365-Day Hospital Readmission Prediction for Ischemic Stroke

Repository Overview

This repository presents an end-to-end machine learning pipeline for predicting 365-day hospital readmission among patients diagnosed with ischemic stroke. The study leverages the MIMIC-IV clinical database and evaluates multiple machine learning models to assess predictive performance and identify clinically relevant features.

---

Repository Contents

1. Team Contract – ML4Health_TeamContract_17.pdf
2. Project Proposal – ML4Health_proposal_Group17.pdf
3. Cohort Extraction and Exploratory Data Analysis – Data_Handling.ipynb
4. Generic Data Processing Pipeline – Process_Data_Generic.ipynb
5. Machine Learning Modelling – ML_Models.ipynb
6. Feature Importance Analysis – Feature_Importance_ISP_Dataset.ipynb
7. Pre-processed Full Cohort Dataset – full_cohort.csv
8. Environment and Dependency Requirements – requirements.txt

---

Objective

The objective of this study is to predict 365-day hospital readmission for patients admitted with ischemic stroke and to identify key predictive clinical features that may inform clinical decision-making and post-discharge risk stratification.

---

Materials and Methods

Dataset

• MIMIC-IV (version 2.2)
• Adult patients diagnosed with ischemic stroke

Machine Learning Models

• Random Forest
• XGBoost
• PCA-based Clustering

Methodology

• Class imbalance in readmission outcomes was addressed using SMOTE oversampling for the Random Forest model.
• Hyperparameters were optimized using grid search with cross-validation.
• Model performance was evaluated using AUC-ROC and additional classification metrics.
• Feature importance was examined using permutation importance and PCA-based clustering techniques.

---

Results

• Random Forest achieved an AUC of 0.64.
• XGBoost achieved an AUC of 0.78.

Overall predictive performance was moderate, indicating limited discriminative ability despite the use of imbalance correction techniques.

Key Predictive Features

• Length of hospital stay
• Troponin T levels
• Maximum systolic blood pressure

Feature importance analyses suggest complex, non-linear interactions among clinical variables contributing to readmission risk.

---

How to Run the Code

Step 1: Obtain Access to MIMIC-IV

Access to the MIMIC-IV database and BigQuery setup is required. Instructions are available at: https://physionet.org/content/mimiciv/2.2/

Step 2: Environment Setup

Install the required dependencies using:

pip install -r requirements.txt

Step 3: Cohort Extraction and Exploratory Data Analysis

Run Data_Handling.ipynb to extract the ischemic stroke cohort and perform exploratory data analysis.
Alternatively, the pre-processed dataset full_cohort.csv may be used directly.

Step 4: Model Training and Evaluation

Execute ML_Models.ipynb to perform data preprocessing, model training, model evaluation, and performance analysis.

Step 5: Feature Importance Analysis

Run Feature_Importance_ISP_Dataset.ipynb to analyze feature contributions to model predictions.

---

Conclusion

This study highlights the challenges associated with predicting hospital readmission using real-world clinical data. Constructing a clinically relevant, sufficiently large, and complete ischemic stroke cohort proved challenging. The moderate performance observed across models is consistent with existing literature, suggesting that 365-day stroke readmission prediction remains a complex and difficult task even with advanced machine learning techniques.

---

Group 17