Approved Research

Algorithmic fairness and bias in the prediction of cardiometabolic and brain health-related outcomes

Lay summary

This research project aims to answer important questions about the fairness of current risk scores used to predict cardiometabolic (type 2 diabetes, heart disease) and brain-related diseases (neurological disorders, depression, anxiety). We want to find out if these models, which were largely developed in majority populations, work equally well for e.g., minority groups with higher disease burdens or for population subgroups who are otherwise disadvantaged. Additionally, we want to explore the possibility of creating new prediction models that are not only accurate but also fair.

In the first part of the project, we will analyze existing prediction models to see if they are unfair when it comes to predicting diseases. This will help us understand if certain groups are being treated differently or receiving sub-standard care. In the second part, we will develop new prediction models that take into account a wide range of factors, such as demographics, socioeconomic status, diet, lifestyle, and sleep patterns, and other health-related information, such as medication use and disease histories. We will train these new models to perform equally well for different population groups defined by race/ethnicity, sex, and socioeconomic factors like income and education.

The rationale for this research is based on the fact that people with lower socioeconomic status and certain cardiometabolic diseases often face healthcare inequalities, experience more severe complications, and have shorter lifespans. These inequalities not only have a negative impact on individuals but also carry significant costs for governments. By improving the fairness of healthcare processes, we can improve lives and benefit economies.

Currently, many prediction models are developed based on data from majority populations, which may not work as well for minority or marginalized groups. Moreover, if these models are trained on biased data, they can perpetuate existing biases and inequalities.

By developing a comprehensive framework to assess algorithmic fairness and identifying biases in existing models, we can pave the way for new fair models that will help address healthcare gaps. The resulting predictive models will consider performance, fairness, and explainability, and have the potential to replace current risk assessment algorithms and improve preventive strategies for cardiometabolic diseases.

This research project is expected to last for three years and has the potential to make a significant impact on public health by promoting fair and accurate prediction models for cardiometabolic outcomes.