Derivation of a Clinical Prediction Rule for Pediatric Abusive Fractures

点击翻译按钮获取中文摘要

PROJECT SUMMARY Child abuse and neglect represent one of the most serious pediatric public health crises, affecting nearly 1 in 7 children. Fractures are the 2nd most common abusive injury after skin and soft tissue injuries and there is much overlap between the types of fractures caused by abuse and unintentional mechanisms. The diagnosis of child abuse is complex and necessitates an accurate understanding of typical pediatric injury patterns within the context of history, mechanism, socio-demographics, and developmental capabilities. Many studies evaluating the relationship between fractures and abuse focused on specific fracture types, were restricted to children with a pre-defined abusive injury or included only admitted patients, and/or relatively small cohorts, thus limiting conclusions and raising concerns of spectrum bias. Additionally, prior literature has shown implicit and explicit biases related to socio-demographic factors in the identification and evaluation of abuse, likely resulting in over- and underdiagnosis of abuse in some populations. Furthermore, over 75% of children seeking ED care are seen in general ED’s by providers without specialized training in child development and abuse, and up to 1 in 5 children with abusive fractures may be missed in a general ED setting. Despite the frequency of abusive fractures and the potential limitations and biases in making the diagnosis, there are no validated clinical decision rules (CDRs) to assist clinicians in the real-time identification of children with fracture presentations associated with abuse. Our long-term goal is to develop a validated CDR that can be used by clinicians evaluating injured children to assist in the identification of abusive fracture presentations. Our primary objective is to utilize gradient boosted decision tree ensembles to develop a CDR that will identify fracture presentations highly concerning for abuse among patients ≤5 years presenting for emergency department (ED) care. An institutional child protection database that includes outcomes of thorough expert child abuse investigations will be used as a reference standard. The study objectives will be accomplished by 1) analyzing structured variables in the electronic health record (EHR) of patients with fractures evaluated in the Hasbro Children’s Hospital (HCH) ED and HCH Child Protection Program (CPP) using descriptive statistics, 2) applying natural language processing (NLP) techniques to extract data from clinical narratives and radiology reports to generate text-derived variables, 3) employing machine learning (ML) techniques to identify predictor variables to derive and iteratively refine a CDR, and 4) validating this CDR with a different HCH cohort of patients. The expected immediate outcome of this project is the development of a refined CDR to identify fracture presentations that are highly concerning for abuse among children ≤5 years old. This will inform the design of a prospective multi-center follow-up study for broad validation of CDR’s ability to identify high risk patient presentations, improve real-time clinical detection of potentially abusive injuries, and decrease disparities in clinical decision making.