Predicting 30-Day Hospital Readmission Among Patients With Diabetes

View the fully rendered Jupyter Notebook

Project Overview

This project develops and evaluates an interpretable machine-learning workflow for predicting whether a patient with diabetes will be readmitted to a hospital within 30 days of discharge.

The analysis emphasizes careful validation, patient-level leakage prevention, probability calibration, and clinically interpretable model explanation rather than model complexity alone.

Dataset

The project uses the Diabetes 130-US Hospitals for Years 1999–2008 dataset from the UCI Machine Learning Repository.

The original dataset contains:

• 101,766 hospital encounters
• 71,518 unique patients
• Clinical, utilization, diagnosis, medication, and discharge-related variables
• A three-category readmission outcome: <30, >30, and NO

For this project, the target outcome was defined as:

• 1 = readmitted within 30 days
• 0 = not readmitted within 30 days

Analytical Workflow

The workflow includes:

1. Data loading and initial audit
2. Binary outcome construction
3. Missing-value assessment
4. Patient-level train–test splitting to prevent leakage
5. Clinical eligibility filtering
6. Diagnosis-code grouping
7. Feature encoding and preprocessing
8. Leakage-safe cross-validation
9. Comparison of logistic regression, random forest, and gradient boosting
10. Protected test-set evaluation
11. Probability calibration
12. Risk-threshold selection using development data only
13. Permutation-based predictor importance
14. Clinically interpretable risk summaries with confidence intervals

Models Compared

Model	Cross-Validated ROC-AUC
Balanced logistic regression	0.663
Random forest	0.670
Gradient boosting	0.673

Gradient boosting provided the strongest overall performance and was selected for final evaluation.

Protected Test-Set Results

Metric	Result
ROC-AUC	0.671
Average precision	0.215
No-skill average-precision benchmark	0.109
Calibrated Brier score	0.093
Calibrated log loss	0.325
Selected calibrated risk threshold	0.125
Sensitivity	0.547
Specificity	0.696
Balanced accuracy	0.622
Percentage flagged as higher risk	33.03%

Calibration Insight

The raw class-balanced gradient-boosting probabilities substantially overestimated absolute risk.

Sigmoid calibration improved probability accuracy:

Measure	Raw probabilities	Calibrated probabilities
Mean predicted probability	0.451	0.114
Observed readmission rate	0.109	0.109
Brier score	0.217	0.093
Log loss	0.623	0.325

Calibration improved probability interpretation without changing risk-ranking performance.

Key Predictive Insights

Prior Inpatient Utilization

Prior inpatient utilization was the strongest predictor of 30-day readmission risk.

Prior inpatient encounters	Observed readmission rate
0	8.77%
1	14.00%
2	20.02%
3	22.12%
4 or more	30.92%

Discharge Destination

Discharge destination was the second strongest predictor. Encounters involving rehabilitation or continued institutional care showed higher observed readmission rates than routine discharge to home.

Publication-Style Figures

Readmission Risk by Prior Inpatient Utilization

Readmission Risk by Discharge Destination

Predictor Importance

Precision-Recall Curve

ROC Curve

Interpretation Boundary

This project demonstrates an interpretable ML workflow for risk stratification.

The findings represent predictive associations, not causal effects. The model is a portfolio demonstration and should not be treated as a clinically validated decision tool.

Tools Used

• Python
• pandas
• NumPy
• matplotlib
• scikit-learn
• Google Colab

Repository Files

• Diabetes_Readmission_Prediction_Project.ipynb
• readmission_predictor_importance.png
• readmission_risk_by_prior_inpatient_utilization_with_ci.png
• readmission_risk_by_discharge_destination_with_ci.png
• readmission_precision_recall_curve.png
• readmission_roc_curve.png
• readmission_model_performance_summary.csv
• readmission_model_key_findings.csv

Professional Positioning

This project supports the following professional positioning:

PhD Statistical Consultant | Advanced Quantitative Modeling (R, Mplus)

Supporting capability: interpretable machine learning and reproducible statistical computing in Python.

Professional website: https://drimransarmad.com