A machine learning project that analyzes 10 years of patient data from over 130 U.S. hospitals to predict the likelihood of early diabetes-related hospital readmissions.
Tech used: Python, Pandas, Scikit-learn, Imbalanced-learn (SMOTE-ENN), XGBoost, Neural Networks (MLP), Matplotlib, Seaborn, Jupyter Notebooks
The project leverages a large clinical dataset to develop a robust classifier for predicting early readmissions of diabetic patients. The pipeline includes:
- Loaded 100K+ patient records from CSV files
- Cleaned missing values and standardized categories (e.g. medications, diagnosis codes)
- Merged multi-source hospital data into a unified feature matrix
- Applied encoding techniques for categorical variables
- Evaluated feature importance using mutual information and tree-based models
- Selected top-performing features using variance thresholding and recursive selection
- Applied SMOTE-ENN (Synthetic Minority Oversampling with Edited Nearest Neighbors)
- Balanced the dataset to improve minority class prediction performance
- Logistic Regression, Support Vector Machines, Random Forests
- XGBoost for ensemble modeling
- Multilayer Perceptron (MLP) for deeper learning
- Used F1-score, Precision/Recall, Confusion Matrix, and AUC
- Visualized model comparison through bar plots and confusion matrices
- SMOTE-ENN improved minority class recall by over 15%
- Tuning hyperparameters with
GridSearchCVfor SVM and XGBoost - Compared models across metrics to choose the best trade-off between precision and recall
- Focused on clinical relevance: avoiding false negatives for patient safety
- Hands-on experience working with real-world imbalanced healthcare datasets
- Learned the tradeoffs in oversampling techniques like SMOTE vs. SMOTE-ENN
- Understood the importance of model explainability in medical ML
- Gained insight into feature selection's impact on performance and training time