Racial Bias in a VA Algorithm for High-Risk Veterans

PROJECT SUMMARY African-American Veterans are at particular risk of adverse outcomes, including mortality and hospitalization, due to adverse social determinants of health (SDoH) including poor transportation access and housing instability. Identifying individuals at risk of adverse outcomes has been a priority at the Veterans Health

Administration (VA), which has implemented novel predictive analytic tools in clinical care settings to target care resources efficiently and equitably. The VA has invested an average of 5% of total VA spending towards health information technology to support such algorithms. One predictive algorithm implemented nationwide

and commonly used by VA clinicians is the Care Assessment Needs (CAN) score, which predicts risk of future hospitalization and/or death for over 5 million Veterans receiving primary care. The CAN score is currently used by patient-aligned care teams (PACTs) and nurse care navigators to direct clinical programs and

resources, including telehealth, palliative care, and home-based primary care, to high-risk Veterans. The CAN score is primarily based on laboratory, demographic, utilization, and other administrative data. Recent studies have shown that similar algorithms used in non-VA settings may mischaracterize risk for

vulnerable patient subgroups – including African-Americans – whose health is heavily influenced by disproportionate exposure to adverse SDoH. Importantly, race and SDoH are not routine inputs into the CAN score. There is a growing concern that algorithms like the CAN score could generate “algorithmically unfair”

predictions that systematically mischaracterize risk for subgroups – particularly African-Americans – whose care is heavily influenced by SDoH. However, there has been no systematic investigation into unfairness of the CAN score between African-American and White Veterans. In this project, we will systematically examine algorithmic unfairness in the VA CAN algorithm and develop

approaches to mitigate it, including testing the incorporation of SDoH metrics. Our preliminary investigations into the CAN score show that it underestimates risk for African-Americans compared to White Veterans, which may lead to fewer referrals of high risk African-American Veterans to clinical programs. In Aim 1, we will

develop methods to mitigate algorithmic unfairness in the CAN score using its existing variables. In Aim 2, we will incorporate race and select metrics of SDoH that are available through VA screening efforts into the CAN score to improve algorithmic unfairness. In Aim 3, we will use the “Fair” CAN score generated in Aim 2 to

investigate how mitigating unfairness would change the racial composition of Veterans enrolled in clinical programs targeted at high-risk Veterans.