Algorithmic fairness in predictive models to eliminate disparities in adverse infant outcomes: A case for race

点击翻译按钮获取中文摘要

PROJECT SUMMARY Non-Hispanic Black infants have twice the rates of low birthweight births as non-Hispanic White infants. As disparities in adverse birth outcomes drive disparities in infant mortality and adverse outcomes across the life course, improving inequities in birth outcomes is a national priority. Despite this longstanding inequity, many public and private payers are unable to address disparities in adverse infant outcomes because of a lack of race/ethnicity data. This K01 fills a critical need for evidence-based recommendations for collection and use of racial/ethnic data among payers to enable population health management programs to develop predictive algorithms that could be used to reduce adverse birth outcomes. Failure to include race/ethnicity in predictive models used for resource allocation may ultimately lead to biased algorithms that exacerbate health disparities. Aim 1 of this study will use an algorithmic fairness framework to test multiple algorithms for developing predictive models for low birthweight birth. In addition to testing model accuracy, predictive models will be tested for seven measures of algorithmic fairness to assess whether having race/ethnicity improves algorithmic fairness (e.g., equal [or better] predictive accuracy for non-White relative to White women) after applying four fairness-enhancing approaches. This project will utilize medical claims, birth certificates, and beneficiary information from the Arkansas All Payer Claims Database. Linkage to the birth certificates uniquely allows this study to have race/ethnicity, which are absent in the commercial claims given lack of collection by most payers. The seminal Institute of Medicine Report Unequal Treatment recommended collection of race/ethnicity to mitigate disparities in health and healthcare delivery; however, it is well known that payers fear accusations of redlining and rarely collect race/ethnicity in most states. Research on payer and provider views regarding collection of race/ethnicity has been conducted, but similar research on the views of minority beneficiaries are severely lacking. Aim 2 of this study will conduct racially-homogenous focus groups among Black, Hispanic, and Marshallese women in Arkansas regarding attitudes on acceptability of collecting and using race/ethnicity data as well as administrative aspects (e.g., when to collect the data), with an emphasis on perinatal programs. These aims will provide an evidence-base and serve as a national model for collecting and using racial/ethnic data with community input. Large third-party payers have the infrastructure to improve health disparities, but lack a community-engaged approach to inform collection and use of these data to guide development of algorithms using an equitable framework. The K01 will allow the investigator to build on her expertise in insurance claims analysis to acquire skillsets in predictive modeling, community engagement, and qualitative methodologies. These important skillsets will allow the researcher to achieve her long-term goals of becoming a productive and independent researcher with a focus on identifying and mitigating factors that serve as drivers of racial/ethnic disparities in adverse infant and maternal outcomes.