A multimodal tool-augmented agent for evidence-based clinical decision support in longitudinal AMD follow-up

Clinical management of age-related macular degeneration (AMD) requires longitudinal decisions on disease activity, treatment response, and anti-VEGF injection intervals. These decisions depend on integrating multimodal retinal imaging findings across visits, yet manual synthesis is time-consuming and may omit relevant information, limiting transparency and consistency in follow-up planning. Large language models (LLMs) have been explored to summarize longitudinal AMD information, but existing approaches may lack explicit quantitative grounding and evidence attribution, reducing their auditability in clinical decision support. To address these limitations, we developed AMDagent, an evidence-grounded clinical agent that integrates quantitative imaging analysis with guideline retrieval to generate structured, clinician-facing decision-support outputs. The system links longitudinal imaging findings to actionable management considerations while preserving traceable evidence sources. In a 20-case proof-of-concept evaluation of curated longitudinal AMD cases, AMDagent was rated clinically appropriate in 92.0% of predefined decision-support units under expert adjudication. In addition, 83.3% of audited major clinical claims were explicitly grounded in tool-derived quantitative measurements and/or retrieved guideline evidence, enabling clinician verification. These preliminary findings suggest that evidence-grounded, clinician-facing agents may support more transparent and auditable longitudinal AMD decision support, although larger prospective multi-center studies are needed to assess generalizability and clinical utility.