Recent advances in medical artificial intelligence have highlighted the potential of large language models (LLMs) in performing general clinical tasks such as medical report generation and question answering. However, their performance in intensive care units (ICUs) remains limited due to the sophisticated nature of the knowledge involved and the interactions among multiple organs. To address these issues, we propose EagleCare, a novel LLM-based Clinical Decision Support System that integrates knowledge graph reasoning to enhance multi-organ assessment and clinical inference in the ICU. EagleCare constructs an ICU-specific knowledge graph (ICU-KG) aligned with patient electronic health records (EHRs), from which our system retrieves clinically relevant reasoning paths. These reasoning paths are then synthesized with patient-specific data through retrieval-augmented generation(RAG) and LLM-based inference to support interpretable and context-aware decisions. This combination of symbolic medical knowledge and neural representation learning improves the reliability and transparency of the decision process. Experimental results on a real-world ICU dataset and two medical QA datasets demonstrate the superiority of our framework, achieving an average improvement of 5.3% compared to baselines with SOTA LLMs. Furthermore, EagleCare assessment outputs exhibit strong alignment with expert clinical reports, indicating its practical utility in high-stakes ICU environments.
Paper
Full text
EagleCare: Augmenting ICU Monitoring with LLMs and Knowledge Graph Reasoning
Semantic Scholar · Computer Science · 2025
Abstract
Recent advances in medical artificial intelligence have highlighted the potential of large language models (LLMs) in performing general clinical tasks such as medical report generation and question answering. However, their performance in intensive care units (ICUs) remains limited due to the sophisticated nature of the knowledge involved and the interactions among multiple organs. To address these issues, we propose EagleCare, a novel LLM-based Clinical Decision Support System that integrates knowledge graph reasoning to enhance multi-organ assessment and clinical inference in the ICU. EagleCare constructs an ICU-specific knowledge graph (ICU-KG) aligned with patient electronic health records (EHRs), from which our system retrieves clinically relevant reasoning paths. These reasoning paths are then synthesized with patient-specific data through retrieval-augmented generation(RAG) and LLM-based inference to support interpretable and context-aware decisions. This combination of symbolic medical knowledge and neural representation learning improves the reliability and transparency of the decision process. Experimental results on a real-world ICU dataset and two medical QA datasets demonstrate the superiority of our framework, achieving an average improvement of 5.3% compared to baselines with SOTA LLMs. Furthermore, EagleCare assessment outputs exhibit strong alignment with expert clinical reports, indicating its practical utility in high-stakes ICU environments.