README.md

metadata

title: PIE Med
emoji: 📈
colorFrom: gray
colorTo: purple
sdk: streamlit
sdk_version: 1.42.1
app_file: app.py
pinned: false
license: cc-by-sa-4.0
short_description: '🩺 PIE-Med: Predicting, Interpreting and Explaining Medical'

🩺 PIE-Med: Predicting, Interpreting and Explaining Medical Recommendations

Welcome to the repository for PIE-Med, a cutting-edge system designed to enhance medical decision-making through the integration of Graph Neural Networks (GNNs), Explainable AI (XAI) techniques, and Large Language Models (LLMs).

🎥 Demo (or GIF)

Watch our demo to see PIE-Med in action and learn how it can transform healthcare recommendations!

📊 Data Source

We use the MIMIC-III dataset, a freely accessible critical care database containing de-identified health information, including vital signs, laboratory test results, medications, and more. You can find more details about the dataset here:

🛠 Technologies Used

• Python: Core programming language
• Pandas: Data manipulation and analysis (Pandas Documentation)
• PyHealth: Medical data preprocessing (PyHealth Documentation)
• PyTorch Geometric: Building and training GNNs (PyTorch Geometric Documentation)
• Integrated Gradients & GNNExplainer: Interpretability techniques (PyTorch Geometric Documentation)
• Streamlit: User interface development (Streamlit Documentation)
• Py AutoGen Multi-Agent Conversation Framework: Multi-agent collaboration and explanation (Py AutoGen Documentation)

The PIE-Med system's computational requirements depend on the configuration used. For resource-limited environments, the light configuration with an Intel i7 CPU and 16GB RAM offers a basic but functional setup, suitable for testing on small datasets. However, more demanding tasks, such as working with larger datasets or leveraging advanced machine learning techniques (e.g., Graph Neural Networks), benefit from cloud setups like the complete configuration, which includes a GPU (NVIDIA Tesla T4). In resource-constrained contexts, optimizing models and reducing dataset size would be crucial to ensure feasible performance.

🔬 Methodological Workflow

PIE-Med follows a comprehensive Predict→Interpret→Explain (PIE) paradigm:

1. Prediction Phase: We construct a heterogeneous patient graph from MIMIC-III data and apply GNNs to generate personalized medical recommendations.
2. Interpretation Phase: Integrated Gradients and GNNExplainer techniques are used to provide insights into the GNN's decision-making process.
3. Explanation Phase: A collaborative ensemble of LLM agents analyzes the model's outputs and generates comprehensive, understandable explanations.

🌟 Key Features

• Integration of GNNs and LLMs: Combining structured machine learning with natural language processing for robust recommendations.
• Enhanced Interpretability: Using XAI techniques to make the decision-making process transparent.
• Collaborative Explanation: Multi-agent LLMs provide detailed and understandable recommendations.

🚀 Getting Started

Follow these steps to set up and run PIE-Med on your local machine:

Prerequisites

Ensure you have the following installed:

• Python 3.7+

Installation

1. Clone the repository:

   git clone https://github.com/picuslab/PIE-Med.git
   cd PIE-Med
   

2. Create a virtual environment:

   python -m venv venv
   source venv/bin/activate  # On Windows use `venv\Scripts\activate`
   

3. Install the required packages:

   pip install -r requirements.txt
   

Running the Application

1. Run the Streamlit application:

   streamlit run dashboard.py
   

   Open your web browser and go to http://localhost:8501 to interact with the application.

📈 Conclusions

PIE-Med showcases the potential of combining GNNs, XAI, and LLMs to improve medical recommendations, enhancing both accuracy and interpretability. Our system effectively separates prediction from explanation, reducing biases and enhancing decision quality.

⚖ Ethical considerations

PIE-Med aims to support medical decision-making, but is not a substitute for professional medical advice. Users should confirm recommendations with authorised healthcare providers, as limitations of AI may affect accuracy. The system ensures transparency through interpretability techniques, but all results should be considered complementary to expert advice. ⚠️ Please note that the following repository is only a DEMO, with anonymised data used for illustrative purposes only.

🙏 Acknowledgments

We extend our gratitude to the creators of the MIMIC-III database, the developers of the Python libraries used, and our research team for their contributions to this project.

👨‍💻 This project was developed by Antonio Romano, Giuseppe Riccio, Marco Postiglione and Vincenzo Moscato