11Research
Diabetes Prediction Algorithm Comparison
A comprehensive research study that compares various machine learning algorithms for predicting diabetes patient hospital readmission.
Project preview01 / 01
The idea behind it
This project compares various machine learning algorithms for predicting hospital readmission within 30 days for diabetic patients, using the UCI Diabetes 130-US hospitals dataset (1999-2008). The workflow includes data collection, preprocessing, exploratory analysis, supervised learning (Logistic Regression, Decision Tree, Random Forest, KNN, Gradient Boosting, MLP, XGBoost, LightGBM), ensemble learning (Voting, Stacking), unsupervised learning (K-Means clustering), model evaluation, and visualization.
What it does
- Comparison of 8 supervised learning algorithms for readmission prediction
- Ensemble learning models (Voting and Stacking)
- K-Means clustering for patient profile analysis
- SMOTE balancing for imbalanced dataset handling
- Feature engineering (comorbidity score, prior utilization, hba1c_attention)
- Comprehensive model evaluation with ROC curves and confusion matrices
- Feature importance analysis and visualization
- Patient cluster profiling with distinct readmission risks
Challenges & solutions
- Handling imbalanced dataset with low readmission rates
- Feature engineering from complex medical data
- Optimizing hyperparameters across multiple algorithms
- Managing missing and invalid values in healthcare data
- Balancing model complexity with interpretability for clinical use
Outcomes & impact
- Ensemble models (Voting, Stacking) outperformed individual models
- Identified key predictors: prior utilization, comorbidity score, A1C result
- Successfully clustered patients into distinct risk profiles
- Comprehensive evaluation metrics across 8 different algorithms
- Generated actionable insights for hospital readmission prevention