Preliminary development of a child physical abuse algorithm using emergency department records

Area of Science:

• Pediatric Emergency Medicine
• Data Science in Healthcare
• Child Abuse Prevention

Background:

• Child physical abuse (CPA) identification in emergency departments is challenging.
• There's a critical need for tools to help providers recognize child abuse.
• A predictive algorithm could improve CPA identification and risk factor awareness.

Purpose of the Study:

• To develop and evaluate a predictive algorithm for child physical abuse (CPA).
• To assess the algorithm's performance using statistical and machine learning techniques.

Main Methods:

• Utilized discharge data from a pediatric hospital (May 2017-March 2022).
• Employed machine learning models, including XGBoost and lasso regression.
• Evaluated model performance using AUROC, sensitivity, and PPV, stratified by age and gender.

Main Results:

• The XGBoost model achieved an 82.6% AUROC.
• At 90% specificity, sensitivity was 57.7% and PPV was 1.2%.
• High-risk young children presented with somatic symptoms; older children had behavioral health diagnoses.

Conclusions:

• Developing a CPA predictive algorithm using discharge data is feasible.
• Behavioral health professionals play a key role in recognizing CPA.
• Further investigation into age and gender intersectionality can enhance interventions for high-risk children.