Health Risk Classification for Insurance Premium Optimization

Overview

Health insurance providers rely on accurate risk assessment to determine applicant eligibility and premium pricing. This project develops a machine learning classification system capable of predicting whether an insurance applicant is Healthy or Unhealthy using demographic, lifestyle, and medical attributes.

The project demonstrates a complete machine learning workflow including data auditing, exploratory data analysis, preprocessing, model development, model comparison, ROC-AUC evaluation, learning curve analysis, feature importance analysis, and business recommendations.

---

Problem Statement

Anova Insurance seeks to optimize health insurance premium pricing based on the health condition of applicants.

The objective is to develop a predictive model that classifies applicants into:

• Healthy (0)
• Unhealthy (1)

This classification can support:

• Insurance eligibility assessment
• Risk-based premium pricing
• Early identification of high-risk applicants
• Data-driven underwriting decisions

---

Machine Learning Workflow

Raw Dataset
     │
     ▼
Data Audit
     │
     ▼
Exploratory Data Analysis
     │
     ▼
Data Preprocessing
     │
     ▼
Logistic Regression
     │
     ▼
Decision Tree
     │
     ▼
Random Forest
     │
     ▼
ROC-AUC Analysis
     │
     ▼
Learning Curve Analysis
     │
     ▼
Feature Importance Analysis
     │
     ▼
Final Model Selection

---

Dataset Summary

Attribute	Value
Records	10,000
Features	23
Problem Type	Binary Classification
Target Classes	Healthy / Unhealthy

Numerical Features

• Age
• BMI
• Blood Pressure
• Cholesterol
• Glucose Level
• Heart Rate
• Sleep Hours
• Exercise Hours
• Water Intake
• Stress Level

Lifestyle & Medical Features

• Smoking
• Alcohol
• Diet
• Mental Health
• Physical Activity
• Medical History
• Allergies

---

Data Audit & Preprocessing

Key Findings

• No duplicate records identified.
• Five columns contained missing values.
• 96 records contained invalid Age values equal to zero.
• Dataset classes were nearly perfectly balanced.

Preprocessing Steps

• Removed invalid Age = 0 records.

• Median imputation:

  • Blood Pressure
  • Cholesterol
  • Glucose Level

• Mode imputation:

  • Medical History
  • Allergies

• Verified zero remaining missing values.

Final dataset size:

9,904 records

---

Exploratory Data Analysis

The following analyses were performed:

Univariate Analysis

• Numerical feature distributions

• Boxplot-based outlier analysis

• Categorical feature distributions

Bivariate Analysis

• Numerical features vs Target

• Categorical features vs Target

• Correlation analysis

Key Findings

• BMI showed the strongest positive relationship with health risk.
• Blood Pressure, Cholesterol, Age, and Glucose Level were influential predictors.
• Most categorical variables exhibited relatively weak standalone predictive power.

---

Models Evaluated

1. Logistic Regression

Baseline classification model.

Results

• Accuracy: 78.0%
• ROC-AUC: 0.862

---

2. Decision Tree

Tree-based classifier capable of learning non-linear relationships.

Results

• Accuracy: 80.6%
• ROC-AUC: 0.806

Training accuracy reached 100%, indicating overfitting.

---

3. Random Forest

Ensemble learning approach using multiple decision trees.

Results

• Accuracy: 85.6%
• ROC-AUC: 0.938

Random Forest achieved the best overall performance.

---

Model Performance Comparison

Model	Accuracy	ROC-AUC
Logistic Regression	0.780	0.862
Decision Tree	0.806	0.806
Random Forest	0.856	0.938

Performance Visualization

---

ROC-AUC Analysis

ROC curves were generated for all models to evaluate discrimination capability across classification thresholds.

Results

Model	ROC-AUC
Logistic Regression	0.862
Decision Tree	0.806
Random Forest	0.938

ROC Curve

---

Learning Curve Analysis

Learning curve analysis was performed for the Random Forest model.

Observations

• Training accuracy remained close to 100%.
• Validation accuracy improved steadily as additional data was introduced.
• The model demonstrated strong generalization capability.

Learning Curve

---

Feature Importance Analysis

Random Forest feature importance analysis identified the most influential predictors.

Top Features

1. BMI
2. Blood Pressure
3. Cholesterol
4. Age
5. Glucose Level

Feature Importance Plot

---

Business Recommendations

Based on the findings:

• Implement risk-based premium pricing strategies.
• Prioritize BMI, Blood Pressure, Cholesterol, and Glucose Level during underwriting.
• Introduce wellness programs for high-risk applicants.
• Use machine learning-assisted screening to improve underwriting consistency and efficiency.

---

Final Model Selection

Selected Model: Random Forest Classifier

Reasons:

• Highest Accuracy (85.6%)
• Highest ROC-AUC (0.938)
• Balanced Precision and Recall
• Strong generalization performance
• Interpretable feature importance

---

Technologies Used

• Python
• Pandas
• NumPy
• Matplotlib
• Seaborn
• Scikit-learn

---

Repository Structure

health-risk-classification-insurance/
│
├── Health_Risk_Classification.ipynb
├── Health_Risk_Classification_Report.pdf
├── dataset/
│   └── Healthcare_Data_Preprocessed.csv
│
├── images/
│   ├── numerical_distributions.png
│   ├── categorical_distributions.png
│   ├── model_comparison.png
│   ├── roc_curve.png
│   ├── learning_curve.png
│   └── feature_importance.png
│
└── README.md

---

Author

Ruthuraraj R

Assistant Professor | Mechanical Engineering Applied AI, Machine Learning, Generative AI, AI Agents, and Engineering Intelligence Systems