Using Machine Learning to find a life saving needle in a haystack of children's emergencies

SUMMARY Adverse safety events (ASEs) resulting from medical care are a leading cause of preventable injury and death in the United States. The National Academy of Medicine recommends that hospital and Emergency Medical Services (EMS) systems “implement evidence-based approaches to reduce errors in emergency and trauma

care for children,” but acknowledges that implementation is limited by the “paucity of high-quality data on the epidemiology of medical errors in children, particularly within the emergency care system.” Our research team developed and validated an EMS chart review tool to identify ASEs in the care of children and has begun to

describe the epidemiology of these events. We have identified pediatric out-of-hospital cardiac arrest (OHCA) as a particularly high-risk condition for ASEs and poor survival. EMS plays a critical role in the health and outcomes of Americans during cardiac arrests. Receipt of effective treatment in the first few minutes of cardiac

arrest can double or triple survival. However, while survival from adult OHCAs and in-hospital pediatric OHCAs have both increased significantly over the last 10-15-years, survival from pediatric OHCA remains largely unchanged. We focus on identifying preventable ASEs occurring over the entire episode of OHCA, recognized

to be a major contributor to mortality and morbidity. The status quo, manual chart reviews, considered the gold standard for evaluating safety and quality of care, are costly and labor-intensive. The main goal of this proposal is to computationally detect ASEs associated with pediatric OHCA at a population level from electronic EMS

charts through the following Study Aims: Aim 1. Identify adverse safety events in the prehospital care of children with OHCA via rules- and regression-based computational processing of structured data in pediatric EMS charts. Aim 2. Extract cardiac arrest-related indicators from EMS chart narrative text using deep learning

NLP techniques and weak supervision techniques to augment the rules-and regression--based automatic screening of EMS charts. Aim 3. Prospectively demonstrate the scalability of automated detection of ASEs in OHCA at the scale of statewide populations. This proposal leverages the strengths of an experienced

multidisciplinary research team that includes informaticians and clinician-scientists with expertise in pediatric patient safety and American Heart Association Guideline development. Successful completion of the project aims will create the foundational elements of an automated tool capable of screening EMS charts on a large

scale to identify, monitor, and ultimately mitigate preventable pediatric prehospital patient safety events. Additionally, the computational tools and annotated dataset created in the course of this project will serve as valuable infrastructure to support future clinical and computational research.