🏥 Health Insurance Premium Prediction System

🌐 Live Application

🚀 Try the app here:
👉 https://ml-project-health-insurance-premium-predictor.streamlit.app/

No installation required — works directly in your browser.

---

🚀 Highlights

• 🔢 Real-time insurance premium prediction
• 🧠 Segmented ML architecture (age-based models)
• 🧬 Custom Genetic Risk feature engineering
• ⚙️ End-to-end preprocessing pipeline (encoding + scaling)
• 🎯 Clean and interactive UI built with Streamlit
• 🔄 Fully functional reset (true form reset using key-versioning)
• ⚡ Smooth user experience with responsive design

---

🧠 Machine Learning Approach

1. Problem Type

• Supervised Learning
• Regression (predicting continuous insurance premium)

---

2. Feature Engineering

A key highlight of this project is the custom normalized medical risk score:

• Converts medical history into numerical values
• Handles multiple conditions (e.g., Diabetes & High BP)
• Weighted scoring system:
  • Heart disease → high risk
  • Diabetes / BP → medium risk
  • Thyroid → lower risk
• Normalized to range [0, 1]

---

3. Model Segmentation

Instead of using a single model:

• Model 1 → Young Users (≤ 25)
• Model 2 → Adults (> 25)

👉 This improves prediction accuracy by capturing different behavioral and health risk patterns.

---

4. Preprocessing Pipeline

• Manual one-hot encoding for categorical variables
• Insurance plan encoding:
  • Bronze → 1
  • Silver → 2
  • Gold → 3
• Age-based scaling using separate scalers
• Ensures consistent feature structure for model input

---

📊 Model Performance & Impact of Feature Engineering

🔹 Young Model (≤ 25) — Before Genetic Risk

Model	Train Score	Test Score
Linear Regression	0.6020	0.6047
XGBoost	~0.603	~0.60

• ⚠️ Extreme Error (>10%): 73%

---

🔹 Young Model (≤ 25) — After Genetic Risk

Model	Train Score	Test Score
Linear Regression	0.9882	0.9887
XGBoost	~0.987	~0.98

• ✅ Extreme Error (>10%) reduced to: ~2%

---

🔹 Rest Model (> 25)

Model	Train Score	Test Score
Linear / Ridge Regression	~0.953	~0.953
XGBoost	~0.994

• ✅ Extreme Error (>10%): ~0.3%

---

🎯 Key Insight

The introduction of the Genetic Risk feature led to a massive performance improvement for the young age group:

• 📉 Extreme error reduced from 73% → 2%
• 📈 Model accuracy improved from ~0.60 → ~0.98

👉 This demonstrates the critical role of domain-specific feature engineering in machine learning performance

---

📊 Application Workflow

1. User enters personal, health, and policy details
2. Medical history → converted into normalized risk score
3. Data → encoded and scaled
4. Model selection based on age
5. Prediction generated
6. Result displayed via interactive UI

---

🛠️ Tech Stack

• Frontend: Streamlit
• Backend: Python
• Machine Learning: Scikit-learn, XGBoost
• Model Storage: Joblib
• Data Processing: Pandas, NumPy

---

🔁 Reset Functionality (Key Highlight)

A true reset functionality was implemented using dynamic key versioning:

• Streamlit widgets retain state by default
• Instead of clearing state, widget keys are dynamically changed
• This forces reinitialization of all inputs

👉 Result: behaves like a full page refresh

---

📦 Installation & Setup

git clone https://github.com/Saurabh136/ML_Project_Health_Insurance_Premium_Predictor.git
cd <project-folder>
pip install -r requirements.txt
streamlit run main.py