CareMind · Clinical Decision Support (MVP)

CareMind is a clinical decision support system (CDSS) prototype built with Streamlit.
It integrates retrieval-augmented generation (RAG), guideline databases, and drug knowledge bases to provide evidence-backed, explainable, and bilingual (中/EN) answers to clinical questions.

---

🚀 Features

• Clinical Q&A
  Ask questions like:

  “Can β-blockers be used in hypertensive patients with bronchial asthma?”
  The system retrieves guideline evidence and generates structured suggestions.

• Evidence Retrieval (RAG)
  Uses ChromaDB vector store for guideline chunks and SQLite for structured drug data.

• Structured Outputs

  • 📚 Evidence snippets
  • 💊 Drug interactions / adverse effects (from DB)
  • 🧭 Draft clinical recommendations with compliance disclaimer

• UI

  • Streamlit frontend
  • Download buttons for recommendation reports and evidence exports (Markdown format)
  • Compact bilingual interface (English / 中文)

---

📂 Project Structure

caremind-streamlit/
├── app.py                  # Streamlit frontend
├── rag/
│   ├── retriever.py        # Guideline & drug database retriever
│   ├── pipeline.py         # RAG pipeline & LLM integration
│   └── prompt.py           # Structured prompt templates
├── ingest/
│   └── create_db.py        # Ingest guideline files into ChromaDB
├── data/
│   ├── guidelines/         # Raw guideline PDFs
│   ├── guidelines.parsed.jsonl  # Parsed guideline text
│   └── drug_db.sqlite      # Drug structured database (optional)
├── chroma_store/           # Vector DB persistence (local / demo branch)
├── requirements.txt        # Python dependencies
└── .streamlit/config.toml  # Streamlit UI configuration

⚙️ Installation

1. Clone the repository git clone https://github.com//caremind-streamlit.git cd caremind-streamlit

2. Create environment conda create -n caremind python=3.10 -y conda activate caremind pip install -r requirements.txt

3. Set environment variables export CHROMA_PERSIST_DIR=./chroma_store export CHROMA_COLLECTION=guideline_chunks_1024_v2 export EMBEDDING_MODEL=BAAI/bge-large-zh-v1.5

4. Build the vector DB python ingest/create_db.py
   --in data/guidelines.parsed.jsonl
   --collection $CHROMA_COLLECTION

5. Run Streamlit streamlit run app.py The app will be available at http://localhost:8501.

🌐 Deployment

Local:

Works on Windows/WSL + GPU (RTX 4070 SUPER tested)

Streamlit Cloud:

main branch: minimal demo (no full dataset)

demo-data branch: full dataset (≈ 18MB SQLite + 17MB Chroma index)

🛡️ Compliance & Disclaimer

This tool is for research and demonstration purposes only. It does not replace professional medical judgment. All outputs include compliance disclaimers.

📌 Roadmap

Improve evidence ranking with rerankers

Enhance prompt templates for structured outputs

Support larger LLMs (H100/A100 inference-ready)

Multi-user demo via Docker or Cloud Run

Paper publication on clinical RAG systems

🤝 Contribution

Pull requests are welcome. For major changes, please open an issue first to discuss.

📜 License

MIT License © 2025 Samuel Huang

---

📚 Documentation

• Docs index: docs/00_README.md
• Streamlit Cloud deployment: docs/guides/deployment_cloud.md
• Testing & validation: docs/guides/testing_validation.md