BACKGROUND: Many countries lack comprehensive systems or standardized use of diagnostic codes to track child abuse. This study evaluated the validity of the International Classification of Diseases, 10th Revision (ICD-10) diagnostic codes and developed algorithms to identify physical abuse using administrative data. METHODS: We analyzed children under 10 years hospitalized for injuries between April 2013 and March 2023 at a children's hospital in Japan. Child protection team records notifying local services were used as the reference standard. We assessed diagnostic, procedure, and medication codes using machine learning. RESULTS: Among 1375 injury cases, 53 involved physical abuse, but only one used an abuse-specific ICD-10 code, indicating limited reliability. Combining X-ray and fundus examinations achieved 73.6% sensitivity (95% CI, 59.7-84.7) and 85.7% specificity (95% CI, 83.7-87.5) among children aged < 10 years and 80.9% sensitivity (95% CI, 66.7-90.9) and 79.1% specificity (95% CI, 76.2-81.7) among children aged < 6 years. Several machine learning approaches, including Lasso regression, random forest, and bootstrap classification-and-regression-tree models, yielded broadly consistent findings. CONCLUSIONS: ICD-10 codes alone are insufficient for identifying physical abuse combining specific procedures may improve case identification for future epidemiological surveillance.
Obikane et al. (2026) studied this question.