Predictive modeling · MLOps · 2023 – 2025

30-Day Decompensation: Maintaining the Longest-Running ML Pipeline in VA History

Maintained and extended the VA's most mature predictive model, which flags discharged veterans at highest risk of 30-day readmission, with MLflow-based MLOps.

AI/ML Lead Data Scientist · U.S. Department of Veterans Affairs, Office of Connected Care, via InnoVet Health

Randolph Schilke

By , Senior AI Engineer
Updated

Continuously running ML pipeline in VA history
Longest
Readmission risk window
30-day
Scheduled pipelines: dev and prod
2

Architecture

How it works

  1. 01
    CDW and SDOH data
    Corporate Data Warehouse
  2. 02
    Feature store
    Scheduled preparation
  3. 03
    Inference
    Staging version in dev
  4. 04
    Validation
    Scheduled validation step
  5. 05
    MLflow registry
    Staging to Production
Staged promotion with MLflow model stages

Context

The 30-Day Decompensation program predicts which discharged veterans are most likely to be readmitted to an inpatient setting. It is the VA's most mature data science model and has run in continuous automation since inception.

It draws on the VA Corporate Data Warehouse (CDW) and social determinants of health data.

What I built

  • Maintained and extended the model and its MLOps infrastructure on Azure Databricks.
  • MLflow model management with staged promotion: the candidate version runs in a scheduled dev pipeline while the production version runs in a scheduled prod pipeline.
  • Each pipeline runs feature store preparation, inference, and validation.
  • Version history in the MLflow registry, with superseded versions archived.
MLflow registry for the Decompensation model showing a Staging version in the scheduled dev pipeline and a Production version in the scheduled prod pipeline
MLflow model stages: the Staging version runs in dev, the Production version runs in prod.

Results

  • The pipeline has stayed in continuous automation since inception, the longest run of any ML pipeline in VA history.
  • Candidate versions run in a scheduled dev pipeline alongside the production version before promotion.

Contact

Hiring for ML platform or healthcare AI work?

I'm open to Senior and Staff AI engineering, ML platform, and data science leadership roles, and to healthcare AI consulting. Email is the fastest way to reach me, and I reply personally.

Target roles
Senior or Staff AI Engineer · ML Platform / MLOps Engineer · Data Science Lead
Location
Remote-friendly; based in the Jacksonville, FL area
Work authorization
U.S. citizen
Focus
ML platforms, MLOps, LLMs, and healthcare AI
Email meConnect on LinkedInDownload résumé (PDF)

Also as Word or the web résumé.