A new study presented at the Pediatric Academic Societies 2025 Meeting demonstrates that artificial intelligence substantially enhances the detection of physical child abuse in emergency departments. Researchers designed a machine learning model that analyzes diagnostic codes related to high-risk injuries and physical abuse indicators, and compared its accuracy to conventional methods relying strictly on diagnostic codes assigned by healthcare providers or administrative staff.
The traditional diagnostic coding system routinely misses approximately 8.5% of child abuse cases, often due to limitations in how healthcare providers assign specific codes. The machine learning model, however, provided more reliable estimates and identified trends more effectively, offering reduced estimation errors across multiple hospital sites. The study involved 3,317 emergency department cases concerning children under age 10, with nearly three-quarters of cases involving children under 2 years old. The research revealed that using only abuse-specific codes overestimated prevalence, while the machine learning approach kept error rates at a minimum, ranging from -3.0% to 2.6% (average absolute error 1.8%).
Researchers conducted a secondary analysis of cases evaluated between February 2021 and December 2022 at seven children’s hospitals participating in large-scale child abuse research networks. Their methodology utilized LASSO logistic regression models to predict likelihood of physical abuse, integrating both injury and abuse-specific codes. The improved accuracy could enable earlier intervention, better treatment, and improved outcomes for vulnerable children. Study authors emphasize that artificial intelligence-powered tools offer significant potential for advancing both clinical care and research in sensitive and complex domains like child abuse detection.
