Echo Chambers: Bias and Representation in Cardiac Imaging Datasets for Artificial Intelligence

We examined the landscape of publicly available cardiac imaging datasets to assess how their distribution and construction shape bias and equity in cardiovascular AI.

Across 38 publicly available echocardiography, cardiac magnetic resonance imaging, and cardiac computed tomography datasets, nearly 80% originate from high-income countries, with no public cardiac imaging datasets from Africa or South America. Fewer than two-thirds of these public datasets report any demographic information, and only a small minority are linked to clinical outcomes. Studies reveal consistent gaps in model performance across racial and ethnic groups, substantial variability in labeling and disease definitions, and strong evidence that harmonization and label quality often improve performance more than changes in model architecture. Cardiovascular AI developed using publicly available cardiac imaging datasets is largely constructed on geographically narrow, demographically uncharacterized data from well-resourced health systems. This creates an "echo chamber" that may limit representation of the global majority within publicly available data resources. While these findings are specific to publicly available datasets and may not necessarily extend to private or commercial data resources, they highlight the importance of considering equitable data infrastructure, prioritizing globally representative outcome-linked datasets and local capacity building.