Predicting and explaining recurrent child abuse using interpretable machine learning: Evidence from national-level child abuse data in South Korea (2017-2020).

OBJECTIVE

This study aimed to develop machine learning models to predict the risk of each type of recurrent child abuse using national-level child abuse data. Additionally, it sought to identify key factors based on abuse type and explain recurrence risk in individual cases.

METHODS

The analysis employed a broad range of factors related to child, perpetrator, the initial episode of child abuse, and service, covering 51,517 abused children, 46,497 perpetrators, and 64,774 reported cases. Distinct predictive models were developed for each type of recurrent abuse: recurrent physical abuse (RPA), recurrent emotional abuse (REA), recurrent sexual abuse (RSA), and recurrent neglect (RN).

RESULTS

The highest AUC-ROC score was 0.793 for RN, followed by 0.749, 0.702, and 0.700 for RSA, REA, and RPA, respectively. Younger perpetrator age likely increased the risk of all types of recurrent child abuse. Planning or completing counseling for abused children and their perpetrators was an important factor in mitigating the risk of RPA, REA, and RSA. The initial episode of abuse being of the same type was the most influential factor for RPA, REA, and RN, while the perpetrator being man was the most significant factor for RSA. Additionally, factors associated with parenting and family environment-including inappropriate parenting attitudes, a lack of parenting knowledge and skills, and conflicts with family members-were crucial factors for RSA.

CONCLUSION

These findings can assist child protection agencies and related organizations in facilitating the early prevention of child abuse recurrence and designing interventions tailored to each abuse type.