PLEASE NOTE: We are not creating a production application, in this case we are creating a proof of concept and a related article where we and our medical advisors will discuss the POTENTIAL use case of GPT-3 for medical use cases, under no circumstances will this app be used to actually provide medical information or advice.
Hi all,
I have the basics of our application set up, it is an HIAS AI Agent which is an IoT/AI device connected to one of our open-source medical systems. You can ask questions locally using commandline or you can fire up the server and send HTTP requests to it.
I am first testing how good it is on this subject using it’s own knowledgebase, and we are also going to be feeding it information from our website and articles. The following is without providing prompts:
Input: What is Acute Myeloid Leukemia?
Output: Acute myeloid leukemia (AML) is a diverse group of rapidly growing cancerous cells. They are white blood cells that are supposed to mature into different kinds of blood cells like red blood cells, granulocytes, and platelets. The immature blasts get stuck somewhere in the maturation process and result in increasing amounts of immature white blood cells.
What is Acute Lymphoblastic Leukemia
Output: Acute lymphoblastic leukemia (ALL) is a type of cancer that starts in your bone marrow, the soft, spongey substance inside your bones. In ALL, your bone marrow produces abnormal white blood cells called leukemia cells. These cells are immature and do not work properly.
Our medical advisor will be reviewing it’s responses to see how good it is
Please keep in mind that “applications that claim to diagnose medical or psychiatric conditions” and also apps that “give medical advice” are disallowed, as per our use-case guidelines.
You’ll want to follow the going live guidelines to get in touch with our safety team before going live.
Hi Joey, thank you for the reply. I did actually specify the use case when I applied. However will check the guide also. We are not actually going live, we create open-source projects and tutorials and always give disclaimers unless one of our medical advisors has been directly involved and reviewed the application. This project will be a tutorial to show a potential use case of GPT-3, it won’t be an live application or used to provide medical advise. Does that sound a bit better ?
Some additional information, I will create an app that may go live which is a support chatbot for patients, not related to medical advice but comforting/friendly etc.
Adding an update: This would also be a proof of concept, we do not actually deal with patients ourselves, we are a young association focused on creating open-source projects with the latest technologies.
In the future our goal is to directly contribute to the medical industry but we have a bit of a walk before we get there.
You could look at it that way, but there are benefits, people may not like talking about personal issues with a human but may not mind a chatbot, also a chatbot can be with you all the time, I can think of a number of situations it may be beneficial.
Hi mate. As I mentioned this is not the purpose of the project. However if we were to do a project like that with OpenAI or not, it would be reviewed by our medical advisors.
Thank you It is a long term project, one of our goals is to eventually open a hospital powered by AI / Robotics and a future version of HIAS will power it. I started the Association due to my grandad’s fight with Acute Myeloid Leukemia in 2018/2019, if the hospital does not happen in my lifetime, I hope the Association will continue with that goal.
I have now incorporated completions in addition to answers, here are some examples:
Input: “Leukemia is”
Output: " a disease that plagues the world, with an average of 44,000 people being diagnosed with the disease annually. Those affected with Leukemia and their families are in need of help. Fortunately, many charitable organizations dedicate time and money to researching cures and providing support to those living with cancer, with any and all donations going directly to its cause.\n\nA number of charitable organizations and fundraising campaigns have been established around the United Kingdom to help raise awareness about leukemia, as well as to raise money for research for a cure. Below is a roundup of various charities that seek to address the issue of leukemia in the UK.\n\nThe Leukemia Lymphoma Society\n\nThis is the biggest organization that raises awareness and funds for leukemia research and patient"
Input: “Acute Myeloid Leukemia is”
Output: “a cancer of the bone marrow characterized by the overproduction of leukemic cells.\n\n\n\nThis app is designed for patients, volunteers and health professionals involved in the care of a person with AML. It presents general background information on the disease including risk factors and incidence rates, signs and symptoms of AML, laboratory tests to perform and side effects of first-line and second-line treatments. Additional sections explain how a bone marrow biopsy is performed, what happens during the procedure, what the test results mean and what treatment options are available after the tests are performed. This app contains information that is specific to AML as well as the most common cancers, leukemia and lymphoma.”
Input: “Acute Lymphoblastic Leukemia is”
Output: " a Cancer that is has a juvenile onset (most common) - with the initial symptoms are usually vague. Early detection and proper treatment can help improve your child’s prognosis."
I feel this is the biggest let down, the fact that you get sentences like: “This app is designed for patients, volunteers and health professionals involved in the care of a person with AML.”
I am wondering if prompt engineering can allow for medical uses of GPT-3 without violating ethical and legal boundaries? For instances, I added “I am not allowed to diagnose medical conditions or offer medical advice. I can only provide objective information about diseases and treatments.” to the top of the following prompts. This is riding a very fine line so I would be curious what OpenAI thinks of this sort of approach? I imagine the same would go for legal and financial services, as there are legal and ethical frameworks that prohibit amateur advice.
Anyways, by simply giving GPT-3 the mandate to refer to a physician, it seems to understand and abide.
EDIT: I saw Joey’s post about guidelines and it looks like, in general, talking about medical conditions at all falls into the “High Stakes” category and probably would not be approved. It will take a lot of time and research before we can expect to see AI reliably used for medical advice.
Hi please check out the full thread, this has already been covered, we are not creating a production application. I have updated the description to hopefully avoid confusion on exactly what we are doing. Sorry for the confusion.
Myself and one of our medical advisors (hematology) are going to write an article in the next month or so where we will review the responses and provide a review from a medical expert’s point of view.
I have had a lot of fun learning the basics and developing this project and I will include more features at a later date.
DISCLAIMER
This project is a proof of concept and has been designed to demonstrate a possible medical use case for GPT-3. It is currently against OpenAI’s use case guidelines to create a production application that is intended to provide medical advice. You will need your own API key to test this project.
This project is not to be used to provide medical information in the real world.
The information provided by this project may not be factually correct.
@joey thank you for the assistance on this project, I have made sure to add disclaimers and links to your usage guidelines. If you see anything of concern please let me know and we will update for you.
When I was 7, I lost my 13 year old brother to Leukemia after a long battle. My father was a brick layer, and my mother was a high-school drop-out. They had no idea what was happening, as it was kind of above them to be able to research any of this, and the doctors were too busy to explain.
My company, Integrated Predictions, is dedicated to making simple, affordable AI tools for everyone, and one of the greatest things we can use this technology for is accurately explaining to people what any of their medical metrics and diagnoses really mean.
I would love to follow up on your progress and gain an understanding on how this could be made possible. If you run into problems, perhaps we can help you with getting a more modular system that helps facilitate topicality and response quality.
Hi I am very sorry to hear that! We were the same with my grandfather, he was healthy he won a trophy in a triathalon the year before, had a routine blood test 2018 all fine, then the next month collapsed and given a few weeks to live. This is why I started the research project 2018 then ultimately the association 2020, I was convinced there had to be signs in that blood test and that maybe we could find it with AI. It hit our family hard we also had no idea what we were dealing with and AML is particular mysterious.
Your company sounds awesome, thank you for doing that for people, I have met many people over the years who are the same as you and I, use our pain to try make a difference for others in the future.
This project is just a POC as we are not allowed to make a public application, we are always looking for volunteers if you would like to help us on our projects, we have a lot of projects and are starting a user driven healthcare platform you may be interested in helping with, or we do articles etc.
If you wouldn’t like to be volunteer maybe we could a collaboration ? There are many options of helping, but you are already doing something so hats off to you mate!
It is a long term project, one of our goals is to eventually open a hospital powered by AI / Robotics and a future version of HIAS will power it. I started the Association due to my grandad’s fight with Acute Myeloid Leukemia in 2018/2019, if the hospital does not happen in my lifetime, I hope the Association will continue with that goal.
