Existing chest radiograph datasets, such as CheXpert and ChestX-ray14, havedriven the development of new machine learning approaches to achieve expert ornear-expert level performance on a variety of tasks. The primary focus ofmodels developed using these datasets has been to replicate human-levelperformance by training on labels computationally extracted from radiologyreports. We propose a different paradigm: pair an existing diagnostic testwith labels from a more accurate, higher fidelity diagnostic test. Thisapproach seeks to ask whether data from a cheaper, lower fidelity diagnostictest contains information for detection of pathologies using more accurate,gold standard labels. In the context of chest X-rays, a good example is theradiologic comment of cardiomegaly, a catch-all term or an abnormally enlargedheart. Cardiomegaly is known to be poorly predictive of cardiac disease anddoes not trigger meaningful clinical action. Instead, we can pair chest X-rayswith gold standard structural heart disease labels derived fromechocardiograms conducted on the same patients. This resource contains 71,589unique chest X-rays from 24,689 different patients paired with keyechocardiography measurements indicative of left ventricular hypertrophy anddilated left ventricle, pathologies which occur during early stage heartfailure. The data also includes information about the relative times of thechest X-rays, the age/sex of the patient at the time of recording, and relatedmetadata information. This data can be used as a resource for the community tobuild novel approaches to detect clinically actionable labels.

Existing chest radiograph datasets, such as CheXpert and ChestX-ray14, havedriven the development of new machine learning approaches to achieve expert ornear-expert level performance on a variety of tasks. The primary focus ofmodels developed using these datasets has been to replicate human-levelperformance by training on labels computationally extracted from radiologyreports. We propose a different paradigm: pair an existing diagnostic testwith labels from a more accurate, higher fidelity diagnostic test. Thisapproach seeks to ask whether data from a cheaper, lower fidelity diagnostictest contains information for detection of pathologies using more accurate,gold standard labels. In the context of chest X-rays, a good example is theradiologic comment of cardiomegaly, a catch-all term or an abnormally enlargedheart. Cardiomegaly is known to be poorly predictive of cardiac disease anddoes not trigger meaningful clinical action. Instead, we can pair chest X-rayswith gold standard structural heart disease labels derived fromechocardiograms conducted on the same patients. This resource contains 71,589unique chest X-rays from 24,689 different patients paired with keyechocardiography measurements indicative of left ventricular hypertrophy anddilated left ventricle, pathologies which occur during early stage heartfailure. The data also includes information about the relative times of thechest X-rays, the age/sex of the patient at the time of recording, and relatedmetadata information. This data can be used as a resource for the community tobuild novel approaches to detect clinically actionable labels.