Fig. 7.

Unsupervised exploration of feature representations (single task model). a,b Scatter plots with marginal distributions for the first four modes of PCA applied to feature representations of the CheXpert test data obtained with the single task disease detection model. A random set of 3000 patients is shown with 1000 samples from each racial group. Different types of information are overlaid in color from left to right including presence of disease, biological sex, racial identity, and age. c Scatter plots with marginal distributions for the t-SNE embedding. d Scatter plots with marginal distributions for the logit outputs produced by the model's prediction layer. No obvious patterns emerge for biological sex and race, while we observe a grouping of younger patients aligned with the ‘no finding’ label.