There is growing evidence that artificial intelligence (AI) technologies like machine learning (ML) can perpetuate or even worsen social inequalities when deployed into real-world settings. This has been demonstrated in many realms, including policing, the court system, banking, social services provision, and there is growing concern the same is true in medicine. At the same time, there has been an outpouring of new AI-based interventions, with a ten-fold increase in the number of Food and Drug Administration (FDA) approvals for AI-based technologies since 2017. However, little research empirically examines the health equity implications of ML-based clinical decision-making tools. One clinical arena in which ML-based tools are already in use is emergency department (ED) triage, as an alternative to the common Emergency Severity Index (ESI) system. Despite its widespread popularity, evidence has shown that ESI-based triage has many problems, including poor acuity discrimination, with up to 50% of patients triaged at the midpoint of the scale, and is associated with racial inequalities, with African-American patients experiencing longer wait-times and lower triage levels controlling for illness severity. This study will use an ML-based ED triage tool that is already in use at a major academic medical center in the United States to explore the extent to which several factors are associated with inequality in predictive performance across patient racial/ethnic groups. This research will take a mixed methods approach to concurrently examine both human and ?machine? elements that affect the triage tool?s final impact on patients.
Aim 1 will be a qualitative study involving ethnographic observation and semi-structured interviewing of triage nurses, to develop a conceptual framework for clinicians? understanding of and interaction with an ML-based tool.
Aim 2 will examine ?label bias?, a type of measurement bias. The Applicant will use synthetic and real electronic health record (EHR) data and simulate different levels of label bias, then examine predictive performance of the triage tool across patient racial/ethnic groups.
Aim 3 will explore different methods for imputing missing EHR data. The Applicant will deploy common, simplistic deletion-based methods as well as a promising new ML-based imputation method called an autoencoder, apply the triage model to generate predictions and examine performance across patient racial/ethnic groups. This project is innovative because it contributes to the development of a ?life cycle? model of ML-based tools and their health equity implications using a mixed methods approach that integrates both human and computational elements, while also providing a rigorous training plan for the Applicant, an MD-PhD student in epidemiology. This training plan is rigorous, synergistic yet diverse, and will include advanced coursework, dedicated 1-on-1 and group mentoring with experts in the field, attendance at seminars and targeted conferences, integration with clinical education and professional development. This project will be an essential step toward the Applicant?s maturation into an independent physician-scientist.
This proposal represents a significant contribution to public health in that it seeks to improve upon current emergency department (ED) triage processes, which are associated with racial inequalities in wait time and triage severity. Specifically, this research will evaluate the performance of an alternative machine learning (ML) -based tool for ED triage across racial/ethnic groups, contributing to a more accurate and equitable triaging of patients and directly improving patient health outcomes. More broadly, this project will enhance the ability of many stakeholders, including data scientists, clinicians, researchers and health policy-makers to evaluate the health equity impacts of any proposed ML-based interventions for health.