Understanding algorithmic fairness for clinical prediction in terms of subgroup net benefit and health equity

Why does fairness in algorithms matter in healthcare?

Doctors are increasingly using computer programs called clinical prediction models to help them make decisions about who is likely to become ill, what a patient’s diagnosis is, who might need extra monitoring, and which patients should get certain treatments.

These tools can be helpful, but they don’t always work equally well for everyone. In some groups (like ethnic minorities or people living in poorer areas), these clinical prediction models can miss illnesses or give less accurate predictions, which can make existing health inequities even worse.

For example: The COAG study showed that some genetics‑based dosing tools for blood‑thinning drugs worked well for people of European ancestry but missed important genetic variants common in African Americans.

What does algorithmic fairness mean for healthcare?

Algorithmic fairness is a way of trying to redress health disparities. It does this by trying to make sure that AI algorithms don’t treat people differently based on racial bias, gender, or income unless there is a valid clinical reason.

Unfortunately, in practice. instead of improving an algorithm’s performance for the groups where it works poorly, algorithmic fairness often leads to worsened performance for the groups where it was working well.

What is “subgroup net benefit”?

We developed a framework called “subgroup net benefit” which could be used to measure fairness in clinical prediction models. The framework asks, “How much real-world clinical value does this model give to each demographic group?”

This means:

• How many people benefit from the model in each group?
• How many people are harmed or missed?
• Does the model help reduce (or accidentally widen) health gaps in society?

Instead of just checking if the maths is fair, subgroup net benefit looks at whether the model is helping people in a fair and useful way.

How the subgroup net benefit framework works in practise – case studies

We tested the framework by using it with two case studies:

1.  A model predicting who might develop type 2 diabetes, which was designed to help decide who should be offered a lifestyle programme to help prevent diabetes.
2. A lung cancer screening algorithm which was used to decide which people should receive a diagnostic scan.

Key takeaways: how can subgroup net benefit help in healthcare?

Subgroup net benefit focuses on clinical impact, not just maths, because treating everyone in the same way doesn’t always result in fairness.

• It gives more help to groups who need it most, which reduces health inequities.
• If a clinical prediction model gives greater benefit to groups with worse health outcomes, this leads to smaller gaps between the groups.
• Even if the model isn’t perfectly equal, it can still make the overall system fairer.

This method allows trade-offs between equity and other healthcare goals, which is important when resources are limited. Subgroup net benefit helps stakeholders and decision makers to see whether a system treats everyone fairly in real‑world medical and social situations, instead of judging fairness only by mathematical formulas.

What does the research recommend?

We suggest that subgroup net benefit should be routinely reported when new clinical prediction models are developed and tested, because it can offer a much clearer picture of how a clinical prediction model affects fairness and health inequality.

Overview

1. Prediction tools in healthcare can unintentionally treat some groups worse than others.
2. Existing fairness methods aren’t ideal because they often don’t go further than lowering performance of the algorithm for advantaged groups.
3. Subgroup net benefit offers a better way: it measures the real benefit each group gets and helps identify whether a model is truly promoting fairness and health equity.
4. We suggest that subgroup net benefit should be routinely reported when new clinical prediction models are developed and tested.