Medical Cost Prediction

Problem Statement

Health insurance costs vary dramatically across individuals, and understanding cost drivers is essential for insurers, employers, and policyholders. The challenge: build an accurate regression model that predicts individual medical costs and, crucially, explains which factors drive those costs — enabling what-if analysis for policy design and personal health decisions.

Technical Approach

Exploratory Analysis & Feature Engineering

The dataset contained demographic and health information for 1,300+ individuals. Key features: age, sex, BMI, number of children, smoking status, and region. Exploratory analysis revealed a powerful interaction effect: smoking multiplied the effect of BMI on costs — smokers with high BMI had 4x the costs of non-smokers with the same BMI. This interaction term became the single most predictive feature in the final model.

Model Selection

• XGBoost: Achieved the best performance with R² of 0.86 after hyperparameter tuning. Automatically captured non-linear relationships and feature interactions without explicit engineering.
• Random Forest: R² of 0.84 — close second, with the advantage of built-in feature importance for interpretability.
• Ridge Regression: R² of 0.74 — linear baseline for comparison. Underperformed due to the smoking-BMI interaction it couldn't capture without manual feature engineering.
• Linear Regression: R² of 0.74 — identical to Ridge, confirming regularization didn't help (low-dimensional dataset with strong signals).

Streamlit Deployment

Built an interactive what-if dashboard using Streamlit. Users adjust sliders for age, BMI, smoking status, and children to see real-time cost predictions. The dashboard displays feature contribution breakdowns showing exactly how each factor impacts the predicted cost. Deployed as a self-contained Python app — no frontend code required.

Key Results

• Achieved R² of 0.86 with XGBoost — explaining 86% of cost variance
• Identified smoking-BMI interaction as strongest cost driver — 4x multiplier effect
• Deployed interactive Streamlit dashboard for real-time what-if cost simulation
• Age and number of children were secondary predictors — additive but weaker effects

Tech Stack