Coded Conversations: The Entity Linking Challenge 

Episode 02: Show Notes

There are vast amounts of patient data captured in unstructured clinical data, making it challenging to analyze them and extract meaningful insights. However, by applying a standardized clinical terminology like SNOMED CT, healthcare organizations can convert this free-text data into a structured format that can be readily analyzed by computers, in turn stimulating the development of new medicines, treatment pathways and better patient outcomes. 

In this episode, SNOMED International Chief Digital Information Officer Rory Davidson is joined by Will Hardman, Artificial Intelligence (AI) specialist at UK-based AI consultancy Veratai, to discuss why SNOMED International has partnered with Veratai and PhysioNet, the research resource for complex physiologic signals that is managed by members of the MIT Laboratory for Computational Physiology, to launch the Entity Linking Challenge. 

You’ll learn what entity linking involves and why the challenge was launched, as well as a few of the challenges the partners faced in designing the competition. Will explains some of the difficulties faced by annotators in dealing with codes that are ambiguous and predicts some of the exciting outcomes that the challenge is expected to produce. He also shares his predictions for the future of entity linking beyond the competition. Tune in to find out more about the Entity Linking Challenge, how you can get involved, and what winners stand to gain – including a share of the $25,000 prize pool!

Key points from this episode:

• How the idea for the Entity Linking Challenge came about.
• The barriers to using available technology to extract valuable information from clinical data.
• What entity linking involves.
• Unexpected challenges participants may encounter in the building of entity linking models.
• The role of each of the partners in the competition.
• Three exciting outcomes from the Entity Linking Challenge that Will foresees.
• The challenge of finding a dataset and how the partners came to use MIMIC-IV.
• How annotators deal with ambiguous codes.
• How participants can access the MIMIC-IV dataset and details on the training required.
• Will’s predictions for the future of entity linking beyond the competition.
• How to find out more about the competition.

Quotes:

“And really, the reason to do this Entity Linking Challenge was to show some of the things we talk about in SNOMED CT, about how do we use new technologies that fall under the AI banner and umbrella to benefit healthcare, improve healthcare data, and get SNOMED into the vast ocean of unstructured healthcare data records that are out there.” — Rory Davidson  [0:01:48]

“The question is, ‘Can you actually use the computer to figure out where the mentions of specific well-defined clinical concepts are in this data and extract them automatically?’” — Will Hardman [0:03:18]

“There's an awful lot of knowledge about health conditions and human health that's locked up in these records that's currently not available for analysis. So that's at the root of this [Entity Linking] challenge.” — Will Hardman [0:03:41]

“I think the linking part is easier. I think it's the entity recognition part that's hard.” — Will Hardman [0:23:03]

“And right now we're focusing on how do we get that structure into the data? Because once we've got that structure, that's when you can apply some of these technologies in even cooler things about the analytics and the clinical pathways and how do you improve patient healthcare and patient outcomes and that’s where we get into the really exciting stuff.” — Rory Davidson [0:23:56]

Links mentioned in this episode:

Will Hardman

Veratai

PhysioNet: MIMIC-IV

DrivenData

Blog post about entity linking 101 (Mentioned at 17:01)

SNOMED CT Entity Linking Challenge - Benchmark

Rory Davidson

Rory Davidson on X

SNOMED International