Fig 2. Model performance.

Overall model performance in predicting both suicidal ideation (SI) and concomitant suicidal ideation and attempt (SA) using each method with and without Boruta feature selection. Logistic regression with feature selection and elastic-net without feature selection were the top performing models for classifying SI from controls (AUC = 0.70; CI 95%: 0.70–0.71), but did not perform significantly better than any other model trained after feature selection. The random forest with feature selection, trained only on SI vs. controls, was able to distinguish a smaller holdout test set of individuals endorsing both SI and SA from controls (AUC = 0.77; CI 95%: 0.76–0.77).