Ovates Aid: building a medical assistant with Open AI

OpenAI now allows to create custom GPTs. They can be thought about as the bias layer added to the ChatGPT with optional ability to call APIs via commands known as actions.
In this post, we'll delve into my journey of creating "Ovates Aid", a specialized medical assistant: https://chat.openai.com/g/g-i7umTDseN-ovates-aid.

Users

First of all, as always, we think about what it is what we want to solve.
The use-case I had in my head was an assistant for a medical professional. Just like in any profession, there are some details which slip out of our minds. I was imagining the four: (1) a pharmacy worker, getting a request for simple health conditions not requiring a prescription, (2) a nurse, needing to quickly check on something what is somewhat suspicious but not suspicious enough to alert doctors, (3) a general practice doctor, needing to relate a combination of factors of a patient, (4) a medical underwriter, working through the insurance case.

For all of the medical cases, user data is highly sensitive for obvious reasons. And because of the regulations, difficulties in anonymising healthcare data -- there is a natural limit in automating this process. Therefore, an AI assistant takes an approach of simply being a working collection of medical knowledge. It does not require any sensitive data and instead assists our user by prompting questions to consider and the probable causes based on symptoms.

No surprise, there are mere mortals like us, who would like to have an understanding about the state of our health before committing to all of those complicated queues and receptions. And with all respect to all of the medical workers -- quite often we would like to have a better and cleaner description of what is the state of our health.

This is what we want Hippocrates AI to solve, but we might have some trade-offs to consider. After all, audience specified in such way is not specified at all, the range is very wide.

Setup

Creating a GPT model with OpenAI is impressively straightforward. Hit a "create" button on the gpts page: ChatGPT (openai.com) 

Follow prompts from GPT creator assistant. It will help to create a logo, specify focus and preferences of the model. While a hit the request limit per hour pretty quickly and had to wait for half an hour after the initial model was set, it was otherwise a straight-forward experience. 

With such a setup, the model was already well tuned for the requests of a kind we would expect. There was no longer a need for a quite long "setup" typing, something what would make it less useful in the actual work environment of a pharmacy or clinic. So, it got a medical focus and became careful -- preferring to ask clarifying questions than to do "bets". However, I had a thirst for something more.

Of course, the most curious part would be to integrate it with Open AI actions. But for this I would need a public AI. Probably it comes as no surprise to us mere taxpayers, but despite of the budgets of public healthcare organisations like NHS or WHO -- I couldn't find yet an API providing an up-to-date information on existing diseases/cures.  While the one could imagine an API listing all of the registered pharmaceutical brands, with registered drugs and drug's compounds for a given country -- it doesn't seem to be there yet. Or similarly, an API with known infections and known active compounds countering them. Those kinds of API would been perfect for our bot, as it would be able to gather the latest information. 

I had to settle for less, so I went for a collection of medical books I could find. Books are obviously going to grow out of date, but if we are looking for a general knowledge -- they should suffice. After a bit of digging, I found an open collection of books here: InfoBooks.org 
The last thing to do was to select 20 books to be used as a knowledge base for our GPT. I haven't seen a strong change in prompted responses, but it might be that the books dataset needs more polish.

Publish

Publishing the GPT came with a disappointment of sorts. It wasn't possible to just "publish" it to be visible to everyone. OpenAI requires a domain name to verify identity first. As I wasn't that eager for the assistant to become known fast just yet, I went for an option of the sharable link for now. If one day you would see this assistant in the public listing, it probably simply means that I won over my laziness.

Happy service to our AI overlords!