An Explainable Ophthalmic Treatment Recommendation System via Knowledge-Graph-Guided Reasoning with Large Language Models

LLM hallucination in clinical decision support: Deployment of LLMs in ophthalmic decision support is hindered by hallucinated reasoning.

Limited interpretability in medical LLMs: LLM-based ophthalmic decision support faces limited interpretability.

High computational cost of medical LLMs: LLM deployment for ophthalmic decision support is constrained by high computational cost.

Knowledge-guided LLM reasoning with ophthalmology KG: The paper proposes a knowledge-guided reasoning framework integrating LLMs with an Ophthalmology Knowledge Graph for ophthalmic decision support.

Inference Ophthalmology Graph beam-search reasoning: The paper designs an IOG mechanism in which an LLM-based agent explores graph entities and relations via beam search to find high-confidence reasoning paths.

Structured KG evidence-guided answer generation: Retrieved entity-relation-entity triples are used as structured evidence to guide answer generation.

Ophthalmology knowledge graph and clinical QA benchmark: The paper constructs a large ophthalmology knowledge graph and curates real-world clinical question-answer pairs for evaluation.

Large ophthalmology KG constructed:

Clinical QA evaluation set curated:

Improved over direct LLM baselines: