Selected elements and findings of three recent systematic reviews of approaches to suicide and self-harm prediction
Review | Types of approach included | Primary study populations | Performance measures | Findings and conclusions |
Runeson et al 3 | Psychological rating scales with risk cut-off applied Prediction tools (unweighted variables) | Psychiatric patients (inpatient and outpatient) Individuals presenting to emergency settings Primary care patients | Sensitivity Specificity (NPV/PPV in supplement) | None achieved predefined accuracy threshold (80% sensitivity, 50% specificity). No support for use. Unclear whether may improve prediction as complement to clinical impression. |
Carter et al 4 | Biological measures with risk cut-off applied Psychological rating scales with risk cut-off applied Prediction tools (unweighted and weighted variables) | Psychiatric patients (inpatient and outpatient) Individuals presenting to emergency settings Military veterans Prisoners | PPV LR/CUI* summarised | Combined pooled PPV 26.3% for self-harm and 5.5% for suicide. No individual instrument or pooled subgroup with accuracy suitable to allocate treatment. |
Belsher et al 5 | Prediction models derived by various methods (including machine learning) | Psychiatric patients (inpatient and outpatient) Individuals presenting to emergency settings Primary care patients Military populations General population | AUC Accuracy Sensitivity PPV | Good overall classification, but low PPV. Would result in high false-positive and considerable false-negative rates if used in isolation. At present limited practical utility. |
*Performance metrics primarily applied to diagnostic and screening tests.
AUC, area under the receiver operating characteristic curve; CUI, clinical utility index; LR, likelihood ratio; NPV, negative predictive value; PPV, positive predictive value.