A Comparison of Suicide Risk Scales in Predicting Repeat Suicide Attempt and Suicide

OBJECTIVE: To compare the predictive accuracy of the Suicide Intent Scale (SIS), the Suicide Assessment Scale (SUAS), the Karolinska Interpersonal Violence Scale (KIVS), and the Columbia-Suicide Severity Rating Scale (C-SSRS) for suicide attempts and suicides within 3 and 12 months of an episode of self-harm. METHODS: This prospective multicenter cohort study included patients (N = 804) aged 18-95 years with a recent episode of self-harm assessed in psychiatric emergency settings from April 2012 to April 2016. Suicide attempts and suicides were identified in medical records and in the National Cause of Death Register. Receiver operating characteristic curves were constructed, and accuracy statistics were calculated. A sensitivity of at least 80% combined with a specificity of at least 50% were considered minimally acceptable. RESULTS: At least 1 suicide attempt was recorded for 216 participants during follow-up, and 19 participants died by suicide. The SUAS and C-SSRS were better than chance in classifying the 114 suicide attempts occurring within the first 3 months; a C-SSRS score ≥ 27 yielded a sensitivity/specificity of 79.8%/51.5% (P < .001). During 1-year follow-up, the SUAS and C-SSRS also performed better than chance, but no cutoff on either instrument gave a sensitivity/specificity of ≥ 80%/≥ 50%. The SIS was the only instrument that could classify suicides correctly. At 3 months, the area under the curve (AUC) was 0.94 (95% CI, 0.89-0.99), and a score ≥ 21 predicted suicide with a sensitivity/specificity of 100%/81.9%, based on only 4 suicides. At 1-year follow-up, the AUC was 0.74 (95% CI, 0.61-0.87), and a score ≥ 17 predicted suicide with a sensitivity/specificity of 72.2%/57.9%. CONCLUSIONS: Instruments that predicted nonfatal repeat suicide attempts did not predict suicide and vice versa. With the possible exception of the prediction of suicide by the SIS in a short time frame, the specificity of these instruments was low, giving them a limited relevance in the prediction of suicidal behaviors.