Submitted 2 days ago by misk@sopuli.xyz to technology@lemmy.zip
I feel like this is the self driving car thing again.
How often are human doctors wrong in their diagnoses?
How often are LLM doctors wrong in their diagnoses?
I'm pretty sure the former is close to 75%, and the latter substantially less. I've heard of so many people go to doctor after doctor and not get the right diagonsis or treatment for whatever they have going on, and it takes 5+ to find the one who figures it out and gets them treated.
Username about proof, unsubstantiated claims.
There are no arguments otherwise here.
I'm not here to write beautifully crafted arguments when no one has yet to disagree.
Well, I won’t make up numbers (like you did) but two things
The thing that LLMs are great at is taking a LOT of datapoints and coming to a conclusion based on all of them.
Humans can look at a few but get overwhelmed.
if you feed a ton of diagnostic data including past incidents, blood test, perhaps DNA tests, i'm pretty sure LLMs will be able to better figure out a diagnosis than a doctor using traditional methods.
When users self-diagnose, they're often wrong, because they don't know what the fuck they're doing. Garbage in garbage out regardless of the entity trying to process it.
This study is one that put doctors against a LLM, 90% accuracy for chatgpt, 74% for doctors not using LLM tool. https://www.advisory.com/daily-briefing/2024/12/03/ai-diagnosis-ec
So chatgpt wrong 10% of the time, doctor wrong 26% of the time. 2.6x worse failure rate by real docs... for that one anyway. The better the data for chatgpt, the better it's diagnosis. Humans probably won't get much better, but LLMs? I bet they will.
We're likely to have an intermediary step where HCPs handle the symptoms, testing, etc and then it's fed into a medical focused LLM. The LLM will output potential diagnosis for a doctor to review for sanity, even though the doctor is probably less accurate it will make everyone feel better, and then the doc will slap a diagnosis on their profile.
LLMs will be infinitely better than humans at figuring out drug interactions (it's just a big fucking database), allergies (they can't forget you're allergic to NSAIDs like my wife is, and who routinely has been given them by HCPs who fuck up.) Who knows what else.
and if the argument is "Bbbuut the LLM was wrong once and someone DIED!"
The comparison is the human being wrong over and over and over and over to the result of countless deaths. Malpractice lawsuits must be rare compared to the amount of mistakes that are made, simply because it's difficult to get to the point where you win, and extremely costly if you fail the suit.
We already have people posting on social media for medical advice. LLMs just can't be worse than that.
You can at least sue a doctor for malpractice if they make a mistake. If you follow medical advice from a chatbot and you die, who is liable?
Large Language Models were built to rewrite emails, not provide valid medical advice
Yeah, definitely share your medical records with chatbots.
Here is the really scary part: so many doctors were using Google lately anyways… Now they are turning to medical llms.
vivalapivo@lemmy.today 2 days ago
…as the medical costs surge