• Visitors can check out the Forum FAQ by clicking this link. You have to register before you can post: click the REGISTER link above to proceed. To start viewing messages, select the forum that you want to visit from the selection below. View our Forum Privacy Policy.
  • Want to receive the latest contracting news and advice straight to your inbox? Sign up to the ContractorUK newsletter here. Every sign up will also be entered into a draw to WIN £100 Amazon vouchers!

NHS - best in the world

Collapse
X
  •  
  • Filter
  • Time
  • Show
Clear All
new posts

    #31
    Originally posted by vetran View Post
    So three separate problems.

    1. AI will tell you earlier with greater certainty and with probably less time elapsed & cost that you are very likely to have lung cancer with fewer false positives so fewer biopsies are needed.

    2. There are not enough staff to perform biopsies which are starting to be replaced with less invasive tests and better surgical techniques mean it should be possible with less cost & time

    Lung Biopsy | Johns Hopkins Medicine

    3. Testing the result of the biopsy is currently done by sending the tissue to a lab and highly trained specialist counting and taking pictures of abnormal cells but of course AI can recognise pictures and count them faster and more consistently than any human.

    two of the three could really benefit from AI and save money for the middle one.
    1. "Fewer bio0psies needed. Not exactly, no oncologist at present will rely solely on an AI.

    2. Yes in theory, not really in practice.

    3. Again, no one will rely on a machine.

    Just like many things could already be done automatically, like flying a plane, eventually a human expert (or usually in complex cases a panel of experts) has the final say, especially for life and death situations.
    So your original assertion that AI could replace experts, while strictly true, will not happen in practice.
    And AI is far from being infallible, and never will be, which is why those comittees of experts still decide treatment for complex cases.

    in case you're gnuinely interested, "reinforcement learning" in the hottest thing in medical AI at the moment, it tries to make decisions under conditions of uncertainty (like AlphaGo).
    Fascinating stuff.
    Hard Brexit now!
    #prayfornodeal

    Comment


      #32
      Originally posted by vetran View Post
      Y and you accuse me of not being sharp.

      5-10% improvement with reduced or the same costs would be considered massive by any normal organisation. Key point here is that detection is earlier and with AI (and it seems with oncologists) the more information you give it the better it gets. If the cost falls you can automate more cheaper and do more tests. As detection gets better and you start doing more scans earlier then false negatives start to disappear you do less tests.
      FFS. You never do less tests. that's just the way it is.
      So lets say you had sudden piercing headaches.
      Classic red flag for a brain tumour.
      An MRI scan is taken and passed through a neural network with 95% accuracy (95% is about state of the art nowadays) and it says you probably don't have brain cancer.
      You'd be happy to forego further tests?
      Hard Brexit now!
      #prayfornodeal

      Comment


        #33
        Originally posted by sasguru View Post
        1. "Fewer bio0psies needed. Not exactly, no oncologist at present will rely solely on an AI.

        2. Yes in theory, not really in practice.

        3. Again, no one will rely on a machine.

        Just like many things could already be done automatically, like flying a plane, eventually a human expert (or usually in complex cases a panel of experts) has the final say, especially for life and death situations.
        So your original assertion that AI could replace experts, while strictly true, will not happen in practice.
        And AI is far from being infallible, and never will be, which is why those comittees of experts still decide treatment for complex cases.

        in case you're gnuinely interested, "reinforcement learning" in the hottest thing in medical AI at the moment, it tries to make decisions under conditions of uncertainty (like AlphaGo).
        Fascinating stuff.

        Ah its Turkeys voting for Christmas you are worried about so Oncologists are the modern tube drivers?

        If the machine is proven again & again to be faster & better then shouldn't they start believing it?

        Years ago I was taught metalwork, the minute I saw a CNC machine it was obvious such work would be taken over but the engineers I worked with assured me I was wrong. Now CNC tools perform stuff that no man could achieve at a speed that would be considered impossible before and at a cost considerably cheaper. However you still need someone who understands the material & process to design the cuts but not to make thousands of items.

        Autopilot is used routinely leaving the Pilots fresh to do the difficult stuff. Though recent Boeing failures highlight just how hard you need to work to make it pilot & failure proof.

        Now gathering & presenting data faster, better & cheaper is something computers can do, if we present the data and opinions well the decision is easier, more consistent and normally better because you can have the expertise of the top 100 oncologists applied to every biopsy.

        The first large report I wrote 25 years ago needed experts to interpret it, now reports can have expert analysis built in that highlights the things experts feel are important whilst including the underlying data. They can also be shared internationally anonymised if need be to get opinions & inform other experts.

        The other things machines can do is correlate billions of facts and find patterns, if we scan, analyse and present them well a human can look at that and go ooh that is interesting maybe we will try that.

        That is why we need to push from the government to develop & install the DLR and then let the unions moan.

        Thanks I will have a look.
        Always forgive your enemies; nothing annoys them so much.

        Comment


          #34
          Originally posted by sasguru View Post
          FFS. You never do less tests. that's just the way it is.
          So lets say you had sudden piercing headaches.
          Classic red flag for a brain tumour.
          An MRI scan is taken and passed through a neural network with 95% accuracy (95% is about state of the art nowadays) and it says you probably don't have brain cancer.
          You'd be happy to forego further tests?
          What happens if an oncologist reviews the scan and decides there is no evidence of a tumour in the scan?( Is that 100%? Or is it more like 80-90%? I am assuming the quality of medical experts varies considerably.

          The point they made is AI is 5% better at spotting positives and 11% better at ignoring false positives and that is without decades of training by the best oncologists in the world or the millions of scans taken annually.)

          No they probably say lets wait & see.
          Next scan is also negative.
          wait & see
          Third scan - yep we see it.

          If you run AI in parallel and on scan 1 it suggests I see something suspicious here and blows up the image with a big red circle will the expert change their mind? When that has happened in a few thousand cases they will trust it more. When the biopsies tie up with the AI findings then yes they will do fewer biopsies & tests because they have a gold standard.
          Always forgive your enemies; nothing annoys them so much.

          Comment


            #35
            Originally posted by vetran View Post
            because you can have the expertise of the top 100 oncologists applied to every biopsy.

            .
            That's not how it works. We aren't using Knowledge-based expert systems, that was abandoned in the 1990s.
            Typically we use convolutional neural networks (these are the best performing models for medical imaging, I've built one or two to accuracy of the 90s percent) to "learn" the training images then let them loose on the test data.
            In real terms you wouldn't be able to explain exactly why one image was selected as cancerous and the other wasn't.
            Neural networks are black boxes, you can explain them in general principles but you can't say why they came to a certain decision.
            No one in the world can, that's their big downfall in medical diagnosis.
            So no you still need oncologists.

            Anyway you sound like an expert. The going rate is 1000-1500 Euros per day (on the continent at least). Fill yer boots. I am. Kerching
            Last edited by sasguru; 12 September 2019, 15:25.
            Hard Brexit now!
            #prayfornodeal

            Comment


              #36
              Originally posted by sasguru View Post
              That's not how it works. We aren't using Knowledge-based expert systems, that was abandoned in the 1990s.
              Typically we use convolutional neural networks (these are the best performing models for medical imaging, I've built one or two to accuracy of the 90s percent) to "learn" the training images then let them loose on the test data.
              In real terms you wouldn't be able to explain exactly why one image was selected as cancerous and the other wasn't.
              Neural networks are black boxes, you can explain them in general principles but you can't say why they came to a certain decision.
              No one in the world can, that's their big downfall in medical diagnosis.
              So no you still need oncologists.

              Anyway you sound like an expert. The going rate is 1000-1500 Euros per day (on the continent at least). Fill yer boots. I am. Kerching
              I thought they were now reverse engineering the black boxes:

              https://www.sciencedirect.com/scienc...13158219303535

              It would seem its the thing you want on medical imaging, the where & what. So its definitely something we should be pursuing.

              We still need pilots but we reduce risk by letting computers doing the boring bits so they can concentrate on the difficult bits. Why not do the same in medicine?


              Not an expert and despite your rudeness I do learn stuff when talking to you. Hopefully some of the things I say make a sort of sense as well.
              Always forgive your enemies; nothing annoys them so much.

              Comment


                #37
                Originally posted by vetran View Post
                I thought they were now reverse engineering the black boxes:

                https://www.sciencedirect.com/scienc...13158219303535

                It would seem its the thing you want on medical imaging, the where & what. So its definitely something we should be pursuing.

                We still need pilots but we reduce risk by letting computers doing the boring bits so they can concentrate on the difficult bits. Why not do the same in medicine?


                Not an expert and despite your rudeness I do learn stuff when talking to you. Hopefully some of the things I say make a sort of sense as well.
                Indeed reverse engineering CNNs is an ongoing research activity.
                Not something I'm interested in though.
                Currently working on a Reinforcement Learning model to treat sepsis. Struggling would be more like it but Ill get there in the end.
                We do use AI in medicine, its going to be important.
                But IMO the AI hype is unjustified. Machine Learning is actually pretty simple stuff, underneath, computational statistics in essence.
                It's just pattern matching.
                Deep Learning for example will never replace a computer programmer (according to Cholet, the guy who wrote the Python Deep learning library, so he should know).
                I'm also working on genetic algorithms to see if thats true or not.
                Hard Brexit now!
                #prayfornodeal

                Comment


                  #38
                  But getting back to the theme of this thread. The NHs isn't investing as much as it should in AI screening systems for cancer because (and I heard this from salesperson in clientco) because there aint the time and cash for mass screening. Imagine giving half the population an MRI. The AI software to process those scans is the cheap part of the equation.
                  In the US of course you can pay for it personally or pick a premium plan that does so, as you can if you have deep pockets here.
                  Hard Brexit now!
                  #prayfornodeal

                  Comment


                    #39
                    Originally posted by sasguru View Post
                    In the US of course you can pay for it personally or pick a premium plan that does so, as you can if you have deep pockets here.
                    Private MRI Scan | How Much Does An MRI Cost? | Vista Health

                    £199

                    Even the great unwashed spend that on a couple of PlayStation games and some Dominoes pizzas at the weekend.

                    Comment


                      #40
                      Originally posted by DimPrawn View Post
                      Private MRI Scan | How Much Does An MRI Cost? | Vista Health

                      £199

                      Even the great unwashed spend that on a couple of PlayStation games and some Dominoes pizzas at the weekend.
                      Really, is it that cheap? I've had one on private a few years ago, assumed it was about 1K. Maybe it was then and prices have fallen.
                      Still for the NHS (199 x say, 35 million) is too much.
                      Hard Brexit now!
                      #prayfornodeal

                      Comment

                      Working...
                      X