Health Plan Cost Risk Analysis

Predicting High-Cost Patients Using Demographic & Behavioral Data.

This project analyzes U.S. medical insurance records to understand the drivers of medical spending and predict high-cost patients, a critical use case for payer organizations and health plans.

The analysis is split into two notebooks:

1. 01_Health_Plan_EDA.ipynb - Exploratory Data Analysis

2. 02_High_Cost_Risk_Model.ipynb - Predictive Modeling

---

Business Problem

Health plans must manage risk while maintaining care quality. Identifying members likely to incur high medical costs helps with:

• Early outreach and care management
• Budget forecasting
• Fair pricing and risk adjustment
• Targeted member education

This project demonstrates how statistical analysis and machine learning can support those goals.

---

Dataset

Source: Kaggle – Medical Cost Personal Dataset https://www.kaggle.com/datasets/mirichoi0218/insurance

Observations: 1,337

Target: Annual medical charges

Engineering Label:

• high_cost = 1 if annual charges > 75th percentile
• high_cost = 0 otherwise

---

Notebook 1 — Key EDA Insights

Factors associated with higher spending:

• Smokers spend roughly 3.7x more than non-smokers.
• Charges increase sharply with age (especially after 45).
• Higher BMI correlates with higher spending (obese group is the most expensive).
• Regional impact exists but is weaker (Southeast is highest).
• Cost distribution is heavily right-skewed (few very high spenders).

Business Takeaways:

• Smoking cessation could reduce claims cost.

• Preventive care for obesity and chronic illness has strong ROI.

• Age bands help segment risk for pricing and outreach.

• Region differences may matter with additional contextual data.

  ---

Notebook 2 — Predictive Modeling

Goal: Predict which members will be high-cost (top 25% of total charges).

Models Trained:

• Logistic Regression
• Random Forest Classifier

Random Forest

Model Performance (Test Set):

• ROC-AUC: approximately 0.92
• Recall (high-cost): approximately 82%
• Accuracy: approximately 96%

Feature Importance (ranking):

1. Smoking status
2. BMI
3. Age
4. Number of children
5. Sex and region (smaller influence)

Interpretation: High-cost risk is primarily driven by lifestyle and health behavior factors.

---

Tech Stack

• Python
• pandas, numpy, matplotlib, seaborn
• scikit-learn
• Jupyter Notebook
• Git and GitHub

---

Repository Structure

Health-Plan-Cost-Risk-Analysis/
│
├── data/
│   └── insurance.csv
├── notebooks/
│   ├── 01_Health_Plan_EDA.ipynb
│   └── 02_High_Cost_Risk_Model.ipynb
├── README.md
└── .gitignore

---

Next Steps (Future Enhancements)

• Add comorbidity features (hypertension, diabetes, etc.)
• Test Gradient Boosting / XGBoost models
• Build a Power BI dashboard for business users
• Deploy a lightweight scoring API