Extended Data Table 2 Distribution shift quantification in the ID settings between demographic groups on MIMIC-CXR

From: The limits of fair medical imaging AI in real-world generalization

  1. a, Prevalence shift P(Y|A) was derived using the total variational distance between the probability distributions of Y conditioned on different groups. P values were computed using a two-sided proportion z-test. b, Representation shift P(X|A) was derived by first encoding input into representations from a frozen foundation model f (that is, MedCLIP79) and then computing the MMD distance with a Gaussian kernel81. P values were computed using a two-sided permutation test using this distance as the test statistic81. All P values were adjusted for multiple testing using Bonferroni correction78.
Back to article page