Algorithmic fairness and health equity in clinical prediction

Clinical prediction models are increasingly used to guide decisions in healthcare, such as identifying patients for screening or preventive treatment. However, there are growing concerns that these models may not work equally well for all groups, potentially reinforcing existing health inequalities.

Traditional approaches to algorithmic fairness often focus on mathematical definitions, such as ensuring equal performance across groups. In practice, these approaches can be limited, sometimes reducing performance for some groups rather than improving outcomes for those who are disadvantaged.

This research introduces a different approach: subgroup net benefit, which assesses how much real-world clinical benefit a prediction model provides to different groups. Rather than focusing only on statistical fairness, it evaluates how benefits are distributed across populations and whether a model helps reduce health inequalities.

The study demonstrates how this approach can be applied using examples, including a diabetes prevention model and a lung cancer screening tool. It also highlights that improving fairness may involve trade-offs, particularly when healthcare resources are limited.

Overall, the research argues that subgroup net benefit offers a more meaningful way to assess fairness in healthcare by focusing on clinical impact and equity, and recommends that it should be routinely reported when developing and evaluating prediction models.

Read more on the THIS Institute research pages
Read the full paper