Using machine learning to classify suicide attempt history among youth in medical care settings

Background The current study aimed to classify recent and lifetime suicide attempt history among youth presenting to medical settings using machine learning (ML) as applied to a behavioral health screen self-report survey. Methods In the current study, 13,325 (mean age = 17.06, SD = 2.61) pediatric primary care patients from rural, semi-urban, and urban areas of Pennsylvania and 12,001 (mean age = 15.79, SD = 1.40) pediatric patients from an urban children's hospital emergency department were included in the analyses. We used two methods of ML (decision trees, random forests) to (a) generate algorithms to classify suicide attempt history, and (b) validate generated algorithms within and across samples to assess model performance. We also employed ridge regression to evaluate performance of the ML approaches. Results Our findings demonstrate that ML approaches did not enhance our ability to classify lifetime or recent suicide attempt history among youth across medical care settings, suggesting that relationships may be mainly linear and non-interactive. In line with prior research, a history of suicide planning, active suicidal ideation, passive suicidal ideation, and nonsuicidal self-injury emerged as relatively important correlates of suicide attempt. Limitations The cross-sectional nature of the current study prevents us from determining the extent to which the important variables identified confer risk for future suicidal behavior. Conclusions The present study underscores the importance of suicide risk screenings that focus on the assessment of active and passive suicidal ideation and suicide planning, in addition to nonsuicidal self-injury, across pediatric medical settings.