🌍 Life Expectancy Prediction Using Machine Learning

This Project is part of a MATH 444 class at CSUN - Project 2

This project analyzes global life expectancy data using various statistical and machine learning techniques. Our goal is to identify key socio-economic and health factors influencing life expectancy and build accurate models to predict it.

---

👨‍👩‍👧‍👦 Team Members

• Brandon Ismalej
• Jittapatana (Patrick) Prayoonpruk
• Mishek Sambahangphe

---

📊 Dataset Overview

• Source: World Health Organization (WHO) - Global Health Observatory (GHO)
• Period: 2000–2015
• Countries: 193
• Size: 2938 rows, 22 columns
• Target Variable: Life Expectancy (continuous)

🔍 Features Include:

• Health Indicators: BMI, HIV/AIDS, Immunization rates
• Socioeconomic Indicators: GDP, Population, Schooling, Income composition of resources
• Demographic Data: Adult Mortality, Infant Deaths, Under-five Deaths

---

⚙️ Preprocessing

• Training Set: Data from 2014

• Testing Set: Data from 2015

• Missing Value Handling:

  • Imputed using missForest (non-parametric Random Forest-based method)
  • 10 iterations, 100 trees per forest

• Post-Imputation Size:

  • Training: 183×22
  • Testing: 183×22

---

📐 Feature Engineering

📌 Stepwise Feature Selection (using AIC/BIC)

Selected Predictors:

• Income Composition of Resources
• HIV/AIDS
• Adult Mortality
• Total Expenditure

🔄 Box-Cox Transformation

• Applied to the target variable to stabilize variance
• Optimal λ (lambda): 1.0303

---

🧪 Model Evaluation

Model	Adjusted R²	RMSE	Key Features
Linear Regression (Baseline)	0.796	—	Adult Mortality, HIV/AIDS, Income Composition
Stepwise Regression	0.8666	—	+ Total Expenditure
Box-Cox Transformed Regression	0.8664	2.99	Same as stepwise
Generalized Least Squares (GLS)	—	2.63	Improved residual handling
CatBoost with PCA	0.8849	2.75	Dimensionality reduction (90% variance kept)
CatBoost without PCA	—	1.92	Best performance overall

---

🤖 Machine Learning: CatBoost

• Algorithm: Gradient Boosting (handles categorical & continuous)
• Feature Selection: All-Subset Regression
• PCA: Used in one version to reduce dimensionality
• Performance: CatBoost without PCA achieved the lowest RMSE and highest R² across all models

---

🧠 Principal Component Analysis (PCA)

• Applied before CatBoost (in one model)
• Retained 90% of original variance
• However, the non-PCA CatBoost model outperformed it

---

🏁 Conclusion

• Generalized Least Squares was the best linear model
• CatBoost without PCA delivered the most accurate life expectancy predictions
• PCA, while helpful for dimensionality reduction, did not enhance CatBoost performance in this case

---

📁 Project Files

• life_expectancy_model.ipynb: Jupyter notebook with all code and analysis
• data/: Cleaned and imputed dataset
• models/: Trained model outputs and diagnostics
• plots/: Visualizations for EDA and model diagnostics

---

💡 Future Work

• Include more recent data post-2015
• Experiment with ensemble stacking and deep learning methods
• Explore regional and continent-based subgroup modeling
• Add interpretable ML tools (e.g., SHAP, LIME) for feature impact explanations

---