EyeAgent: An Agentic AI System for Multimodal Clinical Decision Support in Ophthalmology

Specialized AI systems outperform general ones in medicine: EyeAgent's diagnostic accuracy improved from 69.71% with only 5 general tools to 80.79% with 53 specialized ophthalmic tools, demonstrating that AI systems designed for specific medical domains are significantly more reliable than one-size-fits-all approaches.

Interpretability and tool selection transparency matter for clinical trust: The system achieved 93.7% accuracy in selecting appropriate diagnostic tools and received expert ratings above 88% for interpretability and reasoning, suggesting that clinicians are more likely to trust AI systems when they can understand why the system made specific recommendations.

AI assistance benefits less experienced clinicians most: Junior ophthalmologists saw the greatest improvement (19% boost in report quality) when using EyeAgent as an assistant, while senior ophthalmologists' performance matched the AI alone, indicating that responsible AI deployment should consider how different user expertise levels interact with AI tools.

Human-AI collaboration can enhance rather than replace clinical expertise: EyeAgent matched or exceeded senior ophthalmologist performance and improved overall diagnostic accuracy by 18.51% when used collaboratively, suggesting that responsible AI implementation focuses on augmenting human decision-making rather than autonomous decision-making.

Comprehensive validation across real-world cases is essential: The system was tested on 200 real-world clinical cases with expert evaluation across multiple dimensions (accuracy, safety, completeness), demonstrating that responsible AI deployment requires rigorous, multi-faceted validation beyond benchmark datasets.