Fig. 6. Unfair model performance not resulting from shortcut learning in a dermatology classifier—despite strong attribute-condition correlation.

a Age distribution for examples with Acne (light green) and all other conditions (gray) in the training set. Note the significantly lower mean age for patients with Acne, as would be expected clinically. b AUC vs separation plot for Acne. AUC is binarised by using the score for Acne vs all; separation is calculated on a binarised prediction (top1). c ShorT analysis does not demonstrate that separation is significantly correlated with age encoding in this example (two-sided Spearman correlation). Source data are provided as a Source Data file.