social determinants of health (SDoH) and behavioral health data, which are crucial to developing a more comprehensive and accurate prediction model for youth suicide risk.
This project addresses these gaps by combining machine learning techniques with rich and diverse datasets to develop actionable suicide risk prediction models. The overarching goal is to identify youth at high short-term risk for nonfatal suicide attempts and suicide deaths, focusing on outpatient care settings where most mental health interventions are delivered. The study has three aims: 1) use national Medicaid claims data to develop and validate machine learning models that predict nonfatal suicide attempts and suicide deaths within 90 days of an outpatient visit.2) evaluate model performance across key demographic factors, including age, gender, race, and ethnicity. and 3) conduct focus groups with Medicaid administrators, clinicians, youth/young adults at risk for suicide, and family members to assess the model’s acceptability, perceived usefulness, barriers and facilitators to implementation, and integration into clinical workflows.
We will leverage national Medicaid claims data from all 50 states linked to National Death Index (NDI) data to capture suicide deaths, along with contextual SDoH data from the Area Health Resource File (AHRF), American Community Survey (ACS), and Social Vulnerability Index (SVI). These combined data sources will enable us to examine clinical, social, and systemic risk factors across youth aged 8-25 years who accessed outpatient mental health services between 2010 and 2022 (N = ~13 million). Our analysis will focus on identifying patterns in clinical trajectories, healthcare utilization, and SDoH that predict short-term suicide risk. We will develop and validate four machine learning models to predict nonfatal suicide attempts and suicide deaths, stratified by care setting (mental health specialty vs. primary care). Following best practices, we will engage stakeholders, including Medicaid policy representatives, clinicians, youth/young adults, and families, to ensure the models are actionable and align with real-world needs.
The overall impact of this project will be to allow widespread and effective identification of youth who have the highest short-term risk for non-fatal suicide attempts and suicide and enhance risk mitigation for at-risk youth through claims-based clinical decision support tools to help front-line clinicians make informed choices and provide effective care in outpatient settings
