Approved Research

Studying the relationship between fairness and generalisability to develop fairer clinical prediction modelling for lung cancer and diabetes prognosis

Lay summary

Is artificial intelligence (AI) fair? Every year, more and more AI tools are developed to improve our standard of care. Amongst these, clinical prediction models use information like age or blood pressure, to calculate a patient's risk of having a disease or developing it in the future. This can be a valuable tool for doctors to tailor a patient's treatment or refer them for screening.

These models learn from sources like GP records or medical studies. Yet, they can sometimes carry over prejudices currently existing in healthcare. For example, patients from more deprived areas tend to receive worse care compared to those from wealthier ones. A model could learn to mimic this and incorrectly prioritise the treatment of richer patients, or retain less information about poorer ones. This bias is present in most real-world data, so researchers have developed ways to detect it and ensure that their predictions benefit all patients.

Are these 'fairness tools' always useful? This isn't very well understood. Defining when a model's use of patient's characteristics (like age or smoking status) is fair or unfair isn't clear either. In this project, we explore these questions and extend the researcher's toolkit for tackling them.

In particular, we will create a framework in which the preferences of both doctors and patients are taken into account when deciding which models are fairer than others. The goal of each model will then be to prioritise the health and wellbeing of all sectors of the population. We will use this framework to study how models trained in some sections of the public (like those from wealthier areas of the UK) perform in the rest of the population, and how this relates to fairness. This will finally allow us to develop newer techniques that make models learn relationships that work better on disadvantaged patients.

This project will take 36 months and will help us better understand how we need to develop and test models to make sure that they are fair and trustworthy. We will focus our efforts on clinical prediction models used in lung cancer screening and diabetes prevention.