Research in the assessment of suicide risk has been criticised due to poorly performing tools, which have typically been developed for other purposes. At the same time, new prediction work in the field has tended to have methodological problems with overfitted models, which sometimes lack external validations, and have inadequate sample sizes for multiple candidate predictors. This is compounded by complex models with poor interpretability and transparency.


In this presentation, I will outline two new clinical prediction models that have been developed using multivariable accelerated time failure models that address some of the weaknesses in previous tools. Both have been developed using a cohort of 53,172 people who have presented to healthcare with self-harm episodes in Sweden, and followed up for suicide mortality and repeat self-harm outcomes over the subsequent 12 months. High-quality linked national registers were used to test socio-demographic and clinical predictors and their association with both outcomes (suicide mortality and repeat self-harm). The first model, for suicide risk, contained 11 factors: age, sex, substance misuse, mental health history and treatment, and history of and index self-harm. It showed good discrimination (c-index 0.77, 95% CI 0.75 to 0.78) and calibration in external validation (where there were 178 suicides within 12 months). For risk of suicide within 12 months, using a 1% cut-off, sensitivity was 82% (75% to 87%) and specificity was 54% (53% to 55%). This was translated into web-based risk calculator (Oxford Suicide Assessment Tool for Self-harm or OxSATS).

The second model, for repeat self-harm, included 17 factors in the final model, including sex, age, substance misuse factors, living situation, recent medication, features of self-harm history, recent psychiatric history, and family history of a psychiatric disorder. It showed good calibration (O:E ratio=1.15, 95%CI 1.09 to 1.20) and discrimination with a c-statistic of 0.72 (95%CI 0.70 to 0.73), based on external validation for repeat self-harm within 6 months (n=1,373 repeat incidents). At a 10% cut-off, sensitivity was 51.5% (48.8 to 54.2) and specificity 80.7% (80.1 to 81.4). The model was translated into the OxSET web-based risk calculator.

We conclude that OxSATS and OxSET accurately predict suicide and repeat self-harm over specified time frames. Further validations and linkage to effective interventions are required to examine clinical utility.