A 2022 study evaluated deep learning models and found the models could predict self-reported race from medical images with high performance: x-ray imaging area under the receiver operating characteristic curve (AUC) range 0.91–0.99, chest CT imaging AUC range 0.87–0.96, and mammography AUC 0.81.
January 01, 2022
high
temporal
Performance quantification of deep learning models across multiple imaging modalities.
A 2022 study found that several hypothesised non-imaging confounders had low ability to predict race in isolation: body-mass index (BMI) AUC 0.55, disease distribution AUC 0.61, and breast density AUC 0.61.
January 01, 2022
high
temporal
Assessment of possible confounding variables tested with regression models.
A 2022 study reported that deep learning models' ability to predict self-reported race from medical images persisted across all anatomical regions and frequency spectrums and remained when images were corrupted, cropped, or noised, indicating robustness that may make it challenging to prevent race prediction in medical imaging AI.
January 01, 2022
high
temporal
Investigation of model robustness to image corruptions and analysis across anatomical regions and frequency spectrums.