Toward Multimodal Conversational AI for Age-Related Macular Degeneration.

PURPOSE: To evaluate OcularChat, an age-related macular degeneration (AMD)-specific multimodal large language model for interpreting color fundus photographs. METHODS: A general-purpose multimodal large language model was fine-tuned using 705,850 simulated patient-physician dialogues paired with 46,167 AREDS images, then tested on separate AREDS and AREDS2 datasets. RESULTS: In AREDS, OcularChat correctly classified advanced AMD, pigmentary abnormalities, and drusen size in 95.4%, 84.9%, and 67.8% of images, respectively. Retina specialists rated its responses more highly than those of the same model without fine-tuning. CONCLUSIONS: OcularChat gives the potential to support clinician-supervised, interpretable AMD image review, research annotation, and education, requiring prospective validation before clinical deployment.