LLM-Based Multi-Agent System for Neuro-ophthalmic Diagnosis and Personalized Treatment Planning

WENMIAO WANG^*

• National University of Singapore College of Design and Engineering, Singapore, Singapore

We present an LLM-based multi-agent framework for neuro-ophthalmic decision support. While our experiments focus on retinal diseases, the architecture is motivated by established links between ocular pathology and nervous-system processes, enabling descriptions and outputs that are actionable for neuroscience-oriented screening and referral. The system separates data ingestion, diagnosis, and planned treatment planning into coordinated agents. A multi-LLM ensemble—optionally combined with a CNN image classifier—produces probability distributions over common retinal conditions, retaining multiple differential hypotheses rather than a single label. Two fusion strategies are implemented: a rank-based baseline (baseline(1/r)) and a reliability-plus-entropy aggregator calibrated on a small subset of training data. On a 200-case ophthalmic dataset comprising authentic and partially synthesized records, the aggregator improves top-1 diagnostic accuracy from single-model baselines of 22–56% to approximately 78%. The design supports downstream personalization: multi-candidate outputs and calibrated uncertainties can be mapped to neuro-relevant triage, monitoring, and treatment pathways. Limitations include dataset size and partial synthesis, and the current absence of integrated Bayesian/knowledge-graph reasoning. Overall, the framework illustrates how uncertainty-aware, multi-candidate diagnosis in ophthalmology can serve neuroscience goals such as early risk stratification and referral without additional data collection.