Predicting first onset of suicide attempt among children with suicidal ideation or non-suicidal self-injury using machine learning: a prospective population-based cohort study

While children with suicidal ideation or non-suicidal self-injury (NSSI) are at high risk of suicide, most do not attempt suicide. This study aims to identify predictors of first suicide attempts among children with suicidal thoughts or NSSI. We utilized longitudinal data from the Adolescent Brain and Cognitive Development study to examine a broad range of risk factors in children reporting suicidal ideation (N = 344, mean age = 9.93) or NSSI (N = 261, mean age = 9.88) at baseline, predicting first attempt of suicide during a 4-year follow-up. Predictive models were developed through bivariate feature selection followed by Discrete-Time Random Survival Forest machine learning, identifying predictors of subsequent suicide attempts. During the follow-up, 40 of 344 children (11.6%) with suicidal ideation, and 32 of 261 children (12.3%) with NSSI at baseline initiated suicide attempts. For suicidal ideation, risk for attempting suicide increased with caregiver-reported NSSI, witnessing domestic violence, severity of suicidal thoughts, being female, online social screen use, and less parental supervision. For NSSI, risk of suicide attempt increased with witnessing domestic violence, anxiety disorders, caregiver-reported NSSI, being female, and disruptive, impulse control, and conduct disorders. Our findings shed light on the development of population-based suicide prevention strategies for children.