Predicting Hospital Readmission using Clinical Notes

A machine learning project to predict 30-day hospital readmission risk, helping hospitals identify high-risk patients for targeted interventions. This project leverages NLP techniques on unstructured EHR data.

1. Problem Statement

Unplanned hospital readmissions are costly for healthcare systems and often indicate poor patient outcomes. By proactively identifying patients at high risk of readmission, hospitals can implement post-discharge plans to improve care and reduce costs. This project builds a binary classification model to predict which patients will be readmitted within 30 days of discharge.

2. Dataset

This project uses the publicly available MIMIC-III dataset. Specifically, it utilizes the NOTEEVENTS.csv file containing de-identified clinical notes.

• Preprocessing: The text data was cleaned by removing stop words, punctuation, and converting to lowercase.
• Feature Engineering: TF-IDF was used to vectorize the clinical notes into a numerical format suitable for machine learning.

3. Methodology

A Logistic Regression model was chosen as a strong, interpretable baseline. The pipeline consists of:

1. Loading clinical notes.
2. Text cleaning and preprocessing.
3. Vectorization using TfidfVectorizer.
4. Training the Logistic Regression classifier.

4. Results

The model's performance was evaluated on a held-out test set. Given the class imbalance inherent in readmission prediction, metrics beyond accuracy are crucial.

Metric	Score
AUC-ROC	0.78
F1-Score	0.65
Precision	0.72
Recall	0.59

5. How to Run This Project

1. Clone the repository:

   git clone [https://github.com/shivaheidari/Readmission.git](https://github.com/shivaheidari/Readmission.git)
   cd Readmission

2. Install dependencies:

   pip install -r requirements.txt

3. Run the analysis: Open and run the Readmission_Analysis.ipynb notebook.
4. Future Work

6. Future Work

• Advanced Models: Advancing model with multi-modal predictions models.
• Deployment: Containerize the prediction pipeline with Docker and deploy it as a REST API using Flask/FastAPI.
• Feature Engineering: Incorporate structured data (lab values, demographics) alongside the text data.