For decades, our ability to predict suicidal thoughts and behaviors has been at near-chance levels (Franklin et al., 2017). A recent study, led by Dr. Jessica D. Ribeiro at Florida State University, was designed to take steps toward advancing our ability to predict suicidal thoughts and behaviors by addressing major methodological issues pervasive to prior suicide prediction efforts.
Dr. Ribeiro and her team addressed two major methodological constraints. First, they focused on short-term prediction. The vast majority of suicide prediction studies have focused on the prediction of eventual (i.e., long-term) rather than imminent (i.e., short-term) suicidal thoughts and behaviors. Recent meta-analytic evidence indicates that effects of risk factors on eventual suicidal thoughts and behaviors are modest at best. However, Dr. Ribeiro and colleagues reasoned that it was possible that the same risk factors would have considerably stronger effects over shorter, more clinically useful timeframes. Second, they considered more complex conceptualizations of suicide risk. Most prior studies have considered fairly simple and determinate conceptualizations of suicide risk, often examining risk factors in isolation or within small sets of risk factors combined in fairly rudimentary ways (e.g., screeners; sum scores). Dr. Ribeiro and her team posited that suicide risk may instead be complex and indeterminate. As such, they examined the predictive utility of applying machine learning methods for prediction of suicidal thoughts and behaviors. Dr. Ribeiro and colleagues followed a sample of over 1,000 suicidal and/or self-injuring adults 3, 14, and 28 days after baseline assessments. Participants completed a large battery of self-report and implicit measures at each timepoint. Retention rates were over 90% at each follow-up; close to 10% of participants attempted suicide over the course of the study. Results indicated that short-term prediction alone did not improve prediction for suicide attempts at follow-up – even commonly cited “warning signs” for suicide (e.g., insomnia; agitation; hopelessness) produced predictive accuracy only marginally better than chance. Machine learning risk algorithms demonstrated improved predictive performance relative to univariate prediction. Critically, although many univariate predictors did not demonstrate strong performance in the short-term, this does not mean that these factors have no relevance to STBs; rather, these findings suggest that these factors, when considered in isolation, may be insufficient to produce accurate prediction of imminent suicidal thoughts and behaviors.
References