Senior Investigator Kaiser Permanente Washington Health Research Institute Seattle, Washington, United States
Background: Rates of preteen (youths aged 8-12 years) suicide have increased 4-fold in the last 10 years. The American Academy of Pediatrics recommends targeted screening for preteens when clinically indicated; however, the rubric relies on chief complaint, parent or self-report or “warning signs”. Identifying preteens at risk could be significantly improved using risk prediction models. Few studies report on machine learning (ML) models for suicide risk prediction in preteens and their utility in identifying those in need of further assessment and/or safety planning. This study examined whether a ML model trained and validated using data from adolescents performs as well for preteens. Objective: Our objective was to test a suicide risk prediction model for preteens that was previously validated for teens. Design/Methods: We used healthcare data for 287,073 specialty mental health outpatient visits among 59,531 preteen individuals across 7 health systems. The prediction target was 90-day risk of suicide attempt following a specialty mental health visit. We used logistic regression with least absolute shrinkage and selection operator (LASSO) and generalized estimating equations (GEE) to predict risk. Performance of the existing model in a new sample of preteens, without retraining, was compared to performance of the identical model in adolescents using area under the receiver operating curve (AUC). Results: The AUC produced by the existing model in the new sample of preteens making specialty mental health visits was 0.835. Model performance was slightly better in preteens compared to teens where the AUC was 0.796.
Conclusion(s): Prediction models already in operational use by health systems can be reliably employed for identifying preteens in need of further evaluation for suicide risk.
