Tackling health inequity using machine learning fairness, AI, and optimization

Qi, Miao
Thumbnail Image
Other Contributors
Dunn, Stanley
Hendler, James A.
McGuinness, Deborah L.
Bennett, Kristin P.
Issue Date
Computer science
Terms of Use
This electronic version is a licensed copy owned by Rensselaer Polytechnic Institute (RPI), Troy, NY. Copyright of original work retained by author.
Full Citation
Health inequity, which leads to unfair and preventable disparities across individuals in opportunities to achieve optimal health, has been brought back into the national spotlight by global COVID-19 pandemic. As artificial intelligence (AI) is increasingly being applied within the health domain, this work aims to develop a new fairness-aware framework, based on machine learning (ML) fairness metrics, AI technologies, and optimization, to help clinical researchers, healthcare providers, and policy makers identify, quantify, reduce, and eventually eliminate potential biases in data-based decision making and implement evidence-based practice to improve patient outcomes. The ultimate goal is to enhance diversity, equity, and inclusion (DEI) in population health in support of better health outcomes for all. We developed a set of health equity metrics to identify and quantify disparities between research sample learnt by AI models and the real-world population that eventual research findings will be applied to. These health equity metrics were derived from existing fairness metrics applied in other areas such as machine learning. Unlike reference-group based metrics measuring bias against a golden truth defined by researchers, these scalable metrics quantify bias against target populations who should have equal opportunity for selection. This research proves that equity metrics could be effectively applied to multiple health domains and shed light on clinical and policy implications. We applied our novel health equity assessment framework, embedded with the proposed equity metrics, to three use cases in population health: randomized clinical trials (RCTs) in Chapter 2, clinical trial recruitment planning in Chapter 3, and healthcare utilization including prescription drugs and vaccines in Chapter 4. To turn health data into usable information that can be understood by observers, we present key equity evaluation results both analytically and visually. In RCTs (Chapter 2), equity metrics, which act as representativeness metrics, enable users to determine overrepresentation, underrepresentation, or exclusion of subgroups with respect to a target population indicating potential limitations of RCTs. Additional statistical tests quantify the significance of observed subgroup inequities with consideration of study sizes and estimation errors of ideal rates. These metrics can measure the level of inequity for all possible protected subgroups of patients defined using multiple protected attributes and provide a single visualization that incorporates and compares these subgroup measures. For clinical trials recruitment planning (Chapter 3), a goal-programming-based multi-objective optimization approach, integrating quantitatively defined enrollment goals, was designed to make equitable enrollment plans for RCTs. The method can prospectively produce equitable enrollment plans in the experiment design stage and retrospectively evaluate inequities in clinical trial enrollment during and after the experiment. It provides opportunities for researchers to demonstrate validity of investigation and to examine disparities across subgroups defined over subjects' characteristics of interest. Furthermore, equity metrics can be used as measures of effects of demographic and socioeconomic determinants on healthcare access and utilization (Chapter 4). They enable users to find differences in healthcare services associated with vulnerable subpopulations such as overprescription and underprescription to medications and insufficient accessibility and utilization of healthcare services. The findings suggest that different determinants exist regarding to the resources/service of different health needs. This method can be valuable assistance in decisions regarding healthcare and provides an opportunity to promote equitable access to healthcare and improved health outcomes. Finally, we developed an interactive web-based R-Shiny prototype toolkit called TrialEquity to address the equity problem of supporting health using our fairness-aware approaches described in the above cases (Chapter 5). To move toward greater health equity, we expect our AI-empowered health equity evaluation framework can be an important and fundamental tool to guide the way.
August 2022
School of Science
Dept. of Computer Science
Rensselaer Polytechnic Institute, Troy, NY
Rensselaer Theses and Dissertations Online Collection
Restricted to current Rensselaer faculty, staff and students in accordance with the Rensselaer Standard license. Access inquiries may be directed to the Rensselaer Libraries.