LLM-based multi-agent system for neuro-ophthalmic diagnosis and personalized treatment planning.

Introduction: Ophthalmic findings can non-invasively reflect nervous-system status. We present an LLM-based multi-agent framework that preserves diagnostic uncertainty to support neuro-ophthalmic screening and referral. Methods: Heterogeneous inputs (clinical text/PDFs and optional fundus/OCT images) are normalized by an Information Collection Agent. A Diagnosis Agent ensembles multiple LLMs and, when available, a CNN image branch; outputs are aggregated with an uncertainty-aware fusion. Results: Across a curated ophthalmic corpus, the multi-agent framework improves robustness over single-model baselines and produces multi-candidate distributions suitable for downstream triage and monitoring. Discussion: Uncertainty-aware, multi-candidate predictions align with clinical decision-making under ambiguity and suggest future work on calibration and knowledge-layer fusion.