Predictive analytics platform for reducing preventable 30-day hospital readmissions
ReadmitRisk is a full-stack care management platform that identifies high-risk patients and prioritizes post-discharge interventions to reduce preventable hospital readmissions.
Key Metrics:
- 📊 282K+ patients analyzed across two clinical datasets
- 💰 $1.5B in cost exposure identified
- 🎯 122K high-risk members flagged for intervention
- 🏥 205 hospitals benchmarked with CMS penalty data
- 📈 63% ROC-AUC model performance on MIMIC-IV data
Business Impact:
- Reduces Medicare penalties (up to 3% of payments)
- Improves HEDIS scores and Star Ratings
- Optimizes care management resource allocation
- Targets $10K-$25K preventable readmission costs
Prioritizing high-risk patients and viewing personalized intervention recommendations
Real-time patient risk scoring with machine learning models trained on 282K+ patient records.
Capabilities:
- Multi-dataset analysis (UCI Diabetes + MIMIC-IV)
- Risk tier segmentation (Critical, Very High, High)
- Dynamic cost exposure calculations
- Interactive data visualizations with Recharts
Switch between UCI and MIMIC-IV datasets to compare model performance and patient populations
Prioritized patient worklist with clinical reasoning and actionable recommendations.
Features:
- Top 50 high-risk patients with sortable views
- Multi-factor cost calculations (meds, diagnoses, age, comorbidities)
- Clinical decision support ("Why High Risk?")
- Intervention protocols aligned with CMS guidelines
Calculate potential savings from targeted care management interventions.
Inputs:
- Population size
- Current readmission rate
- Intervention cost per patient
- Expected success rate
Outputs:
- Net annual savings
- ROI percentage
- Break-even analysis
- Patients needed to treat (NNT)
State-by-state CMS penalty tracking and hospital benchmarking.
Data Sources:
- 205 acute care hospitals
- CMS Hospital Readmissions Reduction Program (HRRP)
- State-level readmission benchmarks
- Penalty amount estimates
Connect any MCP-compatible AI assistant to ReadmitRisk data — no clone or install needed.
{
"mcpServers": {
"readmit-risk": {
"url": "https://readmit-risk-production.up.railway.app/sse"
}
}
}7 Tools Available:
- Patient risk lookups across UCI and MIMIC-IV datasets
- High-risk patient filtering with age/threshold controls
- Hospital readmission metrics and CMS penalty data
- Live ML predictions using trained Gradient Boosting models
- Feature importance and dataset comparison analytics
Deployed on Railway via SSE transport. Also runs locally via stdio — see mcp_server/README.md for details.
Transparent ML model evaluation with feature importance analysis.
Analytics:
- ROC-AUC curves and precision-recall metrics
- Feature importance rankings
- Dataset comparison (UCI vs MIMIC-IV)
- Validation methodology documentation
Frontend:
- Next.js 14 (App Router) - React framework with server components
- TypeScript - Type-safe development
- Tailwind CSS - Utility-first styling
- Recharts - Interactive data visualizations
- Dark mode - System preference support
Backend/ML:
- Python 3.11 - Data processing and ML training
- scikit-learn - Gradient Boosting and Logistic Regression models
- Pandas/NumPy - Data manipulation
- SMOTE - Class imbalance handling
- Google BigQuery - MIMIC-IV data extraction
Data Sources:
- MIMIC-IV (211K admissions) - ICU clinical database from MIT
- UCI Diabetes (71K patients) - Hospital readmission records
- CMS HRRP (205 hospitals) - Public penalty data
1. Data Extraction
├── MIMIC-IV: Google BigQuery (PhysioNet credentials required)
└── UCI: Kaggle public dataset
2. Feature Engineering
├── 61 clinical features (MIMIC)
├── 12 diabetes metrics (UCI)
└── Demographic normalization
3. Model Training
├── SMOTE oversampling (8.8% → 50% positive class)
├── Gradient Boosting Classifier
├── 80/20 train-test split
└── Hyperparameter tuning
4. Evaluation
├── ROC-AUC: 63% (MIMIC), 56% (UCI)
├── Precision-Recall curves
└── Feature importance analysis
5. Risk Scoring
├── Probability thresholds (60%, 70%, 80%)
└── Cost estimation ($10K-$25K range)- Node.js 18+ (for dashboard)
- Python 3.11+ (for ML pipeline)
- PhysioNet credentials (optional, for MIMIC-IV data)
# Clone the repository
git clone https://github.com/NateDevIO/readmit-risk.git
cd readmit-risk/dashboard
# Install dependencies
npm install
# Start development server
npm run devOpen http://localhost:3000 to view the dashboard with pre-loaded UCI data.
See MIMIC_SETUP_GUIDE.md for detailed instructions on:
- PhysioNet credentialing
- Google BigQuery configuration
- Data extraction and processing
- Model retraining
readmit-risk/
├── dashboard/ # Next.js frontend application
│ ├── app/ # App router pages
│ │ ├── dashboard/ # Main analytics dashboard
│ │ ├── care-queue/ # Patient worklist
│ │ ├── impact-calculator/ # ROI calculator
│ │ ├── geography/ # State analysis
│ │ └── model-performance/ # ML metrics
│ ├── components/ # React components
│ ├── lib/ # Data and utilities
│ └── public/ # Static assets & reports
├── mcp_server/ # MCP server (deployed on Railway)
│ ├── server.py # FastMCP tool definitions
│ ├── data_loader.py # Lazy-loading data store
│ ├── train_model.py # Model training script
│ └── models/ # Trained model artifacts
├── data/ # Processed datasets
│ ├── processed/ # UCI + MIMIC data
│ └── mimic_*/ # MIMIC raw data (gitignored)
├── notebooks/ # Jupyter analysis notebooks
├── *.py # Python ML pipeline scripts
└── docs/ # Documentation and screenshots
Hospital Readmissions Reduction Program (HRRP):
- Targets 6 condition-specific readmission measures
- Penalties up to 3% of Medicare payments
- Affects 2,500+ hospitals annually
HEDIS Metrics:
- Plan All-Cause Readmissions (PCR)
- Impacts Medicare Advantage Star Ratings
- Influences member retention and revenue
Risk stratification enables targeted deployment of proven interventions:
- Transitional Care: Post-discharge phone calls within 48 hours
- Medication Reconciliation: Pharmacist review to prevent adverse drug events
- Care Coordination: PCP follow-up scheduling within 7 days
- Patient Education: Teach-back methods for self-care
See About Page for full clinical context and citations.
- MIMIC-IV Dataset: 63.0% ROC-AUC (211K admissions)
- UCI Dataset: 56.4% ROC-AUC (71K patients)
- High-Risk Identification: 122K patients (43% of total)
- Cost Exposure: $1.5B identified across high-risk population
- Intervention ROI: 150-250% with $250 intervention costs
- Resource Optimization: Focus care teams on top 10% highest-risk patients
- Health Plans: Medicare Advantage Star Ratings improvement
- ACOs: Shared savings program performance
- Hospitals: HRRP penalty avoidance
- Care Management Teams: Patient prioritization and workload optimization
- Executive Report (Combined) - Comprehensive analysis
- MIMIC-IV Analysis - ICU dataset insights
- UCI Diabetes Analysis - Diabetes readmissions
- Geographic Analysis - CMS penalties by state
- Methodology (MIMIC) - BigQuery pipeline
- Methodology (UCI) - Feature engineering
- MIMIC-IV data requires PhysioNet credentialed access
- Patient data excluded from git repository (see
.gitignore) - Only aggregated statistics and models shared publicly
- Complies with MIMIC-IV Data Use Agreement
- UCI Diabetes dataset (publicly available via Kaggle)
- CMS hospital metrics (public HRRP data)
- Aggregated summary statistics
# Frontend (dashboard)
npm run dev # Start dev server
npm run build # Production build
npm run lint # ESLint check
# Backend (ML pipeline)
python extract_mimic_cohort.py # Extract MIMIC data from BigQuery
python mimic_feature_engineering.py # Process features
python generate_full_mimic_dashboard_data.py # Generate dashboard JSON# Frontend type checking
npm run type-check
# Python environment
python -m venv venv
source venv/bin/activate # Windows: venv\Scripts\activate
pip install -r requirements.txt- Real-time EMR integration (HL7/FHIR)
- Multi-hospital deployment support
- Advanced NLP on clinical notes
- Causal inference models (uplift modeling)
- Mobile app for care coordinators
- Automated intervention tracking
Built by a healthcare data analyst passionate about using predictive analytics to improve patient outcomes and reduce preventable costs.
This project is a demonstration/portfolio project.
Data Licenses:
- MIMIC-IV: PhysioNet Credentialed Health Data License 1.5.0
- UCI Diabetes: CC0 Public Domain
- CMS HRRP: U.S. Government Public Data
- MIT Lab for Computational Physiology - MIMIC-IV database
- UCI Machine Learning Repository - Diabetes dataset
- Centers for Medicare & Medicaid Services - HRRP public data
- PhysioNet - Clinical data access platform
Reducing preventable readmissions through data-driven care management
© 2026 ReadmitRisk. Demonstration project.