Natural Language Screening to Improve Early Recognition of Child Physical Abuse in Emergency Care Settings

PROJECT SUMMARY/ABSTRACT This project will test an innovative program of automated screening using natural language processing (NLP) to improve early recognition of child physical abuse in urgent and emergency care settings. Physical abuse affects approximately 125,000 US children each year, with infants at highest risk. More than 30% of

children who suffer serious abusive injuries have had prior minor injuries that might have raised the concern for abuse. Improving the recognition of these including bruises, fractures, and others, is the best opportunity to prevent recurrent abuse, escalating injury, and death. Missed abuse is especially common in emergency

departments and urgent care settings, and remains a major public health problem despite efforts at education and awareness-building, routine screening by clinicians, mnemonics, clinical decision rules, and guidelines by professional societies. These interventions are resource-intensive and depend on human vigilance to

recognize a condition that is rare in each setting, but which has enormous collective impact. Further, these interventions are vulnerable to bias and may exacerbate racial and ethnic disparities. Automated computer screening occurs in the background of clinical care and can overcome limitations of methods that depend on human effort. We developed and internally validated one automated screener that

uses NLP to analyze unstructured narrative data within the electronic health record. External validation is needed to determine accuracy in other clinical settings and test the association of high-risk injuries with subsequent abuse. Further, testing is needed to determine the effect of the screener on racial disparities and

avoidable Child Protective Services reports. If externally validated and shown to be equitable, this NLP screener could be rapidly adopted in diverse settings. We propose to externally validate this tool and accomplish these other objectives by determining statewide CPS outcomes (referrals, substantiations, services

provided) for >100,000 urgent care and ED visits for infants in two large healthcare systems. Aim 1: Externally validate an automated NLP screener to identify infants with high-risk injuries in urgent and emergency care settings. We will use manual chart review as the criterion standard for visits identified by the NLP screener and

other high-risk encounters. Aim 2: Determine rates of subsequent abuse for infants with high-risk injuries identified by automated NLP screening but not by usual care. We will link statewide clinical and CPS outcomes to identify abuse within 12 months for >1100 encounters with high-risk injuries. Aim 3. Determine the potential

impact of NLP on racial disparities. The expected outcome of the project is the external validation of an automated abuse recognition tool to improve recognition and tertiary prevention of abuse. Our innovative methods could also be used to test other healthcare interventions to improve the recognition of abuse and other conditions.