An AI-powered assistant for diagnostic reasoning using Retrieval-Augmented Generation (RAG) on annotated patient notes from the MIMIC-IV-Ext DiReCT dataset.
Built with ❤️ using Streamlit, SentenceTransformers, FAISS, and Transformers.
RAGnosis combines dense retrieval and large language models to answer clinical questions using real-world diagnostic notes. It retrieves relevant clinical cases and generates AI-assisted diagnostic suggestions using context.
- 🔍 Dense Retrieval of similar patient cases (FAISS + MiniLM)
- 💬 Context-aware Diagnosis Generation (LLMs)
- 📊 ROUGE, BLEU, and BLEU-1 evaluation support
- 🧠 Real-time clinical query interface via Streamlit
- 🧪 Optimized for GPU acceleration using CUDA
- 💅 Beautiful UI with custom styling
data/
├── diagnosis_flowchart/ # Diagnostic KG for each disease
├── finished_cases/ # Annotated notes (JSON)
├── clinical_embeddings.pkl # Precomputed note embeddingsgit clone https://github.com/AbsarRaashid3/RAGnosis.git
cd RAGnosis
# Optional: create a virtual environment
python -m venv .venv
.venv/Scripts/activate
# Install dependencies
pip install -r requirements.txt
⚙️ Usage 🔹 1. Preprocess Clinical Notes
python src/preprocessing.py
🔹 2. Run the Streamlit Frontend
streamlit run app.py
Then go to http://localhost:8501
You can test generation quality using:
python src/evaluation_utils.py
- ROUGE
- BLEU (standard)
- BLEU-1 (unigram precision)
RAGnosis/
├── app.py # Streamlit frontend
├── src/
│ ├── preprocessing.py # Embedding + data prep
│ ├── retriever.py # FAISS-based retriever
│ ├── generator.py # LLM-based generator
│ └── evaluation_utils.py # Evaluation metrics
├── assets/ # Banner / UI assets
└── .streamlit/config.toml # Custom Streamlit theme
- SentenceTransformers
- Transformers (HuggingFace)
- FAISS
- Streamlit
- Evaluate (HuggingFace)
"Does the patient have COPD?"
"Can this be diagnosed as a duodenal ulcer?"
"Is shortness of breath explained by heart failure?"
"Should this patient undergo endoscopy?"
Created by Muhammad Absar Raashid