Ontology-grounded knowledge graphs for mitigating hallucinations in large language models for clinical question...

Large Language Models (LLMs) show strong potential in biomedical informatics but frequently generate hallucinated or factually incorrect responses, limiting their clinical utility. This study aims to develop and evaluate a GraphRAG framework using an ontology-grounded knowledge graph that mitigates hallucinations in clinical question answering.

We designed a domain-specific Resource Description Framework (RDF)/Web Ontology Language (OWL) ontology and knowledge graph using clinical and hospital data from multiple Egyptian institutions. The ontology was integrated with LLMs to enforce structured semantic grounding during question answering. Clinical questions were evaluated under three conditions: (i) baseline ChatGPT-4, (ii) DeepSeek-R1, and (iii) our ontology-grounded framework. Accuracy was evaluated against clinically reported reference answers derived from five peer-reviewed Egyptian hospital studies.

Our GraphRAG framework significantly outperformed baseline models. While ChatGPT-4 achieved 37% accuracy and DeepSeek-R1 achieved 52%, the ontology-grounded approach achieved 98% accuracy (59/60 questions). The hallucination rate was reduced from approximately 63% in ChatGPT-4 and 48% in DeepSeek-R1 to just 1.7% in our framework, representing a relative hallucination reduction of more than 61% relative to state-of-the-art LLMs. The framework further ensured consistent, reproducible answers aligned with clinical knowledge, demonstrating its robustness for healthcare applications.

Ontology-grounded knowledge graphs provide a reliable and verifiable method for mitigating hallucinations in LLM-based clinical question answering. By embedding structured clinical semantics into the reasoning process, the framework enhances factual accuracy, reproducibility, and safety in biomedical informatics. This work highlights the critical role of semantic knowledge representation in building trustworthy Artificial Intelligence (AI) systems for healthcare decision support.