OHDSI Home | Forums | Wiki | Github

LLMs and OMOP-shaped data (other LLM enthusiasts?)

In Martijn’s presentation at 2023 symposium, he presented individual patients beeing fed to an LLM.

I would like to start a discussion how to best provide medical history of a patient to an LLM (pointer to spreadsheet data that is in OMOP shape (tab for each table); conversion to a narrative text; other?).

All this is in the context of Python/R script that calls LLM APIs and uses the most advanced way of providing prompt data. (no mouse and click and simply using the chat on the website).

My goal would even be developing a reference code using Eunomia data.

In proprietary mode, I am sure many are doing it and it would be great to compare notes. (maybe even start some circle of LLM enthusiast). (or have a chapter in Book of OHDSI on use of OMOP shaped data with LLMs).

I am assuming they were doing SQL generation and consequent narrative construction with something like this.

A discussion on strategies and experiences would be useful by those who have done enough of this.

I would wager that this would turn into a research topic of its own in the coming years: The right tool chaining, right mix of external functions, agent models, system prompt characteristics, even as all of these would be very LLM specific.