OA2012. Self-harm recidivism prediction with a gradient boosting ML approach on the European Injury Database

Abstract Background Predicting intentional self-harm recidivism is critical for reducing preventable mortality and optimizing resource allocation in mental healthcare. While existing risk assessment tools rely on clinician-reported factors, machine learning offers data-driven insights into underrecognized predictors. Methods We developed an XGBoost gradient boosting model using 7,375 records from the EU-IDB Full Dataset (41.7% with intentional self-harm recidivism), employing SHAP (SHapley Additive exPlanations) analysis to identify key predictors. The model was optimized via dynamic threshold adjustment (optimal threshold: 0.3479). SHAP values are shown for each feature. Results The model demonstrated excellent accuracy: precision 0.78, recall 0.78 and an F1-score of 0.78, with an AUC 0.86. SHAP analysis revealed critical drivers of recurrence: Mechanism of injury (0.58), Activity when injured (0.33), and Object involved (0.29). Among the most impactful modalities by feature were: Mechanism of injury (foreign body 1.01, exposure to chemical or other substance 0.84); Activity when injured (other specified leisure or play 0.47, unspecified leisure or play 0.43); Object involved (personal grooming utensil 0.81 e.g. razor, razor blade and sedatives or antipsychotics 0.53). Although Age showed a moderate overall importance (0.20), individual-level analysis highlighted a subset of patients aged 46-62 years with significantly elevated SHAP values (up to 0.77). Temporal features such as the COVID-19 period had limited impact (0.06), indicating that pandemic effects were secondary to clinical and situational factors. Conclusions The model’s high accuracy (AUC 86%) supports its potential as a screening tool in emergency departments, a prioritization system for follow-up care, and a template for prevention programs. Future research should validate these findings across diverse healthcare systems. Key messages • Machine learning model predicts self-harm recidivism with high accuracy identifying key risk factors. • Key predictors are mechanism of injury, activity during injury, and objects involved, more than demographic factors. Topic Self-harm injury, Risk prediction, SHAP analysis.