Rainy Tuesday, Regard Chow

Rainy day keeping me pent up / not lifting, and dinner with the Regardians ate the evening (Vicky’s All Day). Did get a small core workout done in the hotel room.

My thoughts have been circling round ChatGPT writing clinical notes. From what I understand, Bing Chat exposes the very same GPT-4 that Microsoft Nuance is entreating to write physicians’ notes. I believe that it’d be possible to understand the fire Microsoft toys with using applying existing synthetic test cases mapping patient data -> clinical notes; whether it has any potential.

The value might be that, given a complete patient as input, the LLM is able to 1) extract the specific data from the patient related to each diagnosis, and 2) write the natural language from those data to document that condition. Test cases can be scaled up from existing clinical notes.

To actually query the LLM various prompting schemes can be tested, including: chain of thought, zero-shot, and few-shot. There’s a guide on prompt engineering that explains the different strategies. Honestly, the scheme of prompting narrowing the model is fascinating.