Predicting Short-Term Risk for Suicidal Behavior after Contact with Outpatient Specialists: A Machine Learning Approach

点击翻译按钮获取中文摘要

Project Summary/Abstract Because suicidal behavior (i.e., suicide and suicide attempt) significantly contributes to global disease and financial burden, the National Action Alliance for Suicide Prevention, the World Health Organization, and prominent researchers have called for the need to develop and improve prediction models for suicidal behavior. The period after contact with a health care professional is a particularly high-risk period, suggesting that prediction of suicidal behavior in the short-term after contact (i.e., defined as within one year in most studies) would address a critical need in the field to aid intervention efforts. However, previous short-term research has been limited by numerous factors, including a lack of research on youth samples, an overreliance on self-report measures or electronic health records, and the frequent examination of bivariate associations between only one predictor and suicidal behavior. The overall objective of the current proposal is to utilize several algorithms and assess their relative performance in the prediction of short-term suicidal behavior after contact with an outpatient specialist (defined as within 1, 6, and 12 months) using an unparalleled dataset. I will use data from a prospective, large-scale register of all outpatient mental health specialist visits among youth in Stockholm County, Sweden, consisting of approximately 160,000 visits by the onset of the current award. These individuals can be linked to population-based registers assessing a broad range of information (e.g., medical problems, academic information, neighborhood factors, parental psychopathology), which is a significant advantage over prior literature primarily studying demographic and psychiatric predictors in isolation. The central hypothesis is that the examination of numerous predictors within a large sample and the use of advanced statistical methods will significantly improve upon previous suicidal behavior prediction, which has remained slightly above chance. To achieve the overall objective, the current proposal is designed to apply machine learning algorithms through two specific aims: Aim 1: Apply variable selection algorithms that determine a limited number of salient predictors and, therefore, maximize interpretability; Aim 2: Apply ensemble algorithms that aggregate machine learning models and, therefore, maximize predictive power. The current proposal will significantly contribute to the field by examining short-term risk using machine learning techniques among a youth, outpatient sample, including varying follow-up windows and predictors across domains. Finally, the results from the current proposal will have positive impact by informing both 1) basic research through the identification of at-risk subgroups based on numerous predictors, and 2) the creation of a prediction tool that will aid in clinical practice.