Designing a Machine Learning-Based Model Integrating Clinical Orders for Child Abuse and Neglect Identification with Focus on Reducing Socio-economic Bias

Abstract Emergency departments (EDs) face challenges in identifying child abuse and neglect, often influenced by biases disproportionately affecting marginalized groups. While combining electronic health records (EHRs) with machine learning holds promise for addressing these biases, previous work has not integrated the Area Deprivation Index (a measure of socio-economic background). We analyzed 33,961 pediatric ED visits, of which 74 had suspected abuse or neglect incidents, to develop a machine learning model that emphasizes clinical orders and mitigates socio-economic biases. Evaluations across socio-economic strata based on the Area Deprivation Index revealed that a combined text and structured EHR model incorporating socio-economic variables achieved a precision of 0.8 for patients from disadvantaged backgrounds and 0.81 for those from more affluent groups. Key findings indicate that certain clinical orders were associated with increased reporting among lower socio-economic groups, whereas more affluent groups exhibited more critical ED visit attributes. These findings emphasize how ED clinical protocols can aid in identifying potential child abuse and neglect, while also highlighting differing indicators of protective service reports across socio-economic groups.