Previous research has suggested that AI can identify a patient’s age, sex and even race from medical images. This study tested whether this holds for echocardiograms, ultrasound videos of the heart, using more than 530,000 videos from two US medical centres, Cedars-Sinai and Stanford.
The models predicted sex well and age with an average error of about 9 years, and these results held across both centres. Race, however, could not be reliably predicted. When the researchers changed the balance of age or sex in the data used to train the race model, its performance varied widely, while changing the racial composition of the data had little effect on the predictions of sex and age.
The authors conclude that much of the model’s apparent ability to predict race comes from “shortcuts” through other features that differ between racial groups, rather than from race itself. They warn that treating AI predictions of race as evidence of biological differences risks reinforcing a false idea that has contributed to medical errors, since differences in health between racial groups often reflect social factors and structural racism. They call for further research into which imaging features carry demographic information, and into the risks of bias this creates in medical AI.
Learn more about this study here: https://doi.org/10.1038/s41746-022-00720-8
Reference:
Duffy, G., Clarke, S.L., Christensen, M., He, B., Yuan, N., Cheng, S., & Ouyang, D. (2022). Confounders mediate AI prediction of demographics in medical imaging. npj Digital Medicine, 5, 188.