Diabetic Patients' Re-admission Prediction

Course Information

This project is part of the DA380A HT24 Machine Learning course at Kristianstad University.

Project Overview

This project aims to predict hospital readmissions for diabetic patients within 30 days post-discharge. Using a dataset of around 100,000 hospital records, we developed a machine learning pipeline to identify high-risk patients and optimize healthcare resource allocation.

Key Objectives

1. Predictive Modeling: Identify diabetic patients at high risk of readmission.
2. Feature Analysis: Determine factors contributing to readmissions to help inform healthcare strategies.

Dataset

The dataset is derived from the Diabetes 130-US Hospitals dataset, which includes patient records from 130 hospitals in the United States between 1999 and 2008. It contains:

• 50 features: Patient demographics, hospital admission details, and medical information, including medications and lab results.
• Target Variable: readmitted - indicates whether a patient was readmitted within 30 days, after 30 days, or not readmitted.

Data Source: UCI Machine Learning Repository - Diabetes 130-US hospitals dataset

Methodology

1. Data Exploration and Preprocessing

• Clustering: K-means and DBSCAN were used to explore potential patterns but showed minimal clustering due to the complex nature of the data.
• Handling Missing Values: Imputed missing values for payer_code and race, and removed features with excessive missing data (e.g., weight).
• Outlier Management: Retained outliers in features like Time in Hospital due to their high predictive importance.
• Feature Engineering: Created age_group and used feature selection techniques to retain top predictors (e.g., number_inpatient, number_diagnoses).
• Encoding: Categorical features were encoded using one-hot encoding, and binary encoding for features with two values.
• Balancing: Applied SMOTE to address class imbalance.
• Scaling: Standardized numerical features for consistency.

2. Model Selection

• Models: We implemented Logistic Regression, Random Forest, and Gradient Boosting (XGBoost) along with AdaBoost and MLP.
• Model Improvement: Hyperparameter tuning (GridSearchCV) was performed for optimal performance. SMOTE was used to improve recall for minority classes.

3. Evaluation

• Metrics: Models were evaluated using accuracy, precision, recall, F1-score, and AUC-ROC.
• Best Performing Model: The Gradient Boosting model achieved the highest accuracy, followed closely by the Random Forest model.

Results

• Gradient Boosting (XGBoost): Showed the highest predictive accuracy and balanced precision-recall performance.
• Balanced Random Forest: Achieved good recall, suitable for applications prioritizing detection of high-risk patients.
• Key Features: Important predictors included number_inpatient, number_diagnoses, and number_emergency.

Discussion

Challenges

• Class Imbalance: The dataset had a small proportion of readmissions, which was addressed through SMOTE.
• Data Quality: Extensive preprocessing was needed due to missing values and high-cardinality features.

Lessons Learned

• Robust preprocessing and feature engineering significantly impact model performance.
• Handling imbalanced datasets is crucial in healthcare prediction tasks to ensure accuracy across all classes.

Future Work

• Deep Learning: Explore models such as LSTM for sequential data analysis.
• Data Enrichment: Integrate additional data sources to improve prediction accuracy.

How to Run the Project

1. Clone the repository.
2. Install dependencies: pip install -r requirements.txt
3. Run the Jupyter Notebook to preprocess the data, train models, and evaluate results.
4. For deployment, we integrated a FastAPI-based web interface allowing users to select models and upload data for predictions.

Dependencies

• Python
• scikit-learn
• XGBoost
• FastAPI
• Pandas, Numpy

Contributors

• Ahmed Radwan
• Sam El Saati