TLDR: CliCARE is a new framework that enhances Large Language Models (LLMs) for clinical decision support in cancer care. It addresses key challenges like processing long patient records, reducing AI ‘hallucinations,’ and improving evaluation reliability. CliCARE achieves this by transforming unstructured Electronic Health Records (EHRs) into Temporal Knowledge Graphs (TKGs) and aligning these patient trajectories with clinical guidelines. Validated against expert oncologists, CliCARE significantly outperforms existing methods, enabling LLMs to provide more accurate and reliable clinical summaries and recommendations.
Large Language Models (LLMs) show great potential for transforming clinical decision support and easing the burden on doctors, especially when dealing with complex, long-term cancer patient records. However, using LLMs in critical medical fields faces several hurdles. These include their difficulty in processing very long and often multilingual patient records for accurate time-based analysis, a high risk of generating incorrect or ‘hallucinated’ clinical information because traditional methods like Retrieval-Augmented Generation (RAG) don’t fully incorporate clinical guidelines, and unreliable evaluation methods that make it hard to validate AI systems in oncology.
To tackle these significant challenges, researchers have introduced CliCARE, a novel framework designed to ground Large Language Models in clinical guidelines for decision support using longitudinal cancer Electronic Health Records (EHRs). The core idea behind CliCARE is to transform unstructured, multi-year EHRs into patient-specific Temporal Knowledge Graphs (TKGs). These TKGs are crucial for capturing the long-range dependencies and the dynamic evolution of a patient’s disease course.
How CliCARE Works
CliCARE operates in two main stages to ensure robust and evidence-grounded decision support:
1. EHR-to-TKGs Transformation: This initial step converts raw, unstructured patient records into structured Temporal Knowledge Graphs. It involves an efficient context processing pipeline that summarizes extensive historical records and the most recent clinical notes. Key clinical facts, such as diagnoses, treatment regimens, and biomarker trends, are extracted and organized into discrete events with precise timestamps. These events are then linked to a standardized biomedical knowledge graph, creating a personalized TKG for each patient. This structured representation helps LLMs overcome the challenge of processing vast amounts of temporal data.
2. Trajectory-Guideline Alignment: This stage is vital for integrating real-world patient data with established medical knowledge. CliCARE aligns the patient’s descriptive TKG with a prescriptive guideline knowledge graph (Gg), which is built from authoritative clinical practice guidelines like NCCN. This alignment is achieved through a multi-step process: initial semantic matching using a BERT model, followed by an LLM-based reranking to ensure clinical plausibility, and finally, an iterative expansion process to enhance coverage and accuracy. This fusion creates a robust, evidence-fused knowledge representation that directly informs the LLM’s generation of clinical summaries and actionable recommendations.
Also Read:
- HypKG: Integrating Patient Data with Medical Knowledge Graphs for Enhanced Healthcare Predictions
- Beyond the Buzz: Understanding Large Language Models in Medicine
Rigorous Evaluation and Promising Results
The CliCARE framework was rigorously validated using large-scale, longitudinal data from a private Chinese cancer dataset (CancerEHR) and the public English MIMIC-IV dataset (MIMIC-Cancer). A key aspect of its validation was a human-validated LLM-as-a-Judge protocol, co-designed with senior oncologists. This protocol assessed factual accuracy, completeness, clinical soundness, and actionability, demonstrating a strong correlation (Spearman’s rank correlation of approximately 0.7) with expert oncologist assessments, thus providing a reliable and scalable evaluation method.
CliCARE significantly outperformed strong baselines, including leading long-context LLMs and Knowledge Graph-enhanced RAG methods. When paired with powerful models like Gemini 2.5 Pro, CliCARE achieved impressive scores for clinical summaries and recommendations, especially on complex datasets. The framework’s ability to structure chaotic, longitudinal patient records proved crucial, leading to substantial performance gains for various LLMs. The research also highlighted that the full, integrated CliCARE framework is essential for optimal performance, particularly on complex datasets, and that it enables advanced models to leverage richer context for enhanced reasoning across all record lengths.
In conclusion, CliCARE represents a significant advancement towards deploying trustworthy AI in clinical practice by providing reliable clinical decision support. For more details, you can refer to the original research paper: CliCARE: Grounding Large Language Models in Clinical Guidelines for Decision Support over Longitudinal Cancer Electronic Health Records.
