Putting Data in Context to Mitigate Pediatric Inpatient Prescribing Errors

Prescribing errors are common in inpatient pediatrics, resulting in preventable harm to up to 6000 US children per year. These errors occur within 3 overlapping contexts: individuals susceptible to human error, interprofessional teams susceptible to ineffective communication, and systems with inadequate safeguards. However, our understanding of the incidence of prescribing errors and the contribution of these contextual factors is limited because neither prescribing errors nor their contextual factors have been reliably measured or assessed at scale. Our central hypothesis is that adding novel context measures to traditional error prediction measures can enable us to predict prescribing errors with high accuracy. Preliminary work by our group and others have built algorithmic error identification tools using electronic health record (EHR) metadata and identified measures for each of these contexts that are predictive of prescribing errors. Building on this work, this proposal aims to validate the algorithmic approach to detecting pediatric prescribing errors at scale, validate the use of EHR metadata to identify individual, team, and system contextual factors which predict errors, and develop a machine learning model that can predict errors. In the first aim, we will build upon existing wrong-patient order identification tools to validate a broad algorithmic approach to detect pediatric prescribing errors, comparing algorithmic detection of prescribing errors to gold-standard blinded expert prescription review. In the second aim, we will validate the use of EHR metadata to measure team composition and interactions through a comparative focused ethnographic research approach incorporating in-depth observations and interview findings in relation to EHR metadata. Finally, in the third aim we will evaluate individual, team, and system context factors in relation to prescribing errors at three large children’s hospitals, using hierarchical logistic regression to quantify associations and testing the ability of machine learning prediction models to predict prescribing errors using EHR metadata with high accuracy. Through these aims, we expect to identify novel approaches to measuring the inpatient pediatric context and to develop a machine learning model incorporating relevant contextual factors to predict prescribing errors. These findings will be readily generalizable, as they are based on ubiquitously-collected data, enabling a multicenter trial of a transformational approach to clinical decision support, in which context factors inform real- time feedback on high risk prescriptions to prescribers, pharmacists, and nurses caring for the most vulnerable children.