This project was developed as part of the AGI Agent Application Hackathon.
It aims to solve a critical bottleneck in data standardization for healthcare systems.
In particular, standardized data is essential not only for accurate insurance claims but also for securing competitive, high-quality datasets in the AI era.
At institutions like Severance Hospital, a dedicated Data Services Team is working on structured data, but unstructured clinical texts such as free-text descriptions remain largely untouched.
This project addresses that gap by enabling the standardization of radiology free-text reports.
RADRAG is a Retrieval-Augmented Generation (RAG) based tool that standardizes free-text radiology reports into SNOMED CT concepts. It is designed for use in clinical settings, integrating external terminology knowledge and extraction models to enable precise concept mapping.
- ✅ Real-time EMR Integration: Enables real-time standardization of clinical notes directly connected to EMR systems, helping to break down data silos.
- ✅ Batch File Standardization: Supports one-click standardization of existing unstructured report archives.
- Frontend: React + Tailwind + Typescript
- Backend: Flask
- Others: Upstage API, Lanchain, sentence-transformer model, Vite, Vercel, cloudtype
📁 project-name/
├── frontend/
├── backend/
├── rag/
├── README.md
- Download the SNOMED CT International version from the UMLS website.
- Registration and license approval are required.
- Once downloaded, store the vocabulary files in the
data/directory.⚠️ The SNOMED CT files are not included in this repository due to licensing restrictions.
SNOMED CT is hierarchical by design. To enable effective embedding and search, a flat version of the terminology is needed:
python process_data.py make-flattened-terminologyThis step creates a dictionary file containing terms related to the flattened concept list:
python process_data.py generate-sct-dictionary --output-path assets/newdict_snomed.txtWe use sentence-transformers/all-MiniLM-L12-v2 as our embedding model. Concepts are grouped by concept_type_subset, and separate FAISS indices are built for each group.
Relevant code: rag/generate_snomedct_faiss.py
We use the Upstage Information Extraction API, which supports key-based entity extraction. Keys are aligned with the concept_type_subset definitions used for SNOMED CT.
The extracted results are mapped to the nearest concepts in the corresponding FAISS index.
Relevant code: rag/extraction.py
| Name | Role | GitHub |
|---|---|---|
| Kim Minseong | Leed Developer | @kimups |
| Lee Junyeong | Backend Developer | @parkstage |
| Kark Minuk | Frontend Developer | @parkstage |
- Last updated: 2025-04-05
This project is licensed under the MIT license.
See the LICENSE file for more details.