Evaluation and Improvement of Algorithmic Fairness for COVID-19 Severity Classification Using Explainable AI-based Bias Mitigation

Lay Summary During the COVID-19 pandemic, machine learning models have been used to predict how severe the disease might be for patients, helping doctors decide who needs urgent care. However, these models can sometimes be unfair, giving biased predictions for certain sociodemographic groups. Our study focused on making these models fairer, especially regarding sex differences (male and female). We used data from the Quebec Biobank to build a machine learning classification model to predict COVID-19 severity. We found that the model was less accurate for men than for women, showing a bias. To fix this, we tested four methods to reduce bias, including a new approach using Explainable Artificial Intelligence (XAI), which helps understand why the model makes certain predictions. Our XAI method identified key factors that affected predictions differently for men and women. By accounting for these differences, we significantly reduced bias while maintaining high model accuracy. This work shows that fairer machine learning models can help ensure equal treatment for all patients, improving healthcare decisions.