Category: Suicide and Self-Injury
Keywords: Suicide | Risk / Vulnerability Factors | Research Methods
Presentation Type: Symposium
Accurate risk prediction is crucial to deliver cost-effective prevention and treatment. However, our ability to predict suicide ideation has remained around chance levels (Franklin et al., 2017). Using a sample of suicide ideators and controls, this study aims to evaluate: 1) whether unsupervised ML provides evidence for the existence of two distinct groups (i.e., ideators versus non-ideators); 2) whether supervised ML can accurately identify suicide ideators; 3) whether supervised ML provides superior identification compared to traditional multiple logistic regression.
The sample included 914 current firefighters who completed a survey on suicide and behavioral health. Suicide ideation status was determined using the Self-Injurious Thoughts and Behaviors Interview (SITBI; Nock et al., 2007). We performed hierarchical and k-means cluster analyses as unsupervised ML and random forests as supervised ML. All analyses were conducted in R with the MICE, randomForest, and pROC packages.
The sample was predominantly white (86.9%) and male (90.0%), with a mean age of 36.90 years old (SD = 10.78). About half of the sample (48.1%; n = 440) reported experiencing suicide ideation. The analyses included 48 predictors and one outcome (i.e., suicide ideation). Dendrograms produced by hierarchical cluster analyses provided evidence for at least two distinct subgroups in the sample. The two subgroups identified by k-means cluster analysis were consistent with participants’ suicide ideation status (AUC=.72 [.70 - .75]). Random Forests algorithms produced acceptable to good prediction for suicide ideation status (AUC=.79 [.76-.83]). Traditional multiple logistic regression models that included the same 48 predictors produced chance-level prediction (AUC = 0.50 [0.49 – 0.51]).
This study suggests both unsupervised and supervised ML can greatly improve our abilities to identify suicide ideators compared to traditional statistical methods. However, the number of predictors (i.e., 48) included in this study were still considered small for ML. Future studies should examine whether including more predictors will result in even higher identification accuracy.
Florida State University
Friday, November 17
3:30 PM – 4:30 PM
The asset you are trying to access is locked. Please enter your access key to unlock.