Suicide risk assessment tools and prediction models: new evidence, methodological innovations, outdated criticisms

The number of prediction models for suicide-related outcomes has grown substantially in recent years. These models aim to assist in stratifying risk, improve clinical decision-making, and facilitate a personalised medicine approach to the prevention of suicidal behaviour. However, there are contrasting views as to whether prediction models have potential to inform and improve assessment of suicide risk. In this perspective, we discuss common misconceptions that characterise criticisms of suicide risk prediction research. First, we discuss the limitations of a classification approach to risk assessment (eg, categorising individuals as low-risk vs high-risk), and highlight the benefits of probability estimation. Second, we argue that the preoccupation with classification measures (such as positive predictive value) when assessing a model's predictive performance is inappropriate, and discuss the importance of clinical context in determining the most appropriate risk threshold for a given model. Third, we highlight that adequate discriminative ability for a prediction model depends on the clinical area, and emphasise the importance of calibration, which is almost entirely overlooked in the suicide risk prediction literature. Finally, we point out that conclusions about the clinical utility and health-economic value of suicide prediction models should be based on appropriate measures (such as net benefit and decision-analytic modelling), and highlight the role of impact assessment studies. We conclude that the discussion around using suicide prediction models and risk assessment tools requires more nuance and statistical expertise, and that guidelines and suicide prevention strategies should be informed by the new and higher quality evidence in the field.