Predicting Hospital Readmission using Logistic Regression

This project applies statistical modeling to identify key factors contributing to hospital readmissions, using a dataset of 25,000 patient records. The objective is to help healthcare providers proactively detect high-risk patients and reduce preventable readmissions through data-driven insights.

Status: Completed Last Updated: March 30, 2025

---

Project Overview

• Objective: Predict hospital readmission using patient demographics, treatment history, and medical visit data.
• Dataset: Kaggle Hospital Readmissions Dataset
• Modeling Approach: Logistic Regression in R
• Toolset: R, tidyverse, caret, pROC, ggplot2

---

Methodology

Data Preprocessing

• Converted categorical variables into factors (e.g., age brackets, diagnoses, test results)
• Transformed target variable readmitted into binary format (0 = No, 1 = Yes)
• Standardized numeric variables using Z-score normalization
• Handled right-skewed distributions using Winsorization to limit outliers

Feature Engineering

• Selected variables based on medical relevance:
  • Time in hospital
  • Number of medications, lab procedures, and primary/secondary diagnoses
  • Prior hospital visits (inpatient, outpatient, emergency)
  • Diabetes medication status and diagnostic test results

Model Training & Evaluation

• Applied logistic regression with stepwise AIC selection
• Evaluated on:
  • Accuracy, AUC (ROC Curve)
  • Precision, Recall, F1-score
  • Confusion Matrix
• Train-test split: 80% / 20%

---

Results

• Training Accuracy: 61.16%
• Test Accuracy: 61.65%
• AUC (Train): 0.647
• AUC (Test): 0.662

Key Findings

• Positive predictors of readmission:

  • Longer hospital stays
  • Higher number of medications and lab procedures
  • Frequent prior hospital visits
  • Prescription of diabetes medication
  • Advanced age (70s and 80s)

• Negative predictor:

  • Number of procedures performed during stay (associated with reduced risk)

• Non-significant variables:

  • Glucose and A1C test results
  • Change in diabetes medication

---

R Script Highlights (Project_1.R)

• Cleanly loads and preprocesses the dataset
• Performs model training and test evaluation
• Uses caret and pROC for model validation and ROC plotting
• Includes confusion matrices and AUC scores for performance assessment

---

How to Run

1. Clone this repository and install R (version 4.1+ recommended)
2. Open Project_1.R in RStudio
3. Set the correct path to the dataset CSV
4. Run all chunks to preprocess data, train the model, and view evaluation metrics

---

Limitations & Future Work

• Predictive performance is modest, likely due to missing behavioral, social, or follow-up data
• Future iterations may explore:
  • Regularized models (LASSO, Ridge)
  • Tree-based classifiers (Random Forest, XGBoost)
  • SHAP/LIME interpretability methods
  • Inclusion of external datasets to improve feature depth

---

Author

Angel Barrera
Master’s Student, Applied Data Science
East Tennessee State University
Email: [email protected]

---

References

1. Hospital Readmissions Dataset (Kaggle)
2. Regression & Advanced Data Analytics Course, ETSU, 2025