Premium Pulse - Medical Insurance Cost Prediction

A supervised machine learning project that predicts medical insurance charges using the Medical Insurance Cost Dataset from Kaggle.

The project benchmarks 10 regression models, ranging from simple linear regression to advanced gradient boosting ensembles, while maintaining consistent preprocessing pipelines and systematic hyperparameter tuning.

---

Project Overview

Insurance charges depend on multiple demographic and lifestyle factors such as age, BMI, smoking habits, and family size.

This project aims to:

• Predict individual medical insurance costs.
• Compare the performance of multiple regression algorithms.
• Analyze the impact of feature engineering on model performance.
• Demonstrate that domain knowledge can significantly improve simpler models.

---

Project Structure

Medical-Insurance-Prediction/
│
├── load_data.py                     # Dataset loading and train/test split
├── pipeline.py                      # Preprocessing pipeline for linear models
├── pipeline_tree_model.py           # Preprocessing pipeline for tree-based models
│
├── eda.ipynb
├── linear_regression.ipynb
├── ridgecv.ipynb
├── lasso_cv.ipynb
├── elastic_net.ipynb
├── elastic_net_grid_searchcv.ipynb
├── svm.ipynb   
├── decision_trees.ipynb
├── random_forest.ipynb
├── adaboost.ipynb
├── gradient_boosting.ipynb
├── xgboost.ipynb
│
├── models/
│   ├── linear_regression_pipeline.pkl   # Trained Linear Regression pipeline
│   └── xgboost_pipeline.pkl             # Trained XGBoost pipeline
│
├── images/                         # EDA plots, model evaluation graphs, and visualizations
│
└── README.md

---

Dataset

• Dataset: Medical Insurance Cost Prediction Dataset
• Source: Kaggle
• Dataset Link: https://www.kaggle.com/datasets/mosapabdelghany/medical-insurance-cost-dataset
• Target Variable: charges

Features

• Age
• Sex
• BMI
• Children
• Smoker
• Region

---

Models Implemented

• Linear Regression
• Ridge Regression (RidgeCV)
• Lasso Regression (LassoCV)
• Elastic Net (CV)
• Elastic Net (GridSearchCV)
• Support Vector Regression (SVR)
• Decision Tree Regressor
• Random Forest Regressor
• AdaBoost Regressor
• Gradient Boosting Regressor
• XGBoost Regressor

---

Scatterplot of charges vs bmi hue='smoker'

Feature Engineering

One of the major findings of this project is that domain-specific interaction features dramatically improved linear model performance.

Engineered Features

smoker_bmi

smoker_bmi = smoker × BMI

Captures the increased impact of BMI among smokers.

smoker_obese

smoker_obese = smoker AND BMI >= 30

Represents smokers who are also clinically obese.

These features model the real-world compounding effect of smoking and obesity on medical insurance costs.

---

Results

Model	Test MAE	Test R²
XGBoost	~1416	~0.8629
Gradient Boosting	~1611	~0.8595
SVM	~1724	~0.8596
Decision Tree	~1875	~0.8527
Random Forest	~1970	~0.8603
Linear Regression	~2391	~0.8690
Ridge CV	~2391	~0.8690
Elastic Net (GridSearchCV)	~2391	~0.8690
Lasso CV	~2400	~0.8689
Elastic Net CV	~2486	~0.8679
AdaBoost	~2551	~0.8576

Key Insights

• Carefully engineered interaction features allowed linear regression to outperform several ensemble methods in R².

• XGBoost achieved the lowest Mean Absolute Error (MAE), making it the best model for minimizing prediction error.

• The data naturally separates into three distinct groups:

  • Non-smokers
  • Smokers with normal BMI
  • Smokers with high BMI

• These groups explain the structured residual patterns observed across nearly every model.

Evaluation Metric: Models were trained and tuned using cross-validation with neg_mean_absolute_error because target column is having outliers.

---

Feature Importance & Residuals (XGBoost)

Setup

You need Kaggle API and kagglehub to run this.

Installation

pip install -r requirements.txt

---

Running the Project

Clone the repository:

git clone https://github.com/YugamdeepGoyal/PremiumPulse.git

Navigate to the project:

cd PremiumPulse

Open any notebook.

The dataset will be downloaded automatically using kagglehub.

---

Skills Demonstrated

• Exploratory Data Analysis (EDA)
• Feature Engineering
• Scikit-learn Pipelines
• Column Transformers
• Hyperparameter Tuning
• GridSearchCV
• Cross Validation
• Model Evaluation
• Regression Analysis
• Ensemble Learning
• XGBoost
• Machine Learning Workflow

---

License

This project is licensed under the MIT License.

This project is intended for educational and portfolio purposes.