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

## TL;DR

This paper introduces a multi-agent AI system using large language models to improve neuro-ophthalmic diagnosis and treatment planning by handling diagnostic uncertainty.

## Contribution

The novel contribution is an LLM-based multi-agent framework that preserves diagnostic uncertainty for neuro-ophthalmic screening and referral.

## Key findings

- The multi-agent framework improves robustness over single-model baselines in ophthalmic diagnosis.
- Uncertainty-aware predictions align with clinical decision-making under ambiguity.
- The system produces multi-candidate distributions suitable for triage and monitoring.

## Abstract

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.

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.

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.

Uncertainty-aware, multi-candidate predictions align with clinical decision-making under ambiguity and suggest future work on calibration and knowledge-layer fusion.