Clear Intent - Suicidology Data and Evidence podcast

Episode 4: - Can we predict suicide?

0:00
37:18
Spol 15 sekunder tilbage
Spol 15 sekunder frem

Can we predict who will die by suicide? In 1983, Alex Pokorny tried with 4,800 psychiatric inpatients and nearly every rating scale available. More than 96% of the men he flagged as high risk did not die by suicide, and more than half of those who did had been rated low risk. Four decades later, we have machine learning, health records on millions of people, and a Harvard smartphone study reported this year as a “weather app” for suicide risk, with accuracy figures as high as 94%. Dr. Tyler Black works through what those numbers mean, starting with a plain-language primer on sensitivity, specificity and positive predictive value, then looks at what happens to the Harvard model when it meets patients it hasn’t seen before. Along the way he covers the 2025 meta-analysis of 53 machine learning studies, the US Army’s 2026 algorithm, and the VA’s REACH VET program, which has been flagging veterans since 2017. He also argues for what clinicians should do instead: give everyone the basics that work, and build every assessment around the individual person’s own risk and protective factors. If you are struggling, in Canada and the US call or text 988.

Reviewing and checking the audio, and responding to feedback, please see the following corrections

* US requirements: There is no US law requiring suicide risk stratification. The requirement comes from the Joint Commission’s accreditation standard (NPSG.15.01.01). It requires accredited hospitals and behavioural health organizations to screen for suicide risk, assess anyone who screens positive, and document each patient’s “overall level of risk.”

* Large et al. (2016): The 4.84 figure is a pooled odds ratio, not a relative risk. The 5.5% figure (the share of high-risk patients who died by suicide) was reported by the authors themselves.

* Spittal et al. (2025): The modelled PPV for suicide death goes above 5–6%. It reached about 17% in people discharged after self-harm. Even there, more than 8 in 10 flags are wrong.

* Nock et al. (2026): The held-out range of 0.32–0.80 is for suicide-related events. The strongest predictors were a recent suicide-related event and agitation; aggression was not a predictor. Participants were paid $1 per survey.

* SPOT (Simon et al., 2022): The hazard ratio of 1.29 for the online DBT skills arm applies to everyone offered the program, not only those who engaged with it. Engagement was 31% in the care management arm and 39% in the skills training arm.

References for this podcast:

Pokorny et al., (1983) Prediction of suicide in psychiatric patients. Report of a prospective study, Archives of General Psychiatry.

Large et al., (2016) Meta-analysis of longitudinal cohort studies of suicide risk assessment among psychiatric patients: heterogeneity in results and lack of improvement over time, PLOS ONE.

Belsher et al., (2019) Prediction models for suicide attempts and deaths: a systematic review and simulation, JAMA Psychiatry.

Spittal et al., (2025) Machine learning algorithms and their predictive accuracy for suicide and self-harm: systematic review and meta-analysis, PLOS Medicine.

Nock et al., (2026) Using smartphone surveys to predict next-week suicide attempts, Journal of Psychopathology and Clinical Science.



Get full access to Suicidology Data and Evidence at suicidology.substack.com/subscribe

Flere episoder fra "Clear Intent - Suicidology Data and Evidence"