Risk of unnatural death following self-harm in South Africa: development and validation of multivariable prognostic models.

backgroundRecurrent self-harm is common and is associated with an increased risk of unnatural death including suicide and fatal accidents. We developed and validated prognostic models to stratify individuals by risk of unnatural death after healthcare presentation for non-fatal self-harm to support clinical decision-making and targeted prevention when capacity to deliver evidence-based intervention is limited.

methodsWe used insurance claims and vital registration data from 6846 South African medical insurance beneficiaries aged ≥10 years who were discharged alive after healthcare presentation for self-harm during 2011-2021 to develop and validate models predicting unnatural death after non-fatal self-harm. We fitted competing-risk regression models to predict unnatural death within 1-3 years after presentation and discharge. Variable selection was guided by the least absolute shrinkage and selection operator. We used bootstrapping for internal validation and estimated optimism-corrected concordance indices (C-index), calibration intercepts and slopes, and for risk stratification the proportions of unnatural deaths captured across predicted-risk groups. Final models included age, sex, encounter characteristics, prior psychotropic medication use and selected prior mental disorder diagnoses.

resultsThe final models achieved optimism-corrected C-indices of 0.74 at presentation and 0.75-0.76 at discharge and identified 86-88% of observed unnatural deaths within 2 years among the 40% of individuals with the highest predicted risk. Models showed little evidence of systematic miscalibration (optimism-corrected calibration intercepts -0.02 to -0.01), but some overfitting (optimism-corrected calibration slopes 0.84-0.92). Using the model with the highest optimism-corrected C-index, the simplified discharge model, the observed 2-year risk of unnatural death was 0.15% (95% CI 0.06% to 0.32%) among the 60% of individuals with the lowest predicted risk, comparable to the risk among 1 249 760 individuals without a self-harm history (0.14%, (95% CI 0.14% to 0.15%)).

conclusionsThe prediction models effectively rank individuals accessing private sector care in South Africa by their probability of unnatural death following healthcare presentation for non-fatal self-harm. CLINICAL IMPLICATIONS: The models are intended to support clinical decision-making and guide the prioritisation of individuals with predicted high risk for targeted interventions, supporting more efficient allocation of mental health care capacity, which is often limited in low- and middle-income settings.