Artificial intelligence for anti-VEGF management in neovascular age-related macular degeneration: from predictive models toward clinical decision agents

Neovascular age-related macular degeneration remains one of the leading causes of irreversible central vision loss in older adults and commonly requires long-term, repeated intravitreal anti-vascular endothelial growth factor therapy. Although anti-VEGF treatment has substantially improved visual outcomes, real-world management is still limited by undertreatment, overtreatment, poor adherence, irregular follow-up, and persistent uncertainty regarding retreatment timing and interval extension. In this setting, artificial intelligence (AI) has rapidly evolved from image recognition toward clinically relevant prediction and decision support. The objective of this narrative review is to examine how current AI literature supports anti-VEGF management in neovascular age-related macular degeneration, where the strongest evidence currently exists, and where claims regarding decision agents remain premature. Deep learning models can now detect and quantify key optical coherence tomography biomarkers, including intraretinal fluid, subretinal fluid, pigment epithelial detachment, hyperreflective foci, and ellipsoid zone disruption. When combined with structured clinical variables and longitudinal follow-up data, machine learning and generative models have shown promise in predicting treatment response, recurrence, persistent activity, injection burden, and tolerated treatment intervals. However, most published studies remain retrospective, single-center, and prediction-focused, with limited external validation and little direct evidence for clinically deployable closed-loop decision systems. In this narrative review, we synthesize the literature across four linked domains: image interpretation, biomarker quantification, prediction of treatment-related outcomes, and agent-enabling technologies such as home monitoring and language-model-assisted decision support. We further distinguish established evidence from future-oriented concepts, discuss methodological limitations including heterogeneity of endpoints and limited reproducibility, and outline cautious directions for translation.