Fig. 7. Learning tasks and fairness metrics.

a Outline of the four learning tasks described in this paper. For each task, a feature extractor backbone (box) is used in conjunction with one or more prediction heads (circles). Clinical and Attribute Prediction tasks use a single head only; Attribute Transfer tasks use a single head with a frozen feature extractor previously trained for a clinical task. Multitask prediction models use both clinical and attribute prediction heads. b Logistic Regression (LR) fit for fairness metrics. In this example, a logistic function was fitted to the predictions of the model for examples with the condition. The gray distributions represent counts of true positives (top) and false negatives (bottom) across age (x-axis). The LR model fits the probability of a true positive as a function of age (TPR, right y-axis in yellow).