fAIrHEALTH: Fair AI models for health?

In 2024, the team of researchers behind fAIrHEALTH developed a new version of ChatGPT, called DelFI, that better understands temporal aspects. These early warnings could help doctors choose the best time and person who should get screening tests, preventative medicines like statins, or limited-supply vaccines to new epidemics.

But the AI is not fair, in that it is less useful for the people who usually experience worse health: ethnic minorities, low income families and people who do not live in big cities. Because most training data come from more affluent, white groups of people, the predictions for others are less exact. If hospitals were to use the tool today, health divides widen.

AI can predict who is at risk of disease but work best for rich, white people. That's not fair. fAIrHEALTH is building the first "fairness map" for medical AI: a clear guide to which social groups and diseases face the biggest risk of unfair treatment, and which fizes work best. By sharing the map, the project will steer developers and health decision-makers toward tools that are fair enough, show how to fix those that aren't, and stop those that still don't meet the fairness mark.

How will it work?

Using secure, linked health and social data about every adult over 40 in Estonia, Finland and Denmark, and Sweden, the researchers will...

• Measure inequality by asking the upgraded DelFI to predict four real problems (breast cancer, colorectal cancer, heart attacks and brain strokes, and death from COVID 19) in each country, and measure where it gets less accurate.
  
• Test and benchmark five new technical solutions that claim to make AI fairer, checking whether they improve fairness and improve the accuracy of predictions for people who are older, female, not born in the country where they live, with no university education, with less than average income, currently unemployed, unmarried, with a disability, or living in the countryside.
• Explore public acceptance for this technology by conducting an online survey of 1,500 people to understand what level of unfairness people accept from AI-based tools.