Abstract

Objective: To describe clinical, structural and functional characteristics of primary glaucoma; assess the relationship between disease severity and Cybersight AI outputs; and determine diagnostic performance at the system-provided 20% threshold at baseline and 12-month follow-up.

Methods: This prospective single-center longitudinal observational study with a diagnostic-accuracy component was conducted from December 2024 to August 2026 at Eye Center, Hue Central Hospital. Baseline analysis included 229 eyes from 121 patients; paired 12-month analysis included 192 eyes from 100 patients. The reference standard integrated specialist examination, optic-disc assessment, RNFL optical coherence tomography and Humphrey 24-2 visual fields. Generalized estimating equations and 5,000 patient-cluster bootstrap replicates accounted for inter-eye correlation.

Results: Primary glaucoma was diagnosed in 133/229 eyes (58.1%). At the 20% threshold, sensitivity was 51.9% (95% CI 41.9-61.8), specificity 90.6% (83.8-96.4), PPV 88.5%, NPV 57.6%, accuracy 68.1% and AUC 0.744. AI positivity increased with glaucoma severity. Among paired eyes, sensitivity was 57.8% and specificity 80.7%; AUC did not change significantly.

Conclusion: At the 20% threshold, Cybersight AI showed high specificity but limited sensitivity. Detection was more frequent in eyes with greater structural and functional damage. A negative AI result was insufficient to exclude glaucoma, and the system should not replace specialist assessment.