🩺 MASH-Assist AI: Clinical Support Tool

MASH-Assist AI is a functional prototype developed as a portfolio project for the MIT Hacking Medicine in São Paulo hackathon. This tool is designed to "give a voice" to Metabolic Dysfunction-Associated Steatohepatitis (MASH), a silent but serious chronic disease, by tackling two of its biggest challenges: underdiagnosis and the lack of readily accessible clinical knowledge.

The project directly addresses Track 1: MASH and the InterSystems GenAI Challenge.

---

✨ Key Features

This project combines a classic machine learning model with a modern Retrieval-Augmented Generation (RAG) system to demonstrate a dual-function clinical support tool.

1. Risk Prediction Model

• Purpose: To stratify a patient's risk of having MASH based on a range of common clinical and demographic data.
• Method: An XGBoost Classifier model trained on the USA's National Health and Nutrition Examination Survey (NHANES) 2011-2018 dataset. The model's target variable is a proxy for MASH risk, where a Fatty Liver Index (FLI) score of >= 60 is classified as 'High Risk'.
• Input: The model uses a core set of demographic, laboratory, examination, and questionnaire variables (e.g., age, gender, ethnicity, glucose, HbA1c, lipids, liver enzymes, blood pressure).
• Output: A risk classification of Low Risk or High Risk.
• Implementation: See notebook_risk_prediction.ipynb for the complete data processing, training, evaluation, and model interpretation using SHAP.

2. AI Knowledge Assistant

• Purpose: To provide healthcare professionals with quick, accurate answers to questions about MASH diagnosis, management, and guidelines.
• Method: A Retrieval-Augmented Generation (RAG) pipeline using Google's Gemini LLM.
• Knowledge Base: The AI's knowledge is strictly limited to a curated set of PDF documents, ensuring answers are contextually relevant and accurate.
• Functionality: Users can ask questions in natural language (e.g., "What are the key recommendations for the pharmacological treatment of MASH?") and receive a detailed answer synthesized from the source documents.
• Implementation: See notebook_ai_assistant_FAISS.ipynb for the setup of the vector store and the question-answering chain.

---

🛠️ Technology Stack

• Backend & Modeling: Python
• Machine Learning: Scikit-learn, Pandas, NumPy, XGBoost, SHAP
• Generative AI: LangChain, Google Gemini API (gemini-1.5-flash)
• Vector Store (Local): FAISS
• Embeddings: Hugging Face Sentence Transformers (all-MiniLM-L6-v2)
• Development Environment: Jupyter Notebook

---

🚀 How to Run the Project

Follow these steps to set up the project and run the notebooks locally.

Prerequisites

• Python 3.9+
• A Google API Key for the Gemini model. You can get one from Google AI Studio.

1. Clone the Repository

git clone [https://github.com/YOUR_USERNAME/MASH-Assist-AI.git](https://github.com/YOUR_USERNAME/MASH-Assist-AI.git)
cd MASH-Assist-AI

2. Set Up the Environment

Create and activate a virtual environment:

# Create the environment
python -m venv venv

# Activate on macOS/Linux
source venv/bin/activate

# Activate on Windows
# venv\Scripts\activate

Install the required dependencies:

pip install -r requirements.txt

3. Configure API Key

Create a file named .env in the root of the project directory and add your Google API key:

GOOGLE_API_KEY=YOUR_API_KEY_HERE

4. Run the Notebooks

Launch Jupyter Notebook or JupyterLab to explore the project:

# To start Jupyter Notebook
jupyter notebook

• To train the risk model: Open and run the cells in notebook_risk_prediction.ipynb. This will process the raw data and save the trained model as mash_risk_model.pkl.
• To test the AI assistant: Open and run the cells in notebook_ai_assistant_FAISS.ipynb. This will build the vector store (if it doesn't exist) and allow you to ask questions against the knowledge base.

---

📂 Project Structure

MASH-Assist-AI/
│
├── nhanes_data/                     # Folder for raw NHANES data (.XPT files)
├── knowledge_base/           # Folder for PDF documents used by the RAG system
├── faiss_index/              # Saved FAISS vector store index
│
├── notebook_risk_prediction.ipynb  # Notebook for data processing and model training
├── notebook_ai_assistant_FAISS.ipynb # Notebook for the RAG AI Assistant
├── requirements.txt          # List of Python dependencies
├── .env                      # File for API keys (not committed to Git)
└── README.md                 # This file

---

🔮 Next Steps

• Develop a User Interface: Build an interactive web application using Streamlit or Flask to host the risk calculator and AI assistant, making it accessible to end-users.
• Implement a Scalable Vector Database: Replace the local FAISS index with a more robust and scalable vector database solution like InterSystems IRIS for production environments.
• Deploy the Application: Package the models and application for deployment on a cloud service (e.g., AWS, Google Cloud, Heroku).
• Expand the Knowledge Base: Incorporate a wider range of clinical guidelines, research papers, and medical literature to enhance the AI assistant's expertise.
• Refine the Prediction Model: Experiment with different machine learning models or test more patient features to improve the accuracy and scope of the risk prediction.