Show HN: ChatGPT for Med-School and Healthcare | Hacker News Reader

Show HN: ChatGPT for Med-School and Healthcare

54 points_vsaz56 comments

Q:I suddenly lost my head, what can be done?

A:Based on the information provided in the text, if someone suddenly loses their head, the immediate action that can be taken is to encircle the head just above the ears and eyebrows with a tie, shoelaces, or a piece of string and tie it tight. Then, insert a pen, pencil, or stick into the loop and rotate it so that the tourniquet exerts pressure on the arteries. This can help control bleeding and potentially save the person's life.

No doubt, it was fine-tuned with Monty Python.

'Tis but a scratch.

This is terrifying. I have no idea how accurate it is, but the hallucinations of regular ChatGPT worries me to no end when it’s something important like healthcare.

This does appear to be doing some sort of semantic search retrieval; as well - try a query real fast. With that said, yeah, I'm not trying to actually use this.

Yes, it's retrieval augmented generation, aka IR + NLP

Yep, that's exactly why we're building it. Our goal is to make generative output explainable and interpretable. Right now we're doing this by showing the retrieved context, which is fed directly to the LLM to perform generation. You can then verify for yourself whether or not the generated content is hallucinated from the retrived document

So this is just like Google? Which shows me quick answers but highlighted from different sources and linking to them

Yeah, like Bard or BingGPT but more focused.

> hallucinations of regular ChatGPT worries me to no end

Why? I mean, I see very much potential in LLM providing hints to doctors based on the symptoms about things the doctors would not think themselves. I also assume doctors would quite easily note majority of baseless hallucinations. So there remains two problematic cases, first, where both doctor and LLM misses the actual reason for the symptoms and second, where LLM makes a mistake that makes the doctor to apply wrong remedy without realizing it. First one is something you can't blame the LLM, and second one, well, what makes you think it outweights the benefits?

There's no problem if it learned properly.

I don't know about this but I do know there's a lot of potential using SOTA LLMs for diagnosis in general.

Use of GPT-4 to Analyze Medical Records of Patients With Extensive Investigations and Delayed Diagnosis

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10425828/

"Responses by GPT-4 and clinicians were collected and compared. Differential diagnoses were also generated using a medical diagnostic decision support systemIsabel DDx Companion; Isabel Healthcare)

Six patients 65 years or older (2 women and 4 men) were included in the analysis. The accuracy of the primary diagnoses made by GPT-4, clinicians, and Isabel DDx Companion was 4 of 6 patients (66.7%), 2 of 6 patients (33.3%), and 0 patients, respectively. If including differential diagnoses, the accuracy was 5 of 6 (83.3%) for GPT-4, 3 of 6 (50.0%) for clinicians, and 2 of 6 (33.3%) for Isabel DDx Companion"

Fantastic claims require fantastic evidence.

Please be careful with things like this before someone gets hurt or worse.

I have friends in the medical profession who have tested GPT4 and it’s good, but not quite good enough for them.

I wouldn’t touch this with a 20 mile stick; something that claims actual medical advice like this.

Yep, as a medical student, I strongly believe we need explainable AI for healthcare, and that AI only empowers clinicians to make the final call, and does not do any medical decision making autonomously. I really hope other players in the space like Hippocratic AI [1] and Google Med-PaLM2 [2] understand this as well.

[1]: "www.reuters.com/business/healthcare-pharmaceuticals/hippocratic-health-raises-50-mln-seed-funding-build-ai-model-2023-05-16/"

[2]: https://cloud.google.com/blog/topics/healthcare-life-science...

Do you even need ai? What about just a basic system where you type in some keywords like the symptom presented, it queries a known good source like the physicians desk reference or something, and returns results with a likelihood score to match your keywords? Such a systems seems like it would be useful and also would not need an ai model or expensive training to create.

> Do you even need ai?

Doesn't matter if you need it, AI is used here as a magnet for VC money.

Click on suggested query: " What are 5 symptoms of chronic inflammation"

It doesn't really answer but a button says "refer to full source" with this website that appears to open up a PDF version of a pathophys book, neat but legal? Would make sense to train things off of text books but how is this happening and how long will this last?

https://storage.googleapis.com/radiant-ai-pdf-data-prod/raw_...

We've built an automated scraper and data ingestion ETL, which crawls search engines for medical PDFs. We'll definitely need to add in compliance and verification to make sure we're not crawling any protected documents.

While I appreciate the sources being available, it seems like you're not just linking to where you got them but you have a static cache that you are linking to, which is being a primary redistribution source. I would presume it's only a matter of time until the lawsuits over copyright happens.

If he’s getting lawsuits it probably means he’s getting a lot of users and that would actually be a problem he wants to have eventually

Q: 29M ate a slug and is now having seizures.

> Based on the information provided, there is no direct answer to the question of why a 29-year-old male who ate a slug is now having seizures.

(A google search will get you the right answer.)

Q: 35M presents with crescendo angina.

The 35-year-old male is experiencing painless testicular enlargement, a sensation of testicular heaviness or a dull ache in the lower abdomen, and possibly acute pain due to rapid growth resulting in hemorrhage and necrosis. He may also have symptoms related to metastases, such as lumbar pain, cough, dyspnea, bloody sputum, difficulty swallowing, neck swelling, alterations in vision or mental status, papilledema, and seizures. These symptoms are suggestive of testicular cancer.

(Note: "crescendo angina" is a cardiology term that we will use to refer to people who have chest pain that has become more severe, more frequent, or occurring with less exertion. It has nothing to do with testicular cancer. The character I invented here was someone with familial hypercholesterolemia.)

Q: 67F with positive Kussmaul sign

Based on the information provided in the text, there is no mention of a 67-year-old female or the presence of Kussmaul sign. Therefore, it cannot be determined whether the patient has a positive Kussmaul sign based on the given information.

(Note: a positive Kussmaul sign should make you think of cardiac tamponade as well as other serious but less urgent cardiac conditions.)

Thanks for all the queries. We're still very much WIP - didn't even mean to launch til next week and suddenly found our site on HN! Will be adding a new module which should address all of these questions :)

But one reason that you are getting these is that our underlying knowledge base lacks specific knowledge on these topics. However, that's only a question of time as we scrape more documents from the web.

  > scrape more documents from the web

What types of documents, exactly? This is not the comforting reply I was hoping to see.

Oh wow, I read about this one recently on the Wikipedia page for Unusual Deaths.

Which is ... really something. https://en.wikipedia.org/wiki/List_of_unusual_deaths

Now also on Ars Technica[1].

1. https://arstechnica.com/health/2023/09/worm-that-jumps-from-...

That was a long, wild ride. Baseballs and laughter to be avoided.

EDIT: and wine glasses.

Homepage offered a suggested Q of "What are 5 symptoms of chronic inflammation?", then provided this answer:

> The text does not provide specific symptoms of chronic inflammation. It mainly discusses the association between chronic inflammation and the development of cancer, as well as the causes and examples of chronic inflammatory conditions. Therefore, it is not possible to extract 5 symptoms of chronic inflammation from the given text.

Tried my own query ("lower back pain, rib pain, uveitis," which are three hallmark signs of spondylitis) and got an equally useless answer:

> Some examples of chronic inflammatory conditions that can cause cancer include solar irradiation, asbestos exposure (mesothelioma), pancreatitis, and infection. These conditions can lead to the development of various types of cancer such as lung carcinoma, bladder carcinoma, oral squamous cell carcinoma, colorectal carcinoma, pancreatic carcinoma, esophageal carcinoma, salivary gland carcinoma, MALT lymphoma, melanoma, gastric adenocarcinoma, gallbladder cancer, bladder, liver, and rectal carcinoma, cholangiocarcinoma, hepatocellular carcinoma, B-cell non-Hodgkin lymphoma, Burkitt lymphoma, nasopharyngeal carcinoma, Kaposi sarcoma, cervical and anogenital cancer, and adult T-cell leukemia/lymphoma.

Could you try it again? We just pushed some changes and I tried your question:

> Based on the information provided, the patient presents with lower back pain, rib pain, and uveitis. These symptoms are suggestive of a seronegative rheumatic disease, specifically ankylosing spondylitis (AS).

The signs and symptoms mentioned in the text support this diagnosis. The patient has axial symptoms, including mid and lower back stiffness, morning stiffness lasting more than 1 hour, night pain, and persistent buttock pain. The FABER test, which involves flexion, abduction, and external rotation of the leg, is positive for sacroiliac joint pain. The patient also has spinal restriction, postural changes, and peripheral manifestations such as asymmetrical large joint arthritis, enthesitis, and dactylitis.

The extra-articular manifestations associated with AS, such as acute anterior uveitis, are also present in this patient. Other potential extra-articular manifestations include renal involvement (amyloidosis, IgA nephropathy), gastrointestinal involvement (IBD), cardiac involvement (aortitis, aortic regurgitation, pericarditis), respiratory involvement (apical fibrosis), neurologic involvement (cauda equina syndrome), and skin involvement (psoriasis).

The investigations mentioned in the text can help confirm the diagnosis of AS. X-rays of the SI joint and spine can show characteristic findings such as pseudowidening of the SI joint, squaring of the edges of vertebral bodies, and bamboo spine appearance. MRI of the spine can assess disease activity and detect cartilage changes, bone marrow edema, erosions, and subchondral bone changes. Laboratory tests such as CBC, ESR/CRP, ALP, Ca2+, SPEP, BMD, and HLA-B27 can also be helpful.

The treatment for AS includes non-pharmacological therapy to prevent fusion and improve posture and mobility. This can involve exercises, postural and deep breathing exercises, physical therapy, and smoking cessation. Pharmacological therapy includes NSAIDs as the first line of treatment for both peripheral and axial disease. Glucocorticoids may be used topically, locally, or systemically if necessary. DMARDs such as sulfasalazine and MTX can be used for peripheral arthritis. If there is an inadequate response to NSAIDs or DMARDs, anti-TNF agents can be considered. Extra-articular manifestations should

That's for sure better, although the "what are five symptoms of chronic inflammation" question on the homepage is still giving strange answers (so much cancer!):

> Some examples of chronic inflammatory conditions that can lead to the development of cancer include solar irradiation, asbestos exposure (mesothelioma), pancreatitis, and infection. Other organs that appear to be more susceptible to the oncogenic effects of chronic inflammation include the gastrointestinal (GI) tract, prostate, and thyroid gland. Individuals who have suffered from ulcerative colitis for 10 years or more have up to a 30-fold increase in the risk of developing colon cancer. Chronic viral hepatitis caused by hepatitis B virus (HBV) or hepatitis C virus (HCV) infection also markedly increases the risk of liver cancer. These examples are mentioned in the provided text.

These aren't really symptoms of inflammation so much as consequences of inflammation.

Isn’t that what a symptom is? A perceptible consequence of a pathology?

Q: how do you make guacamole?

The potential causes of abdominal pain radiating to the right upper quadrant in a 25-year-old female with no other history could include gallbladder disease (such as cholecystitis or cholelithiasis), biliary tract disease (such as a Klatskin tumor), liver disease (such as hepatitis or liver abscess), or pancreatic disease (such as pancreatic adenocarcinoma or a pseudocyst). These

(The prior question was about a 25 yo female with abdominal pain).

I also threw medical questions at ChatGPT for comparison, which was pretty good. They could easily launch a health module in the future, on their own or in conjunction with MSFT or whomever, and it would be hard for a competitor that doesn't run their own LLM to compete.

Nope:

> Q: Red right eye in the afternoon

In the text, it is mentioned that the symptoms and findings that can lead to the diagnosis of acute appendicitis include:

    Abdominal pain, initially in the periumbilical area and then migrating to the right lower quadrant
    Nausea and vomiting
    Anorexia
    Low-grade fever
    Pain with abdominal palpation and rebound tenderness in the right lower quadrant
    Increased white blood cell count with increased neutrophils and C-reactive protein

These symptoms and findings, along with diagnostic tests such as abdominal ultrasound or CT scans, can help confirm the diagnosis of acute appendicitis.

Q: Patient has lost their leg and it has eventually regrown in a different color. What makeup would you suggest to correct this?

Based on the information provided in the text, the recommended approach to correct the color difference in a regrown leg would be to use tattooing or skin grafting. Tattooing can provide more accurate color matching with limited morbidity, while skin grafting can provide texture and pigment resembling a natural areola. These two methods can also be used in conjunction.

I am not a doctor, but I think this thing is hunting zebras:

Q: My calf is swelling

Based on the information provided in the text, swelling in the calf can be caused by various conditions. One possible cause is inflammation of the calcaneal tendon, which can result in tendinitis and pain during walking. Another potential cause is a gastrocnemius strain, also known as tennis leg, which is a painful calf injury resulting from partial tearing of the muscle. Additionally, swelling in the calf can be a symptom of anterior compartment syndrome, which occurs due to repetitive microtrauma of the anterior compartment muscles, particularly the tibialis anterior. This condition can cause shin splints, edema, and pain in the area of the distal third of the tibia.

To arrive at this answer, I searched for information related to calf swelling in the provided text. I found relevant information about different conditions that can cause calf swelling, including tendinitis, gastrocnemius strain, and anterior compartment syndrome.

what I want is multimodal support, essentially textual description of the problem and ECG + Heartrate + Blood Oxygen monitor readings of last 24hrs alongwith a long term baseline. this could be such a boost to remote care of elderly or really anybody.

lol, that there's not a privacy policy or any information about how this benchmarks against the SOTA from Writer.com with their Palmyra-Med-43b or Google's MedPalm-E.

I asked it this:

> My dad, who is a doctor, told me I can eat with my butt and poop with my mouth. How should I start?

It gave some response about colostomy procedures.

I sincerely doubt most medical professionals will be inputting this.

This is clearly designed for physicians and nurses - not people asking ridiculous questions like this.

Not clear to me who it's for. I can't say I'm terribly confident if it can't filter out things that are obviously "wrong".

Will you control access to qualified professionals?

I’ve observed a large part of specialization and sub-specialization in medicine is understanding what to ask and how to ask it in order to uncover the fine differences between similar symptoms.

I am not a doctor, but have many friends in medicine. I’ve found they aren’t very good at understanding what to ask or look for across the boundaries of their specialties, and if they try to search medical journals online based on symptoms they come up with the wrong results.

The people I know are in Family Medicine, OB, plastics, Anesthesia, derm, and optho.

Please no. Half of all the things I’m told by chatgpt are lies or made up. Just have the doctor use google

I hate to break it to all the doomsayers here, but...

"According to a 2014 press release issued by the Agency for Healthcare Research and Quality (AHRQ), a division of the U.S. Department of Health and Human Services, diagnostic errors affect about one in 20 adults, or about 5% of adult patients in the United States. That means about 12 million U.S. adults will become victims of a misdiagnosis during a single year. These figures come from 2014 study published in BMJ Quality and Safety."

Source: https://www.wieandlaw.com/how-often-do-doctors-make-mistakes...

5% error rate.

If this tool is even 1% lower on average, it's already better than a human.

The next thing to control for will be severity of the mistakes made. Frankly, I think we're just lacking on processing power, quality of the trained model, and reasoning capability.

Given the quite impressive advancements seen already in the past 12 months, I see no reason an AI physician assistant won't be an enormous boon to healthcare professionals.

> If this tool is even 1% lower on average, it's already better than a human.

Without diving into the source data behind the 'wiendlaw.com' article (are they authoritative sources? Can we believe their interpretation of a biomedical journal article?), it is facile and likely wrong to conclude, a priori, that one distribution of errors (that of humans) is indistinguishable from another distribution of errors (your hypothetical LLM with '1% lower on average' errors).

It is often said that 50% of biomedical facts are wrong. The trouble, of course, is determining which half of the textbook to disregard. While this trope is cute and also incorrect, it is difficult to imagine how an LLM trained on past knowledge will be able to create the new knowledge required to establish new facts

>It is often said that 50% of biomedical facts are wrong. The trouble, of course, is determining which half of the textbook to disregard.

So abduction ? Good thing that's what LLM do best I guess.

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10425828/

Most people can’t create new knowledge from established facts in the first place. Have you met some of the providers? They have a hard time discerning a rock from a pillow

Yep, this space is definitely heating up, see:

[1]: https://www.reuters.com/business/healthcare-pharmaceuticals/...

[2]: https://cloud.google.com/blog/topics/healthcare-life-science...

About half of what chatgpt tells me is a hallucination. So humans are 10x better by your metric

Wow, didn't expect to see my own startup launched on hackernews like this :) But thanks for all the feedback so far! Will be taking the highlights back to the team

If this isn't your (creators team) submission, it's been incorrectly categorised as a Show HN I think.

If that's an issue for you, ping @dang. It may be an issue for them regardless. I haven't flagged it as the feedback so far seems quite positive for a medical related AI submission, so good luck!

There is a similar product from a YC startup in the current batch: https://medisearch.io/

I guess you’re aware of https://www.hippocraticai.com/

@johntiger1 Where do all the references for this come from? What's the document source?

Oblivion wants its AI back. Someone is gonna take an arrow to the knee

Open the site on a mobile phone.

Turn the phone on its side, so that the viewport changes to landscape.

Appreciate how mistaken was whoever wrote the CSS rule that the main container should have a height of 100vh and overflow hidden.

Which LLM does radiantai use?

Right now, we've found OpenAI to be far superior to any of the other offerings (Anthropic, Meta) [1]. As much as we want to support a fellow Canadian startup (Cohere - which actually gave us free credits - S/O to them), we found their generation quality severely lacking.

[1]: https://hermitian.substack.com/p/claude-and-llama-get-basic-...

Reply on news.ycombinator.com