this wasn’t a smarter diagnosis

本站添加: 2026-05-10

[00:00:00]So, she'd been trying to explain it for 10 years. Every test that came back normal. So, she typed it into a chat window instead.

[00:00:12]Okay, so there's this person that actually posted on Reddit.

[00:00:18]And they said that for over 10 years unexplained symptoms.

[00:00:26]And they saw multiple specialists that did numerous tests here where it was an MRI.

[00:00:35]They put them through the MRI. [snorts] Then they did other screening full blood work testing for Lyme disease.

[00:00:47]>> [gasps] >> And they even landed up seeing a neurologist at one of the best health care systems in America.

[00:00:56]And all of this results came back normal.

[00:01:04]But she knew that something was still wrong.

[00:01:09]She had no diagnosis.

[00:01:12]So, out of options because she had kind of exhausted them all, they went to our good friend Chat GPT and put her full medical system history into Chat GPT.

[00:01:29]And then Chat GPT came back with something. And they were like, "Aha!

[00:01:37]You have a rare genetic mutation that affects how you process vitamin B.

[00:01:46]And it occurs roughly in one in 10 people."

[00:01:50]And what had happened is these test results, read in isolation of this one and this one and this results in the MRI result appeared to be normal.

[00:02:02]But then read in combination all together being able to process all this data, they pointed to somewhere specific.

[00:02:11]And what she did then she took these results to her her doctor and was like Oh.

[00:02:20]Hey.

[00:02:22]Look at this.

[00:02:24]And the doctor ran a test for this exact thing and you know what? It came back positive.

[00:02:29]Treatment started and most of her symptoms were resolved. This post kind of went viral on Reddit.

[00:02:37]So, a lot of news outlets and social media framed this as oh my gosh, chat GPT is better than doctors. And that framing isn't easy way to analyze this problem.

[00:02:52]It's not the framing that I want to talk about when I look at this type of problem because I think that framing misses a lot. Because a lot of the time the doctors they weren't actually wrong.

[00:03:03]They would doing kind of exactly what they were trained or learned how to do.

[00:03:10]And and hear me out, stay with me because I want to talk about training. Yeah. And I know that that term has now been taken over by training large language models and all this kind of stuff. But to become a doctor requires and and we all know this a lot of training. And in that training there's studying, there's learning, there's hands-on practice, there's like I don't know, there's so much that comes into being a doctor. It's what the reason why it's one of those degrees that takes like so many years and then you also need like real world experience, all of this kind of stuff. Being a doctor is hard. So, if we we look at post about medical advice. This user in their own words says, "Not sure how they didn't think to test me for the MTHFR mutation." And that question, it does have an answer.

[00:04:03]And it's not an easy answer to accept when it comes to diagnosis diagnoses. And it's not because the doctor is bad. I am not like there is a point where some doctors are bad and some doctors are good. I'm not going to get into that. It has nothing to do the My discussion today is nothing to do with the doctor, the person themselves.

[00:04:26]It's their training. So, if we look at diagnostic error, there are two patterns that show up quite repeatedly in studies. And the first is anchoring. Because once a doctor forms an early hypothesis of what is potentially wrong with you, everything gets evaluated against that. And contradictory evidence kind of gets discounted because that you kind of want the anchor to hold. It's like saying, "I think this is what you got, but we got to we got to figure out this problem." Remember, doctors are a problem-solving when it comes to humans. Engineers, they problem-solve in the real world and electricity, you know, all this kind of stuff. Doctors are problem-solving humans. And humans are incredibly complicated. So, they start at a point and they kind of find evidence to prove that point or prove Yeah, the point was they had to show their initial standpoint. They anchor. And the second pattern is premature closure. So, that means, "Oh, we know the solution to the problem." prematurely. Because once a plausible diagnosis exists, the search to figure out what the actual problem is tends to slow down, kind of stop. So, there was a study done where they looked at 100 internal med- medicine errors.

[00:05:48]and they found cognitive factors in 74 of them as the the reason for the error.

[00:05:54]And the most common one was premature closure. And these aren't like problems in like intelligence of the doctor or lack of care. It's kind of indicative of how human cognition works under pressure with limited time and across thousands of patients. And this this is where I bring in chat GPT because chat GPT doesn't anchor like humans do. It has no first impression to protect. It processes everything simultaneously and holds all possibilities open at the same time. This is one of the things that makes this AI incredible is that humans can only process a certain amount of information. These AI models can process millions at the same time. And in this if we directly compare humans to chatbots, it's not we can't say that one is better than the other, but we can frame it as some are differently limited.

[00:06:58]Hope that makes sense.

[00:06:59]So, what most of you might not know about me is my first degree was biomedical engineering which meant I did effectively first and second year medicine which meant I had the absolute privilege of being able to dissect a human body.

[00:07:15]And obviously that came with lots of textbooks and understanding where everything is in the body and what I noticed when I was um looking at these textbooks and stuff is that most of the images in the specific textbooks that I got were all of white people.

[00:07:33]And this meant certain skin conditions and everything didn't resonate with a lot of my class which was mostly black people. And this is not my own I'm not the only one to think of this. There was a 2020 study that found that only 18% of images in dermatology textbooks show dark skin. And in this it's kind of obvious to note that redness shows a different color. Early signs of melanoma, it looks different on different skin tones. And the training material here shows one body type. And therefore, a lot of medical students learn from only one body type. And then in the real world, they're interacting with other body types. And I'm not singling out individual doctors here.

[00:08:22]It's kind of about what the system builds upon. Because gaps in training data become gaps in diagnoses diagnosai?

[00:08:32]There is a level of correlation there, or causation, I should say. It doesn't mean that it happens all the time. I'm not overstating this. I'm just saying that there is a potential that that can happen. And if there are the odd percent of people that do fall through these gaps, they're the ones potentially struggling to try and get a diagnosis because it's just not this obvious first thing. Under the original Reddit post, the comments they they there were thousands of comments. Okay, maybe not thousands, hundreds of comments where people were describing their misdiagnosis from doctors, but then ChatGPT or Gemini was able to diagnose them. And like it's nice for them. Like it there's a there's something with being diagnosed that is incredibly like it it lifts a weight off your shoulders. I can't think of the English word right now. Relieved.

[00:09:23]There's a a notion of relief even if the the the information is like bad, there is a notion of relief when it comes to these things. And it's not also limited to humans. There were some cases in the comments where they were talking about a dog and their vet. And the vet suggested that like you should probably put the dog down, but then they went to ChatGPT and ChatGPT found the problem and then they were the vet agreed, "Oh, that's actually true. We can do something about it." So, there is great uses and this has been really great for a lot of people. But, these examples are not about AI being right and doctors or vets being wrong. It's more about how people had exhausted one system. So, they exhausted going to doctors because doctors were like, "We can't figure out what's wrong with you.

[00:10:15]Nothing is wrong with you. We don't know."

[00:10:18]And then they found another means to evaluate the situation and that was AI.

[00:10:24]And that different system had or has different constraints and different gaps in its knowledge and its training.

[00:10:34]But, not the same gaps and as the human.

[00:10:38]Thus, it was able to help find a diagnosis.

[00:10:42]ChatGPT is not a doctor. It can hallucinate and it does hallucinate. And it cannot examine you. It can't touch and palpate and see what's actually going on with its own eyes. It also has no professional accountability, no clinical judgment, and no way to know what it doesn't actually know. And those things don't make the Reddit post that exist there less real. The test that would have actually found this mutation also existed 10 years ago. And the system that was supposed to find it just missed it. So, there is this idea that every diagnostic system, whether it's human or machine, is built on what it's trained on. And training is never fully 100% complete. And these gaps, they are invisible from the inside of the system. And this is a generalization here, but the people that often fall through the system or the gaps here are the people that least fit the model in which the system was created or trained upon. And ChatGPT found that mutation by chance because it had access to so much more data. And maybe it wasn't in its gap in its training data. But it does ChatGPT does have its own gaps that many doctors are way better at diagnosing or sorting out. The problem now exists is which system is the right system for you? We don't know. Maybe the future is a combination of both. And maybe that will just make medicine better.

[00:12:17]I'm not sure.

[00:12:18]So, there is something here where there is a proceed with caution when using AI for medical diagnosis or medical understanding or whatever you want to call it. Um I'm not a doctor and I don't work in the medical field.

[00:12:36]But I am an advocate for always having more information.

[00:12:41]Remember, you have to figure out if that information is true or not, but always more information is better.

[00:12:49]Therefore, I do think that these AIs play a great role for um the medical field. I I as you know, I studied anatomy. And there was so much to remember. Doctors are insane. And maybe they don't remember it how many years later, but it's somewhere in their brain. But imagine like knowing where your like where certain veins, nerves, muscles, everything go in your body and being able to articulate that and understanding how this complex mechanism works. Doctors are incredible. And maybe AI can help with the retrieval of this information in a particular way.

[00:13:34]Um Yeah, this is not to like I really don't think we should take authority away from doctors. I think doctors Yeah, but AI is a tool that can help in this way. I at least think. But if you disagree with me, I'd really like to hear and educate myself as well.

[00:13:51]So, yeah. Anyway, it is Friday.

[00:13:55]Um I won't see you tomorrow, but I'll see you on Monday. Have a lovely weekend.