## ## ##

 

CASE STUDY

Machine Learning on a Medical Ontology

Developing machine learning tools to validate new releases of the world's leading Clinical Ontology: SNOMED-CT.

The Client

SNOMED CT, developed by SNOMED International, is the worlds most comprehensive and precise clinical terminology service. It contains more than 350,000 concepts. It has been adopted by the public healthcare systems of more than eighty countries worldwide.

As adoption of SNOMED CT accelerates worldwide, the community of practitioners is also expanding. This has led to an increase in the number of new biomedical domains that the ontology must cover, with a corresponding increase in new clinical concepts that must be researched and validated.

Responding to this increase in demand, SNOMED International has transitioned from biannual to monthly release cycles, ensuring that the ontology keeps pace with the rapid evolution of clinical research and practice.

THE CHALLENGE

To achieve its ambition of monthly releases to accelerate the evolution of the ontology, SNOMED International faced new challenges, including a greater draw on the skills of its terminologists and scientific experts to perform comprehensive validation of new concepts.

SNOMED International is committed to assuring the quality of SNOMED CT through automated and manual means, leveraging innovative approaches where possible and practicable. It is for this reason that Unai were asked to help develop machine learning approaches to concept validation, focused on supporting terminologists and other experts to increase velocity without sacrificing quality.

What Unai Delivered

  • Natural Language Processing Model. Unai developed a Natural Language Processing (NLP) model capable of identifying both new concepts that bear strong similarities to existing entries, and concepts that might be better suited to a different location within the ontology.

  • Anomaly Detection Model. A machine learning algorithm - trained on historic SNOMED CT releases - capable of assigning an accurate probability that there is an error in the definition of a new concept.

  • Review Process Integration. A REST API that integrated the two machine learning models into SNOMED CT's existing infrastructure and enabled the models to be invoked during the review process.

The Result

The models delivered by Unai provided new validation capabilities for the SNOMED CT ontology, highlighting for the attention of a terminologist the "needles in the haystack": concepts which most likely require correcting.

It is an example of "AI" supporting a human-intensive process and reducing the risk of human error. SNOMED International is therefore able to allocate its specialist resources sustainably, focusing human intervention where it is most needed.

Throughout the span of the project with Unai, our experience has been positive. The Unai team has been professional, responsive and detail-oriented in their production of our agreed-upon deliverables and open to discussion when alternative approaches are proposed. In our experience, Unai has been a valuable project partner.

Rory Davidson. Chief Information Officer, SNOMED International