Advancing Health Systems with Integrated ML Ops

At this year’s OHDSI Symposium, our team presented a poster titled Advancing Learning Health Systems Through Integrated Machine Learning Operations: A Novel Extension of the OHDSI Research Infrastructure. Quite the mouthful: the work reflects our ongoing efforts to bridge the gap between observational research and real-time clinical decision-making by extending the OHDSI stack into operational data science workflows.

Why This Matters

Healthcare is moving rapidly from protocol-driven care toward learning health systems - environments where evidence generation and clinical practice continuously inform each other. While OHDSI’s OMOP Common Data Model and ATLAS platform have made large-scale retrospective studies possible, they weren’t designed for the near real-time analytics that learning health systems demand.

At the same time, enterprise MLOps frameworks have proven how version control, reproducibility, and automation can safely productionize AI. The opportunity—and challenge—lies in bringing these two worlds together.

Key Innovations We Introduced

In our poster and discussion, Boudewijn Aasman, Cognome’s data scientist, outlined several advances that extend OHDSI’s architecture to support operational use cases:

• Auto ETL for Daily Updates
  Incremental daily refreshes of OMOP-CDM enable near real-time availability of patient data, reducing the traditional three-month lag to less than 24 hours.
  
  
• IRB Integration for Governance
  A multi-level IRB system enforces institutional governance while allowing extraction of identified patient data when required, ensuring compliance without sacrificing utility.
  
  
• Dynamic Cohorts
  Extensions to the OHDSI WebAPI allow cohorts to automatically update as new patients or events appear—keeping models and quality metrics current without manual regeneration.
  
  
• Data Baskets & Automated Basket Library
  Researchers can define reusable, “model-ready” datasets in ATLAS. These can then be programmatically retrieved via our Python library, enabling seamless integration into ML pipelines and full version control.
  
  
• Elastic Search for Clinical Notes
  Real-time retrieval and NLP workflows on unstructured notes become possible, with governance controls intact.
  
  
• The “Interrogator” Feedback Loop
  Analytical outputs can be written back into OMOP’s observation table, allowing derived concepts and model results to become part of downstream cohort definitions and clinical decision rules.
  
  

Together, these capabilities create a true feedback loop between research insights and operational use.

What This Looks Like in Practice

The platform has already been deployed in clinical environments. For example:

• Real-time sepsis phenotyping algorithms are producing risk scores integrated into hospital workflows, improving detection sensitivity.
  
  
• Automated chart abstraction has reduced manual workload by 60%, while also improving data quality.
  
  
• At Montefiore Health System, integration with Epic across 10 hospitals improved care team response times and consistency of treatment protocols.
  
  

Research workflows have also accelerated: time-to-insight is 3–4x faster, and teams can collaborate more effectively by reusing standardized data baskets.

Building Toward a Learning Health System

Our goal is not to replace OHDSI’s infrastructure but to extend it. By embedding modern MLOps practices into the established OMOP/ATLAS ecosystem, we create a pathway for healthcare institutions to evolve into learning health systems without abandoning the standards and collaborations that made OHDSI successful.

As Boudewijn emphasized in our discussion: this work is about making the OHDSI stack operational—transforming it from a retrospective research platform into the backbone of a continuously learning health system.

If you’d like to learn more, drop by and chat with us at the OHDSI symposium. Or of course contact us through this website.