Hospital Readmission Prediction Benchmark

This project implements baseline models for predicting 30-day hospital readmissions using MIMIC-III data, with a focus on the LACE index and machine learning approaches.

Project Structure

readmission_prediction/
├── data/
│   └── data_loader.py          # Data loading and preprocessing
├── models/
│   ├── lace_model.py          # LACE index implementation
│   ├── random_forest_model.py  # Random Forest baseline
│   └── base_model.py          # Base model interface
├── utils/
│   ├── feature_engineering.py # Feature extraction and engineering
│   ├── metrics.py             # Custom evaluation metrics
│   └── config.py              # Configuration settings
├── evaluation/
│   ├── evaluator.py           # Model evaluation pipeline
│   └── visualizations.py      # Plotting and visualization tools
└── main.py                    # Main execution script
results/
├── models/                    # Trained model artifacts
├── predictions/               # Model predictions
└── reports/                   # Analysis reports and plots

Models Implemented

1. LACE Index Baseline (Logistic Regression)

• Length of stay (0-7 points)
• Acuteness of admission (0-3 points)
• Comorbidities via Charlson Index (0-5 points)
• Emergency visits in prior 6 months (0-4 points)

2. Random Forest Baseline

• Ensemble method using LACE components and additional clinical features
• Handles non-linear relationships and feature interactions
• Built-in feature importance analysis

Usage

# Install dependencies
pip install -r requirements.txt

# Run all baseline models
python main.py

# Run specific model
python main.py --model lace
python main.py --model random_forest

Evaluation Metrics

• AUC-ROC: Area under the ROC curve
• AUC-PR: Area under the Precision-Recall curve (better for imbalanced data)
• Accuracy: Overall classification accuracy
• Precision/Recall/F1: Class-specific performance metrics
• Sensitivity/Specificity: Clinical interpretation metrics

Dataset

This project uses MIMIC-III demo data containing:

• Patient demographics and admission details
• ICD-9 diagnosis codes for comorbidity calculation
• Length of stay and admission type information
• Historical emergency department visits

Results

Results are automatically saved to the results/ directory with timestamps and model configurations for reproducibility.

Requirements

• Python 3.7+
• pandas >= 1.3.0
• numpy >= 1.21.0
• scikit-learn >= 1.0.0
• matplotlib >= 3.3.0
• seaborn >= 0.11.0