University of Minnesota surgeon Chris Tignanelli, MD, MS, collaborated with researchers to determine if artificial intelligence (AI) can predict self-reported race from chest x–ray images.
In previous research, investigators have demonstrated that neural network models are able to predict a patient’s self-reported race from their medical images. This finding baffled some researchers who claim there are no known features in medical images that can be used to determine race. These same researchers said there might be a built-in racial bias in collecting the information from the x–rays.
Turns out these researchers were wrong.
To answer how the AI-interpreted x–rays so precisely identified race, the researchers counted the number of pixels of each grayscale value, or pixel intensity, in a large dataset of chest x–rays. They then performed a statistical test to see if different racial groups have different pixel intensity counts and trained machine learning models to predict self-reported race based on these counts.
Turns out there are clear racial differences despite claims by social anthropologists that ‘race’ is a purely social construct. Their statistical test shows that there is a significant difference in pixel intensity counts among different racial groups. The models they developed were also able to predict someone’s self-reported race based on their count, but they were less accurate than neural networks developed in previous studies that based their predictions on entire medical images.
This research demonstrates that “race” is not purely a social construct but is embedded in medical images in ways that are not always apparent to human observers.
