Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain
the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in
Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles
All three sets were sampled from the train or validation set of ChestX-ray14 dataset9. The training set was used to train data Shapley algorithm and compute Shapley values, the validation set was used to compute the predictor performance score during training, and the held-out test set was used to report the final results. Because the distribution of pneumonia labels in the ChestX-ray14 dataset is highly imbalanced, we sampled a larger proportion of pneumonia cases in the training set and sampled balanced validation and held-out test sets in this study.