Healthcare Clinical Analytics¶

Source: examples/healthcare-clinical/README.md — this page is rendered live from that file.

Healthcare Clinical Analytics — Patient Outcomes & Readmission Risk¶

Examples > Healthcare Clinical Analytics

[!TIP] TL;DR — Clinical analytics platform that ingests EHR/FHIR encounter, patient, and diagnosis data through a medallion architecture to produce readmission risk scores, length-of-stay optimization insights, and CMS quality measure reporting. All data is synthetic and fully de-identified under HIPAA Safe Harbor; no real PHI is included anywhere in this example.

Overview¶

This example demonstrates a production-ready clinical analytics pipeline built on Azure Cloud Scale Analytics (CSA). It transforms raw EHR encounter data into actionable insights for care quality improvement, readmission reduction, and regulatory reporting through bronze/silver/gold layers.

[!WARNING] No real Protected Health Information (PHI) is included. Every record is synthetic, computer-generated data satisfying HIPAA Safe Harbor de-identification (45 CFR 164.514(b)(2)). Do not load real patient data without a full HIPAA security assessment.

Architecture Overview¶

flowchart TB
    subgraph Sources["Data Sources"]
        S1[EHR / FHIR Server]
        S2[ADT Feeds]
        S3[Claims / Billing]
    end

    subgraph Bronze["Bronze Layer — Raw Ingestion"]
        B1[Raw Encounters]
        B2[Raw Patients]
        B3[Raw Diagnoses]
    end

    subgraph Silver["Silver Layer — Standardized Clinical"]
        SV1[Fact: Encounters]
        SV2[Dim: Patients]
        SV3[Dim: Diagnoses]
    end

    subgraph Gold["Gold Layer — Analytics & Reporting"]
        G1[Readmission Risk Scores]
        G2[Quality Measures Report]
        G3[Length-of-Stay Analysis]
    end

    subgraph Consumers["Consumers"]
        C1[Power BI Clinical Dashboards]
        C2[Care Management Teams]
        C3[Quality / Compliance Officers]
        C4[Executive Leadership]
    end

    S1 --> B1
    S1 --> B2
    S1 --> B3
    S2 --> B1
    S3 --> B3

    B1 --> SV1
    B2 --> SV2
    B3 --> SV3

    SV1 --> G1
    SV1 --> G2
    SV1 --> G3
    SV2 --> G1
    SV3 --> G1
    SV3 --> G2

    G1 --> C1
    G1 --> C2
    G2 --> C1
    G2 --> C3
    G3 --> C1
    G3 --> C4

Prerequisites¶

Azure Resources¶

• Azure Subscription with Azure Data Lake Storage Gen2
• Azure Databricks Workspace or Microsoft Fabric Lakehouse
• Azure Key Vault (for connection strings and secrets)
• Power BI Premium or Pro capacity (for dashboards)

Tools Required¶

• Azure CLI >= 2.50
• dbt-core >= 1.7 with dbt-databricks or dbt-fabric adapter
• Python >= 3.10

Permissions¶

• Storage Blob Data Contributor on ADLS Gen2
• Databricks workspace access with cluster create/attach privileges

Quick Start¶

1. Clone and Navigate¶

git clone https://github.com/your-org/csa-inabox.git
cd csa-inabox/examples/healthcare-clinical

2. Upload Sample Data to Bronze¶

az storage blob upload-batch \
  --destination bronze/clinical \
  --source data/ \
  --account-name csadatalakedev

3. Run dbt Models¶

cd domains
dbt seed --profiles-dir .
dbt run --profiles-dir .
dbt test --profiles-dir .

4. Explore Gold Layer¶

Connect Power BI to the gold schema tables (rpt_readmission_risk, rpt_quality_measures, rpt_los_analysis) or query them directly in Databricks SQL.

Data Pipeline¶

Bronze Layer — Raw Ingestion¶

Raw data lands in ADLS Gen2 preserving the original schema from EHR/FHIR exports. No transformations are applied at this stage.

Silver Layer — Standardized Clinical¶

Cleansed, validated, and enriched with derived clinical attributes. Encounters gain length-of-stay calculations and readmission flags. Diagnoses are grouped into Clinical Classifications Software (CCS) categories.

Gold Layer — Analytics & Reporting¶

Business-ready views designed for direct consumption by Power BI dashboards and clinical operations teams.

Sample Analytics Scenarios¶

1. Readmission Risk Prediction¶

Identify patients at highest risk of 30-day readmission using a composite score that weights diagnosis complexity, prior admission history, length of stay, and patient risk tier. Care management teams use these scores to prioritize post-discharge follow-up calls and home health referrals.

SELECT patient_id, readmission_risk_score, risk_tier, top_risk_factors
FROM gold.rpt_readmission_risk
WHERE risk_tier IN ('High', 'Critical')
ORDER BY readmission_risk_score DESC
LIMIT 25;

2. Length-of-Stay Optimization¶

Compare actual vs. expected length-of-stay by DRG and facility to identify variation drivers. Facilities with consistently higher-than-expected LOS may benefit from care pathway standardization or discharge planning interventions.

SELECT facility, drg_code, avg_los_days, expected_los_days,
       ROUND(avg_los_days - expected_los_days, 1) AS los_variance
FROM gold.rpt_los_analysis
WHERE encounter_count >= 10
ORDER BY los_variance DESC;

3. Quality Measure Reporting¶

Calculate CMS all-cause readmission rates by condition category for value-based purchasing programs. Track trends over rolling 12-month windows and compare against national benchmarks.

SELECT condition_category, measure_period,
       readmission_rate, national_benchmark,
       CASE WHEN readmission_rate > national_benchmark THEN 'Above'
            ELSE 'At or Below' END AS benchmark_status
FROM gold.rpt_quality_measures
ORDER BY measure_period DESC, readmission_rate DESC;

Data Products¶

Data Contract¶

The contracts/clinical.yml data contract defines sources, column-level tests, and quality expectations for the clinical data pipeline.

Key guarantees: - Freshness: Encounters available in Silver within 24 hours of discharge - Completeness: >= 99% of encounters have valid patient and diagnosis linkage - Accuracy: ICD-10 codes validated against CMS reference; readmission flags verified against discharge dates

HIPAA and Compliance Considerations¶

This example is built around synthetic, fully de-identified data and does not contain any real Protected Health Information (PHI).

De-identification Approach¶

All sample data satisfies the HIPAA Safe Harbor method (45 CFR 164.514(b)(2)):

• No direct identifiers: No names, Social Security numbers, medical record numbers, or device identifiers
• Date generalization: Only month/year granularity for admit and discharge dates
• Geographic generalization: ZIP codes truncated to 3-digit prefix
• Age generalization: Ages represented as age groups, not exact values
• Small cell suppression: Any category with fewer than 5 observations is suppressed

Production Deployment Considerations¶

When adapting this example for real clinical data, ensure the following are in place before loading any PHI:

1. HIPAA Security Risk Assessment completed and documented
2. Business Associate Agreements (BAAs) executed with all cloud providers
3. Azure HIPAA/HITRUST compliance configuration enabled (encryption at rest, TLS in transit, audit logging)
4. Role-based access control with minimum necessary access principles
5. Audit logging enabled on all data stores and compute resources
6. Data Loss Prevention (DLP) policies configured to detect PHI in unauthorized locations
7. Breach notification procedures documented and tested

Directory Structure¶

healthcare-clinical/
├── contracts/               # Data product contracts (schemas, tests)
│   └── clinical.yml
├── data/                    # Synthetic sample data (NO real PHI)
│   └── sample_encounters.csv
├── domains/                 # dbt models (bronze / silver / gold)
│   ├── bronze/
│   │   ├── stg_encounters.sql
│   │   ├── stg_patients.sql
│   │   └── stg_diagnoses.sql
│   ├── silver/
│   │   ├── fct_encounters.sql
│   │   ├── dim_patients.sql
│   │   └── dim_diagnoses.sql
│   └── gold/
│       ├── rpt_readmission_risk.sql
│       ├── rpt_quality_measures.sql
│       └── rpt_los_analysis.sql
└── README.md                # This file

Related¶

• CSA-in-a-Box Examples — Full example catalog
• Tribal Health Example — Related health domain with Indian Health Service focus
• CMS Quality Measures — CMS measure specifications
• FHIR R4 Specification — HL7 FHIR data standard
• HIPAA Safe Harbor De-identification — HHS guidance on Safe Harbor method
• CCS Categories — AHRQ Clinical Classifications Software