Background: Youth suicide is a growing global public health problem. Predictive algorithms and in-person screening show promise in identifying patients at risk of suicidal behavior, yet there is little research examining their comparative performance in children and adolescents using both a validated predictive algorithm and screening tool. We compare the performance of suicide risk screening and a risk algorithm in predicting suicide attempts among pediatric emergency department patients.
Design: In this retrospective cohort study, 19,563 10-18 year old emergency department patients from an urban pediatric hospital in the US were screened from September 2019 through August 2021 for suicide risk using the Ask Suicide Questions survey and the Columbia-Brief Suicide Severity Rating Scale, and their historical electronic health records were extracted to train a risk prediction algorithm. We followed patients from their first screen or first visit in the screening period, if not screened, to the end of the study to observe the presence/absence of a suicide attempt. We deployed a previously published modeling procedure (marginal feature screening with multivariable lasso regression modelingusing the glmnet package in R to identify patients at risk of suicide based on predictors extracted from clinical records prior to and during screening.
Results: Among the 19,563 patients, 9,646 (49.1%) were male, 10,447 (53.2%) were 14-18 years old, 7,834 (39.9%) were Non-Hispanic White, and 495 (2.5%) were treated for a suicide attempt. Among patients that screened positive for suicide risk in testing samples (M=8.1% [95% CI: 7.6, 8.6]) and the identical number of highest risk patients identified by the algorithm, the algorithm correctly identified an average of 50.9% (95% CI: 47.1, 54.6) of attempters in contrast to 36.5% (95% CI: 31.9, 41.2) correctly identified by screening. The algorithm uniquely identified 129% more attempters (n=126) than did screening (n=55).

Conclusions and Relevance: Although the risk algorithm was superior to screening across all performance metrics, our results reveal the potential value of drawing on both screening and predictive algorithms in assessing pediatric suicide risk in clinical settings. Future research should explore opportunities to achieve clinical efficiencies by targeting screening efforts to patients whose risk is unlikely to be detected by predictive algorithms.