Information Silos & Clinical Chaos
Patient data remained trapped across fragmented regional Cerner EHR databases, preventing a unified network-wide view of patient risk at the point of care.
AI-Driven Predictive Clinical Risk
& Patient Readmission Mitigation Platform
Architected an end-to-end healthcare data and machine learning platform that predicts 30-day readmission risk, prioritizes vulnerable patients, and supports prescriptive discharge interventions across a multi-site hospital network.
Predictive Clinical Analytics
30-Day Readmission Risk
HIPAA-Aligned Architecture
0.9748 ROC-AUC
$37M Potential Savings
1.
Business Challenge
Unity Health Systems faced substantial clinical, operational, and financial pressure from high 30-day readmissions and CMS reimbursement penalties under the Hospital Readmissions Reduction Program.
Data Inconsistency Across Facilities
Different hospitals recorded identical clinical events using inconsistent structures, syntax, and descriptions, threatening predictive-model accuracy and fairness.
Resource Strain & Blind Discharges
Case managers lacked real-time risk foresight and had to review 100% of discharges instead of targeting limited post-discharge support toward the most vulnerable patients.
2.
Solution
An end-to-end predictive healthcare pipeline shifted the network from reactive descriptive reporting to proactive, prescriptive clinical intervention.
Secure Multi-Site Data Orchestration
Azure Data Factory and a centralized Self-Hosted Integration Runtime securely extracted data from regional on-premises hospitals without exposing internal firewalls.
Protected Clinical Credentials
Azure Key Vault secured connection credentials and patient-data strings within the ingestion framework.
Metadata-Driven Ingestion
A reusable ADF Control Table pattern used Lookup and ForEach activities to run daily incremental loads across hospital sites.
Azure SQL Star Schema
Unified patient and encounter data was loaded into an Azure SQL analytical warehouse centered on Fact_Encounters and detailed clinical dimensions.
Clinical Data Standardization
Local source values were normalized to ISO formats and mapped to ICD-10-CM, LOINC, and RxNorm, with Site_ID tags supporting regional comparisons.
Production ML Tournament
A Scikit-Learn Gradient Boosting pipeline addressed multicollinearity, high-cardinality memory constraints, and missing clinical values through medical median imputation.
Enterprise AI Visualization
A four-page Power BI suite linked serialized model risk probabilities back to the warehouse for financial tracking, specialty risk mapping, patient prioritization, and governance auditing.
3.
Business Impact
Potential Cost Savings Identified
The platform uncovered major network savings opportunities by accurately identifying patients inside the readmission penalization window.
ROC-AUC Model Performance
The Gradient Boosting champion model achieved approximately 97.5% predictive confidence.
Highest-Risk Patients Targeted
Clinical teams could replace a 100% discharge-review model with focused intervention for patients whose AI risk scores exceeded 80%.
Dominant Clinical Risk Signals Identified
Acuity Score accounted for 47.7% of model importance, while Heart Failure diagnoses contributed 16.7%.
Prescriptive Predicted-Actions Workflow
Machine-learning flags gave discharge planners clear guidance such as “Immediate Intervention” and “Enhanced Discharge” before patients left the hospital.
4.
Solution Architecture
Predictive Clinical Risk Architecture
Regional Cerner EHRsMulti-Site Clinical Data
Patients & EncountersDemographics and Visits
Clinical EventsDiagnoses, Labs, Medications
↓↓↓
ADF + SHIRSecure Incremental Ingestion
→
Azure SQL Star SchemaUnified Clinical Warehouse
→
Gradient BoostingReadmission Risk Scoring
↓
Power BI Clinical Risk & Intervention Suite
The architecture securely unifies regional EHR data, standardizes clinical terminology, generates risk predictions, and delivers intervention-ready intelligence.
5.
Predictive Analytics Delivery Flow
Extract
Regional EHR patient, encounter, and clinical-event data
Standardize
ISO, ICD-10-CM, LOINC, RxNorm, and Site_ID mapping
Unify
Azure SQL star schema and daily incremental loads
Predict
Gradient Boosting readmission-risk model
Intervene
Prioritized patient watchlists and predicted actions
A metadata-driven workflow converts fragmented regional data into actionable pre-discharge risk guidance.
6.
Technology Stack
Azure Data FactorySelf-Hosted Integration RuntimeAzure Key VaultAzure SQL Data WarehousePythonScikit-LearnPandasNumPyGradient BoostingPower BIDAXICD-10-CMLOINCRxNorm
7.
Dashboard Gallery
8.
Project Showcase
Live Power BI Dashboard
Replace this placeholder with the public Power BI embed URL after publishing the report.
Video Walkthrough
Replace this generic YouTube URL after publishing the walkthrough.
9.
GitHub Repository
Explore the Predictive Clinical Analytics Project
The repository contains project documentation, pipeline details, machine-learning implementation context, screenshots, and supporting assets.
Explore More Case Studies
Discover more healthcare analytics and enterprise data transformation projects.