MedGPT

Inspiration

Having worked in a hospital setting for the majority of my time at college, I've witnessed firsthand the amount of time wasted by doctors on cookie cutter insurance forms for denial of service appeals. These forms are easily reproducible, and therefore easily GTP-able.

What it does

This web app takes in basic info about the denied treatment, the patient history, and the insurance provider. This info is then fed into an engineered prompt which returns a basic denial of treatment appeal form.

How we built it

We built it using retool and an api call to the open ai davinci model. Code is here: https://gist.github.com/simonnarang/351879958b5787ef9e100795968c36d2

Challenges we ran into

The biggest challenge is hallucination. The program still occasionally produces information which was not provided in the prompt. With that being said, often times these additions are still relevant to insurance providers, and can be edited to remain accurate. Additionally, due to the inherent limitations of the GPT model, the citations included are occasionally inaccurate.

Accomplishments that we're proud of

Having shown the results to a cohort of doctors, it has been confirmed that the responses produced are usable in a clinical context.

What we learned

With regards to prompt engineering, I did not realize how significant one word changes can alter the quality of the response. For example, adding a passage which asks the engine to impersonate a highly qualified doctor significantly improved the results.