608 - Pilot Study Using Machine Learning to Improve Estimation of Physical Abuse in Emergency Departments

Current reported numbers of physical abuse (PA) are likely underestimates due to challenges in accurately measuring abuse. International Classification of Diseases (ICD) billing codes for PA lack sensitivity in emergency department (ED) settings. We propose a machine learning approach to develop an age-specific ICD code schema using both abuse and injury codes to measure the burden of PA more accurately.

Objective:

Determine the feasibility of machine learning to identify ICD-10-CM codes associated with PA.

Design/Methods:

This retrospective study used the Pediatric Health Information System (PHIS) and the child protection team’s (CPT) administrative database at a large midwestern children’s hospital to identify all ICD-10-CM codes and abused and non-abused children evaluated in the ED between 2016-2020. PHIS identified all ED patients < 5 years with an injury or abuse-specific diagnosis code. After excluding repeat visits, vehicular accidents, birth injuries, metabolic bone diseases and bleeding disorders, abused and non-abused patients who had been evaluated by the CPT were identified. In order to optimize accuracy of diagnosis, only patients evaluated by the CPT were included. Analysis was conducted for the entire study group and predictions were stratified by age. Characteristics and ICD-10-CM codes were compared for abused and non-abused patients using t-tests, Kruskal-Wallis tests, chi-squared tests or Fisher’s exact tests. Lasso regularized logistic regression determined a predictive model for identifying abused children. Receiver Operating Characteristics (ROC) curves and Precision-Recall curves were estimated using stratified repeated 10-fold cross validation (CV).

Results:

We identified 196,901 patients < 5 years with an injury or abuse-specific diagnosis code. After exclusion criteria were applied, 194,586 patients remained. Of these, 2028 (1%) patients were evaluated by the CPT and 512 were identified as abused. Using diagnosis codes and patient age, our predictive model was able to accurately identify patients with confirmed PA (mean CV AUC = 0.87). Predictive performance was slightly weaker for patients without existing ICD codes for abuse (mean CV AUC = 0.81).

Conclusion(s):

Our models have reasonable predictive power for distinguishing between abused and non-abused children and could improve population-level estimates of abuse burden over time, and in response to social interventions.