Photo Credit: iStock.com/yogenyogeny

The following is a summary of “Performance of artificial intelligence-based models for epiretinal membrane diagnosis: A systematic review and meta-analysis,” published in the May 2025 issue of American Journal of Ophthalmology by Mikhail et al. 

Epiretinal membrane (ERM) was known to impair central vision by creating a fibrous layer on the inner retina, and the effectiveness of artificial intelligence (AI)–based tools for its diagnosis needs evaluation. ERM often coexists with or mimics other retinal conditions such as macular degeneration, making accurate diagnosis essential for appropriate treatment planning. 

Researchers conducted a retrospective study to assess the overall diagnostic accuracy of AI models for detecting ERM and to identify factors influencing their performance. 

They performed comprehensive searches in Medline, Embase, Cochrane Library, Web of Science, and preprint databases from inception to June 2024. Studies assessing AI models for ERM diagnosis were included. Study quality and bias risk were evaluated using the Quality Assessment for Diagnostic Accuracy Studies 2 (QUADAS-2) tool. A random-effects model was employed to pool diagnostic accuracy, sensitivity, specificity, and diagnostic odds ratio. Subgroup analyses investigated factors influencing model performance. The study protocol was registered with the International Prospective Register of Systematic Reviews (PROSPERO – CRD42024563571). 

The results showed that studies evaluating AI models for ERM diagnosis varied in quality, as assessed by the QUADAS-2 tool. Pooled diagnostic accuracy, sensitivity, specificity, and diagnostic odds ratios were calculated using a random-effects model. Subgroup analyses revealed several factors affecting model performance. 

Investigators concluded that AI models showed strong diagnostic accuracy for ERM but were limited by inconsistent validation and development methods. The presence of overlapping retinal pathologies, including macular degeneration, underscores the need for robust and diverse training datasets. 

Source: ajo.com/article/S0002-9394(25)00277-6/fulltext