We hope that our article provides a helpful overview of algorithmic fairness debates in healthcare. Please engage with us with any comments or questions!

We conclude by arguing for an alternative framework for the design of equitable algorithms that moves beyond scrutinizing narrow statistical metrics and instead foregrounds health outcomes and utility and clarifies important trade-offs.

These concerns are not unique to the case of lung cancer and apply to the other case studies we discuss in the article, including VBAC calculators, CVD incidence and mortality models, kidney function (eGFR) equations, and healthcare need prediction models.

Using lung cancer screening as an extended case study, we unpack these four categories of fairness concerns and discuss popular approaches for addressing them. Ultimately, we show that these approaches, if deployed, may in fact WORSEN outcomes for individuals across all groups.

For each algorithm, we organize the fairness concerns into a taxonomy of four broad categories:
1️⃣ Inclusion/exclusion of race and ethnicity as inputs
2️⃣ Unequal decision rates across groups
3️⃣ Unequal error rates across groups
4️⃣ Label bias

annals.org/aim/article/...

We hope that our work underscores the importance of foregrounding not only improvements in accuracy, but changes in *decisions and utility* in considering the use of race and ethnicity clinical decision-making.

Our study comes with several important caveats. Notably, in resource-constrained settings (e.g. organ transplants), race-aware models are expected to offer more substantial utility gains.

As a result, the overall clinical utility of race-aware models is surprisingly small. Context matters, but the benefits of race-aware models have likely been overstated.

Further, these individuals also experience modest gains in utility from the use of a race-aware model. This is because, in shared decision-making contexts like the ones we consider, the utility of intervention is 0 at the decision threshold.

Yet, despite this miscalibration, clinical decisions (e.g., screening or treatment recommendations) differ between race-aware and race-unaware models for only a small fraction of individuals (~5%). The individuals whose decisions flip are those closest to the decision threshold.

Using cardiovascular disease, breast cancer, and lung cancer as case studies, we show that race-unaware models are often miscalibrated—underestimating risk for some groups and overestimating it for others. This finding is consistent with evidence cited in support of the use of race-aware models.