🔗 RadLink Dataset

RadLink is a medical text entity normalization dataset containing 1250 unique entities extracted from the RadGraph dataset.

RadLink focuses on the crucial task of normalizing entities extracted from radiology reports by linking them to standardized concepts in the Unified Medical Language System (UMLS) with their corresponding Concept Unique Identifiers (CUIs).

The dataset was created by DBIS at RWTH Aachen University (Yongli Mou, Hanbin Chen, Prof. Dr. Stefan Decker) and University Hospital RWTH Aachen (Gwendolyn Isabella Lode, Prof. Dr. Daniel Truhn).

📊 Data Source and Processing

• Data source: Unique entities extracted from the RadGraph training set
• Processing method: Annotation and normalization through three steps:
  1. Morphological Matching 🔍: Using UMLS API to retrieve potential matches based on string similarity algorithms
  2. Semantic Matching 🧠: Employing LLMs (like GPT-4) to select the best match from top candidates
  3. Expert Validation ✅: Final review and verification by medical students from University Hospital RWTH Aachen

🧩 Data Structure

The dataset contains the following fields:

• name 📋: Original entity name
• ui 🔑: UMLS Concept Unique Identifier (CUI)
• normalized_name ✨: Normalized entity name
• semanticTypes 🏷️: Semantic type classification

💡 Application Scenarios

This is a sub-project of MedKGC, focusing on Named Entity Normalization (NEN) tasks. We built this dataset to evaluate the ability to normalize entity names in medical knowledge graphs. Related project:
MedKGC NEN task

📝 Citation

If you use the RadLink dataset, please cite this project as well as the original RadGraph dataset:

@misc{RadLink,
  title={RadLink: A Medical Entity Normalization Dataset},
  year={2025},
  howpublished={https://huggingface.co/datasets/WestAI-SC/RadLink}
}

For more information about RadGraph, please refer to the original paper RadGraph: Extracting Clinical Entities and Relations from Radiology Reports.