Meta-analysis showed that artificial intelligence models achieved high diagnostic accuracy in glaucoma, with a pooled area under the curve of 0.93. Studies demonstrated that artificial intelligence models may help predict visual field deterioration years in advance and support surgical decision-making.
In a systematic review conducted using Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines and registered in the International Prospective Register of Systematic Reviews, investigators evaluated studies published between 2010 and 2025 assessing artificial intelligence (AI) applications in glaucoma. They searched the PubMed, Embase, Scopus, and Cochrane Library databases and included original studies reporting diagnostic, predictive, or clinical decision support outcomes. Data extraction included study design, patient population, imaging modality, and performance metrics, with diagnostic accuracy studies undergoing random-effects meta-analysis.
A total of 42 studies met the inclusion criteria and were included in the qualitative synthesis. Among these, 13 cohort and observational studies were considered central to AI-guided surgical planning applications. Across studies contributing to quantitative synthesis, data sets encompassed more than 75,000 images and over 10,000 patients from retrospective, prospective, and multicenter cohorts.
Diagnostic models—primarily convolutional neural networks and deep-learning systems—achieved area under the curve values ranging from 0.91 to 0.95. Multimodal models integrating fundus photography, optical coherence tomography (OCT), and visual field data demonstrated the highest performance, reaching an area under the curve of 0.95 vs. 0.91 for single-modality systems.
Meta-analysis of four diagnostic studies including more than 64,000 fundus and OCT images showed a pooled area under the curve of 0.93. Sensitivity analyses excluding retrospective single-center studies didn't materially alter the pooled estimate. Subgroup analysis showed superior accuracy with multimodal approaches compared with single-modality fundus or OCT systems.
Progression modeling studies showed that machine learning and deep-learning approaches outperformed traditional statistical methods in predicting visual field deterioration. Several models forecasted progression up to 4 years in advance, enabling identification of fast progressors and facilitating timely surgical decision-making.
Risk stratification and treatment prediction models incorporating electronic health records and multimodal data demonstrated the ability to predict the likelihood of requiring surgical intervention and to distinguish patients suitable for continued medical therapy compared with surgical escalation. Structural imaging tools using deep-learning enabled segmentation of the optic nerve head and lamina cribrosa, supporting presurgical assessment.
Implementation studies indicated that integration of AI into clinical workflows is feasible, with decision support systems supporting referral pathways and surgical triaging in real-world settings.
Methodologic quality across the studies was moderate to high, although limitations included reliance on retrospective designs, heterogeneity in methodologies, and limited external validation. Just a small subset of studies contributed to meta-analysis, and variability in reporting and reproducibility was noted, with incomplete access to code and data sets.
“AI holds transformative potential to optimize decision-making and improve surgical outcomes, marking a paradigm shift toward data-driven, individualized glaucoma care,” wrote lead study author Vidhya Verma, of Ophthalmology at the All India Institute of Medical Sciences, and colleagues.
The study authors reported no conflicts of interest.
Source: Cureus
