ReadmitRisk Data Pipeline Report

Executive Summary

This document describes the data processing pipeline used to generate the risk predictions and analytics displayed in the ReadmitRisk dashboard. The pipeline transforms raw hospital encounter data into actionable risk scores using machine learning.

1. Data Sources

1.1 UCI Diabetes 130-US Hospitals Dataset

Source: UCI Machine Learning Repository
Description: 10 years of clinical care data at 130 US hospitals (1999-2008)
Size: 101,766 hospital encounters
Key Variables: Demographics, lab results, medications, diagnoses, readmission outcomes

1.2 CMS Hospital Readmissions Data

Source: Centers for Medicare & Medicaid Services
Description: Hospital-level readmission rates and penalty information
Coverage: All 50 states + DC (simulated with realistic patterns based on actual CMS HRRP data)

2. Data Cleaning Steps

2.1 Missing Value Handling

Column	Missing Rate	Treatment
`weight`	97%	Dropped
`payer_code`	40%	Dropped
`medical_specialty`	49%	Dropped
`race`	2%	Filled with "Unknown"
Numeric features	<1%	Filled with median

2.2 Data Quality Transformations

Missing Value Indicator: UCI dataset uses ? for missing values - converted to NaN
Duplicate Encounters: Removed repeat encounters per patient (kept first encounter only)
- Before: 101,766 encounters
- After: 71,518 unique patients

2.3 Target Variable Creation

Original: 3 categories (NO, >30, <30)
Transformed: Binary (1 = readmitted within 30 days, 0 = not readmitted within 30 days)
Class Distribution: 11.2% positive (significant imbalance)

3. Feature Engineering

3.1 Numeric Features Created

Feature	Description	Formula
`age_numeric`	Age midpoint	Mapped from age ranges (e.g., `[70-80)` → 75)
`total_visits`	Prior healthcare utilization	`outpatient + emergency + inpatient` visits
`medication_intensity`	Medications per day	`num_medications / (time_in_hospital + 1)`
`num_med_changes`	Medication adjustments	Count of medications with dosage changes
`A1Cresult_abnormal`	Abnormal A1C indicator	1 if A1C > 7 or > 8, else 0

3.2 Final Feature Set

12 Numeric Features:

time_in_hospital, num_lab_procedures, num_procedures
num_medications, number_outpatient, number_emergency
number_inpatient, number_diagnoses, age_numeric
total_visits, medication_intensity, num_med_changes

8 Categorical Features (One-Hot Encoded):

race, gender, admission_type_id, discharge_disposition_id
admission_source_id, diabetesMed, change, A1Cresult_abnormal

4. Machine Learning Pipeline

4.1 Class Imbalance Handling

Method: SMOTE (Synthetic Minority Over-sampling Technique)
Before SMOTE: 63,413 negative, 8,105 positive
After SMOTE: 63,413 negative, 63,413 positive (balanced)
Applied only to training data to prevent data leakage

4.2 Model Selection

Algorithm: Logistic Regression
Rationale: Interpretable coefficients for clinical explainability
Regularization: L2 with C=0.1 (moderate regularization)
Feature Scaling: StandardScaler applied

4.3 Train-Test Split

Split Ratio: 80% training, 20% testing
Stratification: Maintained class proportions in both sets

4.4 Model Performance

Metric	Score
ROC-AUC	0.564
Average Precision	0.110

Note: The modest AUC reflects the inherent difficulty of predicting hospital readmissions. Performance is consistent with published literature on readmission prediction.

5. Risk Score Calculation

5.1 Score Generation

Risk Score = P(readmission | features) × 100

Where P(readmission | features) is the predicted probability from logistic regression.

5.2 Cost Estimation

Estimated Cost = Risk Score × $15,000 (avg readmission cost)

5.3 Risk Stratification

Tier	Risk Score Range	Count	Description
Critical	80-100%	1,931	Immediate intervention needed
Very High	70-80%	2,232	Priority outreach
High	60-70%	2,920	Proactive monitoring
Moderate	40-60%	9,224	Standard care
Low	0-40%	55,211	Routine follow-up

6. Key Risk Factors

The model identified these top factors influencing readmission risk:

6.1 Risk-Increasing Factors

Total Prior Visits (coef: +4.36) - Strong positive predictor
Number of Medications (coef: +0.28) - Polypharmacy indicator
Lab Procedures (coef: +0.10) - Diagnostic complexity

6.2 Risk-Decreasing Factors

Outpatient Visits (coef: -3.12) - Continuity of care
Prior Inpatient Visits (coef: -1.96) - Established care protocols
Emergency Visits (coef: -1.54) - Prior acute care exposure

Clinical Interpretation: Patients with high healthcare utilization history have complex cases but may benefit from established care relationships. Outpatient follow-up appears protective.

7. Output Files

7.1 Generated JSON Files

File	Description	Records
`patient_risks.json`	High-risk patient records (60%+)	7,083
`risk_summary.json`	Dashboard statistics	15 keys
`state_summary.json`	State-level metrics	51 states
`hospital_metrics.json`	Hospital-level data	746 hospitals

7.2 Data Schema

patient_risks.json

{
  "patient_id": 12345,
  "age": 75,
  "time_in_hospital": 5,
  "num_medications": 18,
  "number_diagnoses": 9,
  "number_inpatient": 2,
  "number_emergency": 1,
  "total_visits": 5,
  "num_med_changes": 2,
  "risk_score": 85.5,
  "estimated_cost": 12825.00,
  "readmitted_30day": 0
}

8. Jupyter Notebooks

Three notebooks were created to document the analysis:

01_data_exploration.ipynb
- Dataset overview and statistics
- Missing value analysis
- Target variable distribution
- Feature correlation analysis
02_patient_risk_modeling.ipynb
- Feature engineering pipeline
- SMOTE implementation
- Model training and evaluation
- Risk score generation
03_hospital_analytics.ipynb
- Geographic analysis
- State-level aggregation
- Penalty estimation

9. Limitations & Considerations

Data Age: UCI dataset is from 1999-2008; healthcare patterns may have evolved
Model Performance: AUC of 0.56 indicates modest predictive power, typical for readmission models
Geographic Data: State-level metrics are simulated based on actual CMS patterns
Cost Estimates: Based on $15,000 average; actual costs vary by condition and facility
Single Hospital System: Results may not generalize to other patient populations

10. Reproduction Steps

To regenerate the analysis:

# 1. Ensure dependencies are installed
pip install pandas numpy scikit-learn imbalanced-learn

# 2. Run the analysis script
python run_analysis_v2.py

# 3. Copy outputs to dashboard
cp data/processed/*.json dashboard/lib/

Appendix: Technology Stack

Analysis: Python 3.x, pandas, scikit-learn, imbalanced-learn
Visualization: Next.js 14, Recharts, Tailwind CSS
Data Storage: Static JSON files (pre-computed intelligence pattern)

Report generated for ReadmitRisk Dashboard v1.0 Last updated: January 2026

FilesExpand file tree

DATA_PIPELINE_REPORT.md

Latest commit

History