SIIM 2024 Update: Hackathon Edition
It was terrific to see so many friends and collaborators in Washington last week—and especially such good attendance at our in-person meeting. I'll have a post summarizing that meeting later this week.
In the meantime, a debrief on our SIIM Hackathon project: as you may recall, our team has been working on making a tool that enables us to automatically associate concepts from common ontologies to CDE Set definitions to further improve the value of the CDE Set definitions and make the Observations labeled with the CDE identifiers more useful.
Briefly, we got far enough that we can do both full-text and semantic searches against the Anatomic Locations index, RadLex, and SNOMED. We then managed to demo a front-end that enabled us perform all 6 of those searches on a given term.
Our Demo
Here's where the user enters a search term:
The user can enter the search term, select the ontologies to include, and the number of results to include from each of the 6 searches. Then (after a brief interval), you get a set of ranked results:
...so it works! (At least to a first approximation... 😄)
And the Audience Choice Award goes to...
We didn't win the Hackathon according to the judges, but we did win the Audience Choice Award! Thanks to all who voted for us for your support.
Next Steps
- LLM reranking/filtering of results
- API Backend
- Work on CDE Set definitions
- Load CDE Set definition
- Split into search terms
- Use LLM-based raw search agent
- Use LLM-based filtering agent
- LLM-based tool to attach to CDE Set definitions at levels
- Incorporation into VS Code CDE Set authoring environment
- Update front end to allow users to select true positives
- Use user selections to iterate LLM filtering/reranking agent
Thanks to the Team
- Pawel Paczuski (upmedic)
- Andrei Morariu (Smart Reporting)
- Ankit Singh (Nuance/Microsoft)
- Catie Ladd (Nuance Microsoft)
- Adam Lang (Nuance/Microsoft)
- Heather Chase (Microsoft/Nuance)
- Roshan Fahimi (MGH)
- Morris Green
- Agnes McFarlin