Evaluating the performance and bias of Electronic Health Record transformers for predicting cardiovascular outcomes following COVID-19

Summary

AI is likely to be a big part of healthcare in the future. This includes dealing with future pandemics. AI models created using patient
data might be unfair. Patient data is collected when people contact doctors for support. This means AI models might not work as well
for people who don’t ask doctors for help. We can’t test if these models are unfair using just patient data, because we don’t know what
is missing. Longitudinal Population Studies (LPS) contain extra data about people's health. This extra data could help us test the fairness
of AI models.
COVID-19 is not well-recorded within patient data but is well-recorded in LPS. Heart attack and stroke are well-recorded within patient
data. We will create an AI model using patient data. We will use the model to predict heart attack and stroke risk. We will see if the
model is worse when COVID-19 is missing. We will then add the missing COVID-19 records using LPS data. We will see if this makes the
model fairer. We will also compare it to an existing non-AI model that doctors already use. This will show how trustworthy AI is when
patient data is incomplete.