Prediction of violence victimisation among children and adolescents using machine learning

Background Violence victimisation among United States (USA) youth is a major public health concern, with over 6 600 homicides and 320 000 non-fatal assault-related emergency visits among 10–24 years old in 2022. This study aims to identify socio-ecological risk factors and explore the use of machine learning to predict the risk of violence victimisation, defined as a broad indicator of exposure to crime, violence or assault occurring within a 12-month period among children and adolescents. Method We used seven machine learning algorithms, including support vector machines (SVM), random forest and extreme gradient boosting. We computed SHapley Additive exPlanation (SHAP) values to interpret the results. This study used secondary data from the Adolescent Brain Cognitive Development Study, which includes information on 11 386 children and adolescents observed from 2018 to 2022. Results Random forest using disaggregated variables achieved a sensitivity of 0.81, while non-linear SVM using aggregated variables achieved a sensitivity of 0.83. Both models identified predictors associated with increased risk of violence victimisation, including age, adverse life events and family conflict. In the random forest model, specific life events, such as witnessing a crime or accident, parents getting divorced, and specific family conflicts, including family members trying to one-up each other and frequent fights, were linked to victimisation. Conclusion The findings offer insights into violence victimisation among children and adolescents, highlighting the potential of advanced computational techniques to uncover complex patterns. While the models achieved high sensitivity, further work is needed to assess calibration, improve predictive precision and determine whether these approaches can support real-world decision-making.