Diabetes Readmission Prediction (MLOps Production System)
Data Scientist / MLOps Engineer
Quick Links
On this page
End-to-end MLOps system predicting 30-day hospital readmission risk at discharge with live API + Streamlit dashboard, monitoring, explainability, and ROI analysis.
Visuals
Business value and impact dashboard
Click image to view full-size
Model performance metrics
Click image to view full-size
ROI validation analysis
Click image to view full-size
Cost-benefit analysis breakdown
Click image to view full-size
Clinical outcomes dashboard
Click image to view full-size
Problem
Hospitals need early identification of high-risk patients to reduce preventable readmissions and costs. The solution must be production-ready with monitoring, not just a notebook model.
Solution
Built a full ML pipeline on UCI Diabetes data (101,766 patients), engineered 305 features from 90 raw features, and evaluated models using patient-level grouped split and cross-validation. Trained and optimized models including LightGBM/XGBoost/CatBoost/LogReg with Optuna tuning. Deployed a FastAPI prediction API and Streamlit dashboard, added model tracking/monitoring (MLflow, Evidently, Prometheus/Grafana), and created stakeholder-ready dashboards covering model performance, risk distribution, SHAP explainability, and ROI scenarios.
Key Results
- ROC-AUC: 67.45% | Accuracy: 67.89% | Precision: 25% | Recall: 27%
- Dataset: 101,766 patients; 90 raw features → 305 engineered features
- Live Streamlit dashboard + deployed FastAPI API + docs
- Monitoring: Evidently + Prometheus + Grafana
- Explainability: SHAP-based feature importance and patient-level explanations
- Business scenario: $7.95M annual savings; 1,153.7% ROI; 3.1 month break-even