Employing machine learning for early detection of poly-victimization in rural children: a survey study in China's Chaoshan region.

backgroundPoly-victimization (PV), encompassing multiple forms of victimization including physical abuse, emotional maltreatment, neglect, and peer violence, poses a significant public health challenge among children, particularly in rural areas with high rates of children whose parents have migrated to cities for work, leaving them in rural areas (left-behind children). This study investigates PV among rural children in the Chaoshan region of China, an area with distinct economic and cultural characteristics.

methodsA thematic survey on PV occurrence was conducted among rural children in Shantou and Jieyang areas using a unified strategy. Four machine learning models, including Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Artificial Neural Network (ANN), and Random Forest (RF), were employed to predict PV risk, and SHAP feature importance was utilized to evaluate risk factors. An early-warning index for PV was constructed using linear regression and feature importance.

resultsChildren in Jieyang were 1.84 times more likely to experience PV compared to those in Shantou (22.95% vs 12.49%). Among PV victims, left-behind children showed a notably higher proportion in Shantou (46.09%) compared to Jieyang (24.48%). The study successfully established specific predictive models for PV among rural children, with an overall prediction accuracy exceeding 80% across regions and 82% for left-behind children. The SHAP framework revealed significant risk factors, such as witnessing school bullying (contributing up to 22.72%) and self-harm intentions (up to 16.43%). The early-warning index demonstrated that the region and left-behind status significantly impacted PV occurrence. Specifically, the PV warning indices for Shantou and Jieyang were 0.621 (IQR: 0.558-0.761) and 0.497 (IQR: 0.422-0.658), respectively, significantly higher than the non-PV warning indices of 0.253 (IQR: 0.037-0.380) and 0.161 (IQR: 0.104-0.256). Left-behind children had higher PV warning indices than non-left-behind children.

conclusionsThis study demonstrates the utility of machine learning models in predicting PV among rural children, particularly left-behind children, in the Chaoshan region. Identifying risk factors and developing an early-warning index provide valuable tools for injury prevention, risk assessment, and targeted interventions, with potential applications in public health policy.