Healthcare Cost Prediction Using Neural Networks

Quick Overview

Problem: Predict individual healthcare costs using regression analysis on demographic and health data, achieving Mean Absolute Error (MAE) under $3,500.

Solution: Built a deep neural network regression model with feature normalization and categorical encoding to predict medical expenses from patient characteristics.

Impact: Achieved $2,800 MAE (20% better than the requirement), demonstrating proficiency in regression modeling, data preprocessing, and TensorFlow/Keras implementation.

Completed Project

https://colab.research.google.com/drive/1TzGEH3w3DyVbxumFz2LufzQEgTDMPYgX?usp=drive_link

Step-by-Step Implementation

Step 1: Environment Setup & Data Loading

What: Imported TensorFlow 2.x, Keras, and a healthcare insurance dataset.

Dataset: insurance.csv containing patient information and medical costs.

Key Libraries:

• TensorFlow / Keras for neural networks
• Pandas for data manipulation
• Matplotlib for visualization
• tensorflow_docs for training callbacks

---

Step 2: Data Exploration

What: Examined dataset structure and features.

Features Identified:

• Numerical: age, bmi, children
• Categorical: sex, smoker, region
• Target Variable: expenses (healthcare costs)

Initial Analysis: Used dataset.tail() to inspect data structure and identify preprocessing needs.

---

Step 3: Categorical Data Encoding

Problem: Neural networks require numerical inputs, but the dataset contains text categories.

Solution: Manual mapping of categorical variables to numerical values.

• Sex: male → 0, female → 1
• Smoker: no → 0, yes → 1
• Region:
  • southwest → 0
  • southeast → 1
  • northwest → 2
  • northeast → 3

Key Action: Applied string stripping to handle whitespace before mapping.

Impact: Converted all features to numerical format for model compatibility.

---

Step 4: Train-Test Split

Split Ratio: 80% training, 20% testing.

Method: Random sampling using frac=0.2 for the test set.

Implementation:

• Test dataset: 20% random sample from original data
• Train dataset: Remaining 80%, created by dropping test indices

Label Separation:

• Popped the expenses column to create labels
• train_labels: target values for training
• test_labels: target values for evaluation

---

Step 5: Feature Normalization

Problem: Features exist on different scales (age: 18–64, BMI: 15–50, expenses: $1k–$60k), causing training instability.

Solution: Z-score normalization (standardization).

• Calculated mean and standard deviation from training data only
• Applied formula: (x − mean) / std

Critical Decision: Used training statistics for both training and test data to prevent data leakage.

Impact:

• Faster convergence during training
• Improved gradient descent stability
• Better model generalization

---

Step 6: Neural Network Architecture Design

Model Type: Sequential feedforward neural network for regression.

Architecture:

• Input layer: 6 features after encoding
• Hidden Layer 1: 64 neurons, ReLU activation
• Hidden Layer 2: 64 neurons, ReLU activation
• Output Layer: 1 neuron for continuous cost prediction

Rationale:

• Two hidden layers capture non-linear relationships
• ReLU activation avoids vanishing gradients
• 64 neurons balance learning capacity and overfitting risk

---

Step 7: Model Compilation

Optimizer: RMSprop (learning rate = 0.001)

• Adaptive learning rate suitable for regression
• Handles sparse gradients effectively

Loss Function: Mean Squared Error (MSE)

• Standard loss for regression
• Penalizes large errors more heavily

Metrics Tracked:

• MAE (Mean Absolute Error): primary evaluation metric
• MSE (Mean Squared Error): secondary monitoring metric

---

Step 8: Model Training

Training Configuration:

• Epochs: 500
• Validation split: 20% of training data
• Verbose: 0 with EpochDots callback

Data Used: Normalized training data

Monitoring: Validation loss tracked to detect overfitting.

Result: Model learned relationships between patient features and healthcare costs.

---

Step 9: Model Evaluation

Test Evaluation: model.evaluate() on normalized test data.

Key Metric: Mean Absolute Error (MAE)

Performance Achieved: $2,800 MAE

Target: ≤ $3,500

✅ Beat target by $700 (20% improvement).

Interpretation: Model predictions are, on average, $2,800 away from actual healthcare costs.

---

Step 10: Prediction Visualization

What: Generated a scatter plot of true versus predicted expenses.

Plot Elements:

• X-axis: Actual healthcare costs (test_labels)
• Y-axis: Model predictions
• Diagonal line: Perfect prediction reference
• Equal aspect ratio for accurate comparison

Interpretation:

• Points close to the diagonal indicate accurate predictions
• Scatter distribution highlights model strengths and weaknesses
• Visual confirmation of $2,800 MAE performance

---

Step 11: Final Validation

Automated Test: Built-in evaluation verifying MAE ≤ $3,500.

Result:

✅ "You passed the challenge. Great job!"

Verification: Model successfully generalizes to unseen test data.