CheXpert is a large dataset of chest radiographs, containing 224,316 images from 65,240 patients, labeled for 14 common observations. The dataset includes uncertainty labels to capture the inherent uncertainties in radiograph interpretation. The authors designed an automated labeler to extract these labels from free-text radiology reports. They investigated different approaches to incorporating uncertainty labels into the training of convolutional neural networks, focusing on five clinically significant pathologies: Atelectasis, Cardiomegaly, Consolidation, Edema, and Pleural Effusion. On a validation set of 200 studies annotated by three board-certified radiologists, the best model outperformed at least two of the three radiologists in detecting four of these pathologies. The dataset is publicly available to encourage further development and evaluation of chest radiograph interpretation models.CheXpert is a large dataset of chest radiographs, containing 224,316 images from 65,240 patients, labeled for 14 common observations. The dataset includes uncertainty labels to capture the inherent uncertainties in radiograph interpretation. The authors designed an automated labeler to extract these labels from free-text radiology reports. They investigated different approaches to incorporating uncertainty labels into the training of convolutional neural networks, focusing on five clinically significant pathologies: Atelectasis, Cardiomegaly, Consolidation, Edema, and Pleural Effusion. On a validation set of 200 studies annotated by three board-certified radiologists, the best model outperformed at least two of the three radiologists in detecting four of these pathologies. The dataset is publicly available to encourage further development and evaluation of chest radiograph interpretation models.