Sociodemographic Reporting in Artificial Intelligence Studies of Retinal Diseases: A Critical Appraisal of the...

Artificial intelligence (AI) holds promise for retinal disease diagnosis and management, but its application may exacerbate existing disparities in eye health if equity considerations are not addressed.

The lack of sociodemographic diversity in training datasets and poor reporting of participant characteristics may limit the generalizability and fairness of AI algorithms in retinal care. Assessing these gaps is essential for equitable implementation.

A comprehensive literature search was conducted using Ovid MEDLINE, Ovid Embase, and the Cochrane Central Register of Controlled Trials from January 1, 2018, to July 29, 2023. Sociodemographic reporting was evaluated using an equity framework. Study characteristics, geographic diversity, and algorithm performance stratified by demographic subgroups were analyzed.

A total of 360 studies met inclusion criteria, the majority of which were retrospective (79.2%), imaging-based (99.4%) studies using fundus photos (51.9%) or OCT (43.1%) data from online databases (55.6%) or clinical charts (42.2%). Geographic diversity of studies was minimal, with only a few of the studies using clinical chart data that were collected in countries within the African (0.3%), Eastern Mediterranean (5.0%), and South-East Asian Region (6.7%). Sociodemographic reporting was limited, with 30.8% of studies reporting gender or sex, and only 8.9% reporting race or ethnicity of participants. Only 11.1% of studies tested AI algorithm performance by gender or sex, and 1.7% by race or ethnicity. The quality of racial and ethnic data reporting was poor overall, with significant gaps in how data were collected, determined, and classified.

Artificial intelligence-related retinal studies have limited equity considerations, particularly with regards to underrepresentation of diverse sociodemographic groups and insufficient racial and ethnic data reporting. Enhanced diversity in study populations and improved data collection and reporting methodologies are needed to ensure AI algorithms are trained and perform comparably across heterogeneous populations affected by retinal diseases.

The authors have no proprietary or commercial interest in any materials discussed in this article.