Multimodal AI in retinal disease: from images to decisions and real-world endpoints

Artificial intelligence (AI) has achieved substantial progress in retinal image classification and disease screening, butits clinical value will ultimately depend on whether it can support meaningful decisions across the patient journey. This Opening Editorial introduces the Special Collection "Multimodal AI in Retinal Disease: Imaging, Clinical Data, and Real-World Endpoints" and outlines priorities for the clinical translation of multimodal retinal AI.

Multimodal AI offers an opportunity to integrate fundus photography, optical coherence tomography, OCTangiography, ultra-widefield imaging, and other retinal modalities with clinical characteristics, treatment history, and longitudinal change. Such integration may improve risk stratification, disease monitoring, treatment-response prediction, and individualized follow-up. However, adding modalities or increasing model scale does not in itself establish clinical utility. Future studies should demonstrate incremental value over strong unimodal and clinical baselines, address missing data and distribution shifts, and validate performance across institutions, devices, and diverse patient populations. Evaluation should extend beyond retrospective discrimination metrics to outcomes relevant to patients and health systems, including visual function, referral appropriateness, time to treatment, treatment burden, safety, access, and patient experience. Transparent reference standards, independent external validation, uncertainty assessment, and rigorous reporting are also essential.

The next phase of retinal AI research should move beyond technical performance toward clinically meaningful, externally validated, and patient-centered evaluation. Multimodal AI will have greatest impact when it demonstrates reliable incremental value and supports trustworthy, equitable, and scalable decision-making in real-world retinal care.