Fig. 5. Unfair model performance resulting from shortcut learning in a cardiomegaly classifier.

a The effect of altering the gradient scaling of the binary race prediction head on race encoding (as determined by the race prediction AUROC of subsequent transfer learning). Each dot represents a model trained (25 values of gradient scaling times 5 replicates), with error bars denoting 95% confidence intervals from bootstrapping examples within a model (n = 3818 independent patients). b AUC vs fairness (equalized odds) plot for Cardiomegaly. There exist models that are fairer and as performant as the baseline model (orange dots). c ShorT analysis demonstrates that unfairness is significantly correlated with race encoding in this example (two-sided Spearman correlation, n = 123 technical replicates). Source data are provided as a Source Data file.