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

BACKGROUND: Victims of child physical abuse (CPA) disproportionately use the emergency department, yet identifying CPA remains challenging in this setting. An urgent need to develop tools that aid providers in recognising violence-related injuries in children exists. Developing a CPA predictive algorithm could improve identification and clinical familiarity with risk factors. We aimed to develop a CPA predictive algorithm and assess its performance. METHODS: Using data from May 2017 to March 2022 involving patients aged 0-17 years treated at a large paediatric hospital in Connecticut, we used statistical and machine learning techniques (eg, XGBoost, lasso regression) to develop a CPA predictive algorithm. Model performance was evaluated using out-of-sample area under the receiver operating characteristic curve (AUROC), sensitivity and positive predictive value (PPV). Performance metrics were stratified by age group and gender. RESULTS: Among 138 234 patients, 298 had documented CPA. Most victims were aged 0-9 years, using Medicaid, and of Hispanic/Latino ethnicity. XGBoost demonstrated the best performance. The out-of-sample AUROC was 82.6% (SE=1.6%). At 90% specificity, sensitivity was 57.7% (SE=3.1%), and PPV was 1.2% (SE=0.1%). Among high-risk patients, older children presented with behavioural health diagnoses, while young children presented with non-specific somatic symptoms; no gender differences were identified. CONCLUSIONS: This study demonstrates the feasibility of using discharge data to develop a CPA algorithm. Findings underscore the role of behavioural health professionals in recognising CPA and highlight high-risk groups that warrant further investigation. Examining the intersectionality of gender and age among children at high risk for CPA may reveal additional opportunities for clinical intervention.