🏥 AI-Powered Health Insurance Claim Approval System

📌 Project Overview

This project leverages GPT-4, FAISS, and FastAPI to automate health insurance claim approvals. The system retrieves past similar claims, analyzes them, and generates automated approval or denial decisions with reasoning.

🚀 Features

✅ AI-Powered Claim Processing – Uses GPT-4 for approval decisions.
✅ Retrieval-Augmented Generation (RAG) – Ensures decisions are backed by past claim data.
✅ FAISS Vector Search – Retrieves top similar past claims for analysis.
✅ Interactive Frontend (Streamlit) – Allows users to submit and review claims.
✅ Bias Detection & Override – Ensures fair claim evaluations.

---

🛠️ Tech Stack

Component	Technology Used
Backend API	FastAPI
Claim Retrieval	FAISS (Facebook AI Similarity Search)
LLM Processing	OpenAI GPT-4 (or Mistral-7B)
Vector Embeddings	Sentence Transformers (all-MiniLM-L6-v2)
Frontend UI	Streamlit
Database	CSV File (Kaggle dataset health_claims.csv)
Deployment	Local (Can be deployed on Hugging Face / Cloud)

---

📂 Folder Structure

ai-claim-approval/ 
│── app/ # Backend API logic 
│ ├── init.py # Makes app a package 
│ ├── main.py # FastAPI app logic 
│ ├── claim_model.py # FAISS retrieval & GPT processing 
│ ├── config.py # API keys & configuration 
│ ├── utils.py # Helper functions 
│ │── models/ # Stores FAISS index & embeddings 
│ │ ├── claims_faiss.index # FAISS vector index 
│ │ ├── embeddings.npy # Saved claim embeddings 
│ │── data/ # Dataset 
│ │ ├── health_claims.csv # Health insurance dataset 
│── build_faiss_index.py # Store FAISS indices
│── requirements.txt # Dependencies 
│── Dockerfile # Deployment instructions 
│── frontend.py # Streamlit frontend 
│── README.md # Project documentation 
│── .gitignore # Ignore unnecessary files

---

🏗️ How It Works

🚀 Step 1: Data Preprocessing

1. Load health insurance claims dataset
2. Convert claims into embeddings using sentence-transformers
3. Store embeddings in FAISS for fast retrieval

🚀 Step 2: Backend API (FastAPI)

1. User submits a new insurance claim through API
2. FAISS retrieves the top 3 most similar past claims
3. GPT-4 (or another LLM) analyzes past cases and makes a decision
4. API returns an "Approved" or "Denied" decision with reasoning

🚀 Step 3: Frontend UI (Streamlit)

1. User enters claim details (Age, Diagnosis, Procedure, etc.)
2. Streamlit sends claim to FastAPI backend
3. Backend retrieves past claims and gets GPT-4’s decision
4. Decision is displayed on the UI with past case references

---

🔧 Installation & Setup (Run Locally)

1️⃣ Install Dependencies

pip install fastapi uvicorn openai sentence-transformers faiss-cpu pandas numpy streamlit

2️⃣ Run Backend (FastAPI)

uvicorn app.main:app --host 127.0.0.1 --port 8000 --reload

✅ API will be live at http://127.0.0.1:8000/docs

3️⃣ Run Frontend (Streamlit)

streamlit run frontend.py

✅ UI will be available at http://localhost:8501

---

🚀 How the UI Output would look like: