How accurate are suicide risk prediction models? Asking the right questions for clinical practice

Abstract

Prediction models assist in stratifying and quantifying an individual's risk of developing a particular adverse outcome, and are widely used in cardiovascular and cancer medicine. Whether these approaches are accurate in predicting self-harm and suicide has been questioned. We searched for systematic reviews in the suicide risk assessment field, and identified three recent reviews that have examined current tools and models derived using machine learning approaches. In this clinical review, we present a critical appraisal of these reviews, and highlight three major limitations that are shared between them. First, structured tools are not compared with unstructured assessments routine in clinical practice. Second, they do not sufficiently consider a range of performance measures, including negative predictive value and calibration. Third, the potential role of these models as clinical adjuncts is not taken into consideration. We conclude by presenting the view that the current role of prediction models for self-harm and suicide is currently not known, and discuss some methodological issues and implications of some machine learning and other analytic techniques for clinical utility.