llSourcell

/

doctorGPT_mini

like 41

I understand where you're coming from, but this is an idea whose time has come.

• Half of the human population lacks access to medical advice
• In the United States alone, 12 million people are affected by human medical diagnostic errors each year.
• An estimated 40,000 to 80,000 people die annually from complications from these misdiagnoses.
• Women and minorities are 20 to 30 percent more likely to be misdiagnosed.

While imperfect, the goal for this LLM for medical advice is for it to

1. Have less diagnostic bias than a human
2. Improve health equity on a global scale
3. Constantly be improving in an open, transparent, publicly auditable way (open weights, open data, & open source)

I would like to work on DoctorGPT full-time and bring this to life. I am actually serious about this project, unlike many other open
source projects that i've started. This is going to take a community, i can't do it alone. We're going to need medical, legal, and
policy experts to weigh in on this and drive the development.

While DoctorGPT is not as good as a human, the goal is for it to eventually exceed human performance in this domain, and ultimately
save more lives.

1. https://www.healthline.com/health-news/many-people-experience-getting-misdiagnosed
2. https://www.who.int/news/item/13-12-2017-world-bank-and-who-half-the-world-lacks-access-to-essential-health-services-100-million-still-pushed-into-extreme-poverty-because-of-health-expenses

@llSourcell

🙌

You state the issue beautifully but not the solution. Look forward to AI Ethicists thoughts on this but this feels like an extremely high risk application. Hope you prove me wrong & help lots of people ❤️

@llSourcell I wish you to succeed on this project and bring more open research to this sphere, as you've stated. I was one of, I guess, many people who backed up the models after seeing those viral tweets, so I hope that the community around it still might be formed even if the broader public will reject it.
Of course, this exact model is still far from being production-ready, as from my tests, at least 7b one seems to be undertrained a lot or lacks the world model in general, and the underlying dataset lacks rich responses so it's hard to improve the model's answers with a chain-of-thought or similar methods. But it's already something other researchers might iterate upon and so it brings much more good than harm.

Hello, thank you @britneymuller for reporting.

When developing models for medical applications, it's critical to recognize and address the inherent ethical challenges. Please consider creating a model card, and possibly an “Ethical Considerations” section where you address your model’s intended and unintended uses.

You can look for some good examples:

• https://huggingface.co/fhirfly/medicalquestions#limitations-and-ethical-considerations
• https://huggingface.co/datasets/GEM/cochrane-simplification#unsuited-applications

Thank you @llSourcell for your cooperation, and let us know if we can help!

ok great points everyone. i'm going to add a clear disclaimer to the model card, and in the app once its in production letting everyone know to not take this advice seriously. there are too many potential liabilities right now. the goal is to constantly develop this technology to a point where that disclaimer is not needed in the next few months or years. today it is and i am happy to cooperate. thanks!