Study: Artificial intelligence more accurately identifies child abuse

(Press-News.org) Artificial intelligence (AI) can help better identify prevalence of physical abuse of children seen in the emergency room, a new study found. The research will be presented at the Pediatric Academic Societies (PAS) 2025 Meeting, held April 24-28 in Honolulu. 

Researchers used a machine-learning model to estimate instances of child abuse seen in emergency departments based on diagnostic codes for high-risk injury and physical abuse. The researchers’ approach better predicted abuse rates than those that rely solely on diagnostic codes entered by a provider or administrative staff. Relying on abuse codes alone misdiagnosed on average 8.5% of cases.

“Our AI approach offers a clearer look at trends in child abuse, which helps providers more appropriately treat abuse and improve child safety,” said Farah Brink, MD, child abuse pediatrician at Nationwide Children's Hospital, and assistant professor at The Ohio State University. “AI-powered tools hold tremendous potential to revolutionize how researchers understand and work with data on sensitive issues, including child abuse.”

Researchers studied data from 3,317 injury and abuse-related emergency department visits at seven children’s hospitals between February 2021 and December 2022. All children were under the age of 10 and nearly three quarters were under the age of two.

# # #

EDITOR:
Dr. Farah Brink will present “A Machine Learning Approach to Improve Estimation of Physical Abuse” on Mon., April 28 from 5:00-6:30 PM ET. 
Reporters interested in an interview with Dr. Brink should contact Amber Fraley at  amber.fraley@pasmeeting.org.
The PAS Meeting connects thousands of pediatricians and other health care providers worldwide. For more information about the PAS Meeting, please visit www.pas-meeting.org.

About the Pediatric Academic Societies Meeting
Pediatric Academic Societies (PAS) Meeting connects thousands of leading pediatric researchers, clinicians, and medical educators worldwide united by a common mission: Connecting the global academic pediatric community to advance scientific discovery and promote innovation in child and adolescent health. The PAS Meeting is produced through the partnership of four leading pediatric associations; the American Academy of Pediatrics (AAP), the Academic Pediatric Association (APA), the American Pediatric Society (APS), and the Society for Pediatric Research (SPR). For more information, please visit www.pas-meeting.org. Follow us on X @PASMeeting and like us on Facebook PASMeeting.

Abstract: A Machine Leaming Approach to Improve Estimation of Physical Abuse

Presenting Author: 
Farah Brink, MD

Organization
Nationwide Children's Hospital

Topic
Child Abuse & Neglect

Background
International Classification of Diseases, 10th Revision, Clinical Modification (ICD-10-CM) codes are inaccurate for determining child physical abuse (PA) prevalence, particularly in emergency department (ED) settings. Consideration of injury codes along with abuse-specific codes may enable more accurate PA prevalence estimates.

Objective
To develop a coding schema to better estimate PA using machine learning.

Design/Methods
We performed a secondary data analysis of children < 10 years evaluated by a child abuse pediatrician (CAP) due to concerns for PA during Feb 2021-Dec 2022 at 7 children's hospitals contributing data to both CAPNET, a multicenter child abuse research network, and Pediatric Health Information System (PHIS). We excluded encounters not linked with PHIS and those not evaluated in the ED during the CAPNET encounter. True PA was defined by CAP assigned rating 5- 7 on a 7-point scale of PA likelihood within the CAPNET database. Abuse-specific codes, including suspected codes, were defined as ICD-10-CM codes for PA modified from the Centers for Disease Control and Prevention child abuse and neglect syndromic surveillance definition. All 4-digit injury ICD-10-CM codes were used. We developed LASSO logistic regression models to predict CAP¬ determined PA for encounters with and without abuse-specific codes and used the models to calculate site-specific estimates of PA prevalence. We calculated the estimation error for site estimates based on 1) abuse-specific codes alone and 2) our LASSO predictive models. Estimation error was defined as estimated PA prevalence minus CAP-determined PA prevalence (true value).

Results
3317 of 6178 CAPNET encounters were successfully linked with PHIS and seen in the ED. Median age was 8.4 months with 74% < 2 years and 59% < 1 year. CAP diagnosed PA in 35% (n=l145) of all encounters, 12.7% (n=240) of encounters without abuse-specific codes, and 63.4% (n=905) of encounters with abuse-specific codes. At least one abuse-specific code was assigned for 43% of encounters. Site-specific estimates of PA prevalence based only on assignment of abuse-specific codes overestimated prevalence with estimation errors ranging from 2.0% to 14.3% (average absolute error 8.5%). Estimates of site-specific PA prevalence based on our predictive models had reduced errors from -3.0% to 2.6% (average absolute error 1.8%) (Fig. 1). Absolute error decreased for 6 of 7 sites and increased by 0.6% for the remaining site (Fig. 2).

Conclusion(s)
Our predictive models more accurately estimated the prevalence of PA compared to abuse-specific codes alone.

Tables and Images
PAS Figure 1.ROC curves 20241101.png
PAS Figure 2.estimate plot.png

END