HalluciGuard-eye: knowledge graph-enhanced retrieval-augmented generation for hallucination detection and mitigation in ophthalmic multimodal large language models
Related Experiment Video
Updated: Nov 5, 2026
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
1Ophthalmic Center, The Second Affiliated Hospital, Jiangxi Medical College, Nanchang University, Nanchang, China.
Background:
Diabetic retinopathy (DR) is the leading cause of preventable blindness in working-age adults, yet scalable early detection is limited by the unreliability of AI-assisted screening. Multimodal large language models can interpret fundus photographs and support DR grading, but they frequently hallucinate - fabricating lesions or misclassifying referable DR - which directly threatens timely treatment and prevention of vision loss. We developed HalluciGuard-Eye to make AI-assisted DR assessment trustworthy.
Methods:
HalluciGuard-Eye combines knowledge graph-enhanced retrieval-augmented generation with hierarchical hallucination detection. An Ophthalmic Clinical Knowledge Graph containing 47,832 medical entities and 156,294 semantic relationships was constructed from clinical guidelines and published literature. A dual-stream architecture uses a Visual-Semantic Alignment Module to verify cross-modal consistency and a Clinical Factuality Grounding Network to anchor generated content to reranked evidence. The Hierarchical Factuality Score assesses anatomical, pathological, and clinical-recommendation accuracy. The framework was evaluated on four public ophthalmic imaging benchmarks.
Results:
HalluciGuard-Eye reduced hallucination rates by 42.2% compared with state-of-the-art medical multimodal large language models and achieved 91.6% factual accuracy for clinical entity recognition on EyePACS. For diabetic retinopathy grading, sensitivity was 89.6% and specificity was 92.1%, representing improvements of 2.4% and 2.7%, respectively, over MedRAG. In the component ablation, removing knowledge-graph retrieval lowered EyePACS HFS by 9.4 percentage points. Additional analyses yielded 87.6% critical-evidence recall@20, 89.2% lesion-location accuracy, and 88.9% sensitivity in a patient-clustered cohort.
Conclusions:
These results indicate that knowledge-grounded retrieval and hierarchical factuality assessment can reduce hallucinations while preserving diagnostic performance in ophthalmic multimodal language models.