Individual level prediction of emerging suicide events in the pharmacologic treatment of bipolar disorder

Abstract

Abstract Background Patients with bipolar disorder have a high lifetime risk of suicide. Predicting, preventing and managing suicidal behavior are major goals in clinical practice. Changes in suicidal thoughts and behavior are common in the course of treatment of bipolar disorder. Methods Using a dataset from a randomized clinical trial of bipolar disorder treatment (N=98), we tested predictors of future suicidal behavior identified through a review of literature and applied marginal variable selection and machine learning methods. The performance of the models was assessed using the optimism-adjusted C statistic. Results Number of prior hospitalizations, number of prior suicide attempts, current employment status and Hamilton Depression Scale were identified as predictors and a simple logistic regression model was constructed. This model was compared with a model incorporating interactions with treatment group assignment, and more complex variable selection methods (LASSO and Survival Trees). The best performing models had average optimism-adjusted C-statistics of 0.67 (main effects only) and 0.69 (Survival Trees). Incorporating medication group did not improve prediction performance of the models. Conclusions These results suggest that models with a few predictors may yield a clinically meaningful way to stratify risk of emerging suicide events in patients who are undergoing pharmacologic treatment for bipolar disorder.