Multimodal artificial intelligence in retinal vascular and neovascular macular diseases: a systematic review of diagnostic and prognostic applications.

backgroundRetinal vascular diseases, including diabetic retinopathy (DR) and retinal vein occlusion (RVO), and neovascular macular diseases such as neovascular age-related macular degeneration (nAMD) are leading causes of vision loss worldwide. With the rapid growth of artificial intelligence (AI), multimodal approaches that integrate diverse imaging modalities and clinical data have emerged as powerful tools for improving diagnosis, prognosis, and risk stratification.

methodsThis systematic review, conducted according to PRISMA 2020 guidelines, synthesized evidence on the diagnostic and prognostic applications of multimodal AI in retinal vascular diseases. Six databases (PubMed, Embase, Scopus, Web of Science, IEEE Xplore, and Cochrane Library) were searched for English-language studies published between 1 January 2019 and 1 November 2025. Eligible studies applied AI or machine learning models integrating two or more data modalities for diagnosis, prognosis, or prediction in DR, RVO, or AMD. Data extraction, quality appraisal, and narrative synthesis were performed.

resultsFrom 11,659 identified records, 12 studies met the eligibility criteria. Multimodal AI systems consistently outperformed unimodal models and, in several cases, exceeded expert ophthalmologist performance. Diagnostic accuracy for AMD and polypoidal choroidal vasculopathy (PCV) ranged from 87% to 96%, with fusion-based approaches achieving area under the curve (AUC) values up to 0.989. Prognostic models predicting treatment response or recurrence in nAMD and RVO achieved AUCs between 0.972 and 0.980, surpassing both clinician and single-modality baselines. Hybrid and foundation models integrating imaging, clinical, and textual data demonstrated promising results but variable robustness. Most studies were retrospective, single-center, and exhibited moderate-to-high risk of bias, emphasizing the need for larger, prospective, multicenter validation to establish clinical applicability and generalizability.

conclusionMultimodal AI demonstrates superior diagnostic and prognostic performance compared to unimodal models and, in some cases, outperforms expert clinicians in managing retinal vascular diseases. Integrating complementary data sources, such as OCT, fundus imaging, and clinical information, enhances model accuracy and generalizability.