Is anyone having success with ChatGPT and OMOP?
I was impressed with the quality of code for questions such as:
Write a SQL query against the OHDSI CDM to build a cohort of patients with new onset diabetes
Write a SQL query against the OHDSI CDM to calculate the numerator and denominator for the HEDIS measure for Comprehensive Diabetes Care
In general, the SQL was completely accurate or only slightly off. The codes, on the other hand, were spotty. For the first one, it knew to use a code for HbA1c. For the others, it guessed wrong. This is still an accelerator, because it is easy to use Atlas to find the right set of concept codes and paste them in.
I expect that with custom ChatGPT/LLM training and/or an integration that calls the Phoebe API to get a more appropriate set of codes, it may be possible to prototype some queries via LLM prompts.