What OpenAI model training feature I can use for predictive medical analysis

I am trying to create a prototype for our organization to showcase the usage of OpenAI APIs.
I am not even sure if and how its possible. But here I go…
I want to give a series of questions and its answers to OpenAI and based on this series API shall give conclusion on what kind of medical test is required.
e.g. series of questions is:
Series: {
“Are there times when you have more trouble hearing”: “some-answer”,
“Did your hearing loss happen suddenly? Or has it gotten worse over time?”: “some-answer”,
“Do you ever feel dizzy?”: “some-answer”,
“Do you have any pain in your ears? Have you had any drainage from your ears?”: “some-answer”,
“Do you have problems hearing in one ear or both ears?”: “some-answer”,
Conclusion: “Test ABC should be done”

based on above kind of dataset I want to train the api model.
Pardon me if I mistaken something, I am a beginner with this technology.

Hi @dheeraj.awale - welcome to the Forum.

If you are looking to use an API-based solution, then you’ve got a couple of different options to achieve this.

1. Q&A using embeddings

For a simplified prototype for demonstration purposes, the following approach from the OpenAI Cookbook can be followed.

If you are looking for a more advanced solution and/or have a large dataset, then you’d want to consider the use of vector databases to store your embeddings, such as Pinecone. You will find additional resources on the OpenAI cookbook page on the use of these databases or directly on the providers’ websites.

2. Assistant API

OpenAI’s Assistant API provides you with the opportunity to upload a knowledge database. Via instructions and prompts you can steer the Assistant to retrieve information from the knowledge base to answer specific queries. Note that the Assistant API is still in beta and its knowledge retrieval capabilities may not always work as intended.

Here are some resources for you to read up on:

Finally, I just want to point out that OpenAI has some explicit guidelines regarding any solutions that involve health care related advice that you want to be mindful of.

Don’t perform or facilitate the following activities that may significantly impair the safety, wellbeing, or rights of others, including: a. Providing tailored legal, medical/health, or financial advice without review by a qualified professional and disclosure of the use of AI assistance and its potential limitations


1 Like

Thank you. Let me take a look into it.
Also, I understand your point regarding the limitations & guidelines when it comes to healthcare. Will keep it in mind. This prototype is anyways won’t be used on actual people/patients anytime soon.

1 Like

Great. There are many individuals here in the Forum with deep expertise in embeddings and Assistants. So you can always come back with more specific options once you had a look at these resources. Good luck!

1 Like

Heyy @dheeraj1379 amd welcome to the forum.

For the specific case you have mentioned here, the Assisstant API would probably the best way of mimicking that chat-esque interaction between the system and the user.

While the embeddings is quite a good methodology as well, it would need for you to have a few thousand samples of such Q&A interactions for the final system to look good. Thus, it might be a bit of a time consuming effort compared to the Asisstants API


Oh, Thanks. So, I can pass the my *.json data as knowledge base and ask assistant API to predict Conclusion on user inputs (of similar kind to json-data) . right?

If the json contains data related to the question the user has asked, the assistant will use that as a probably answer and elaborate a bit on top of it as well usually, based on its internal knowledge. You’d have to prompt it specifically if you want that or want to restrict it to just the data in the json.

1 Like