Measurement-Informed Machine Learning for Generalizable Suicide Risk Prediction in Youth

About This Grant

ABSTRACT Suicide is the second leading cause of death among adolescents aged 10 to 19 years, with increasing incidence rates that underscore the pressing need for improved early detection methodologies. Existing machine learning models frequently achieve acceptable accuracy at the population level; however, they are inadequate in reliably predicting suicidal thoughts and behaviors (STBs) across critical developmental and contextual subgroups, where biopsychosocial predictors may exhibit measurement non-invariance, potentially leading to misclassification and missed intervention opportunities. Preliminary findings from the Adolescent Brain Cognitive Development (ABCD) Study suggest that neurobiological predictors alone offer limited predictive accuracy and display variability across relevant subpopulations, thereby emphasizing the necessity of adopting a comprehensive biopsychosocial framework. This K23 proposal aims to develop a transparent, interpretable, generalizable, and developmentally sensitive machine learning model of suicide risk in youth, incorporating neurobiological, psychological, electronic health record (EHR)-derived, and social indicators. Utilizing ABCD study data from ages 9 to 18, the study will assess and address measurement non-invariance in biopsychosocial predictors across subgroups (e.g., sex, socioeconomic status, developmental stage, region, race/ethnicity) through Moderated Nonlinear Factor Analysis (Aim 1). We will then use longitudinal data extending to Year 8 to predict future STBs via interpretable machine learning techniques (i.e., Random Forest, Elastic Net, Support Vector Machines, and Extreme Gradient Boosting) (Aim 2). Finally, we will enhance model calibration and accuracy within relevant subpopulations through targeted refinements (e.g., Trans-Balance, Platt scaling, and stratified modeling), with external validation performed in the National Consortium on Alcohol and Neurodevelopment in Adolescence-Adulthood cohort (Aim 3). We will use principled approaches, including Bayesian Longitudinal Imputation via Multivariate Probabilities, Multiple Imputation by Chained Equations, and delta-adjusted sensitivity analyses to account for potential non-random missingness. All models will be constructed employing nested and forward-chaining cross-validation, adhering to preregistered protocols aligned with TRIPOD and PROBAST guidelines. To bolster clinical applicability, model performance will be evaluated across subgroups using calibration metrics (e.g., Brier score, slope) and discrimination metrics (e.g., AUC). Findings will inform translational suicide risk assessment for research and clinical settings. With institutional access to EPIC Cosmos, a federated EHR network spanning over 180 million patients, this K23 provides a scalable platform for future R01 work focused on real-world EHR-based implementation.