NRAG: A Knowledge-Enhanced LLM Framework for Interpretable Neurosurgical Disease Diagnosis in Outpatient and Emergency Settings

Abstract

Large language models (LLMs) have achieved state-of-the-art performance in numerous domains, yet their clinical deployment faces critical barriers, particularly insufficient reasoning in complex scenarios and limited interpretability. These challenges are exacerbated in neurosurgical diagnosis for outpatient and emergency settings, where time-sensitive decision-making, fragmented data, and complex comorbidities render conventional free-text-based modeling approaches unreliable. To address the limitations of existing LLMs in medical auxiliary diagnosis, particularly in interpretability and predictive performance, this study proposed NRAG, an auxiliary diagnosis method that combines LLMs with knowledge graphs (KGs). It extracts symptom descriptions from clinical records and performs personalized retrieval of associated paths in KG, and supplements potential patient symptoms to optimize the diagnosis model. Comparative experiments involving multiple general-domain and medical-domain LLMs, along with case studies, were conducted to validate the NRAG's effectiveness. Experimental results demonstrate that integrating KG significantly improves diagnosis accuracy, achieving an F1-score of 0.8150. It also substantially improves model interpretability and performs excellently in expert evaluations. Ablation studies and comparative experiments with other general-domain and medical-domain LLMs confirm the superior performance of the proposed NRAG. NRAG effectively supplements missing symptom information and provides knowledge-path-based evidence for diagnosis results, while improving the precision and interpretability of intelligent diagnosis. Furthermore, this approach sets the foundation for intelligent diagnoses in neurosurgery while providing a methodological framework for the integration of in-depth clinical data mining with medical knowledge base resources.

Ähnliche Arbeiten