back to article Now boffins are teaching AI to dial up chemo doses for brain cancer

Machine-learning software has been trained to suggest the frequency and dosage of chemotherapy for patients suffering from glioblastoma, an aggressive form of brain cancer. We know what you’re thinking: this sounds like a horrifically bad idea. It gets worse when you find out that the model was trained using reinforcement …

  1. James 51

    I'd be more excited if doctors were using genetic profiling now both of the tumour(s) and the patient's healthy tissue so we could compare like with like.

    1. Anonymous Coward
      Anonymous Coward

      This is also happening. This work isn't as advanced or as well publicised for a couple of reasons:

      - Large scale genomic datasets aren't yet widely available. Projects like the 100k Genome Project should begin to change this in the next couple of years.

      - Genomes are *big*. Incorporating them into analysis is really hard so is currently the domain of huge pharmas and of the biggest institutes. It's not something you can just knock out with a couple of weeks of python wrangling, but projects like HAIL and GATK and the continually falling price of cloud infra are changing this.

  2. Pascal Monett Silver badge
    Unhappy

    I am wary of this development

    I will always be grateful for advances in medical science, and any and every cure for anything, but I do not agree with the prospect of having my cancer diagnosis made by a computer, if and when I am concerned by that.

    You can wax lyrical all you want about how a Statistical Analysis Machine (or SAM) can be efficient at diagnosing cancer and specifying treatment, I want a human doctor's professional opinion first. Eventually, the doctor can consult the SAM's opinion, if only to ensure that both are in agreement. I would accept that because the SAM might differ in opinion, and I would expect the doctor to check why and if it is relevant to my case.

    But to have some pimply-faced intern show me a printout and say "here's your diagnostic" ? No thank you.

    1. Chris G

      Re: I am wary of this development

      I assume your downvotes came from pimply faced interns, reinforcement learning is great for dogs, monkies and rats but it falls short as a training method for something that impacts someone's continued existence.

      When the learning experience of an AI equals that of a consultant oncologist it will be useful on real patients, until then, this kind of thing is useful research and not a tool to rely on.

      On the other hand, it must be good it's AI!

    2. Bilious

      Re: I am wary of this development

      Response to the drugs in question may be influenced by lots (really) of genetic and nongenetic variables in both the host and the tumour. One part is pharmacokinetics in the host - the overall fate of the drug substance after administration. Another is pharmacokinetics in the tumour - diffusion into areas and cells and organelles where the action takes place. A third is mapping of the pharmacodynamic susceptibility of each malignant clone in the tumour, and the variation of activity-related susceptibility in each cell. And how successful was the surgery?

      Details about the behaviour of old anticancer drugs will have to be mainly guesswork with very poor predictive value for an individual patient. Most of the microdetails needed are unavailable to the doctor ahd his/her computer program. And who knows what is relevant, and the weight of each variable?

      I would not invest in this project or the people behind it.

      1. sodium-light

        Re: I am wary of this development

        Finally, some cogent commentary on this nonsense. There is such a gaping chasm between medical practice and the computing world. The drugs these people have looked at are simply ancient and, in the case of PCV, probably 66% ineffective. The only way forward is for doctors to embrace the future and fashion their own computed assistance, informed by gained intuition and what actually happens to patients.

        The only newish thing here is non-uniform drug dosing and intervals but that is way off being ready for 'AI'.

  3. Fursty Ferret

    Exciting is one way to put it, we guess.

    And once again a Reg hack exposes their ignorance of a complex subject. Given that your specialist subject area is cheap journalism and stuff like managed switches, as opposed to neurosurgery and machine learning, why not just report objectively?

    1. Chris G

      So, Mr Ferret, or may I call you Fursty?

      To which complex subject are you referring? Oncology with regard to brain tumours or ML/AI?

      Both are assuredly complex but I doubt these researchers were fully trained in both.

      My feeling is that many researchers are looking for high profile or attention grabbing subjectet to apply their ML projects to rather than simply trying to make these systems work.

      Why not teach a system to diagnose faults in sewing machines and teach them to fault find all of the problems that can occur in sewing machines both industrial and domestic, it's a massive market with millions of machines worldwide, also unlikely to result in a horrible death for someone.

      Of course sewing machines are not quite the headline grabber as brain tumours. Maybe you don't appreciate the complex job of being able to write about a broad range of isuues in an engaging way that also questions a subject as much as reporting on it.

  4. Nick Kew

    Looks to me like a perfectly routine use of AI. The expensive Quack will do the high-level stuff, while the AI takes the technician role.

    Experiments with AI in such roles help determine whether it's competent, either in an absolute sense or compared to human workers. Hopefully the AI can do a good job of avoiding some traditional problems, such as mistakes of boredom.

  5. Joe W Silver badge
    FAIL

    AI and cancer research...

    wasn't there something recently? Also with artificial data?

    https://www.theregister.co.uk/2018/07/28/ai_roundup_720718/

    All of my applications for $(statistical method) work perfectly well with artificial data, it's just that when I use real world data that stuff falls apart (how... unexpected - not).[*]

    [*] yes, also for my own stuff. I know what I'm talking about...

  6. gurugeorge

    Ahem

    I almost preferred the author's name to the article. If Katayana Quach is a real name then I've got an average sized wonga. Read my other comments for drunken rants that even I dont recall. Is Quach pronounced "quack, quash kwosh kw-aysh, Quake, or queef?" Regardless this was one of those google images days. She buff though admittedly Ive been suffering from the yellow fever for over two decades, from the very first days of BBS alt-images etc.

  7. JDX Gold badge

    Is El Reg uncommon in being a technology news site which is pretty uniformly pessimistic about technology? And is that conservatism, cynicism or realism?

    1. Robert Helpmann??

      Is El Reg uncommon in being a technology news site which is pretty uniformly pessimistic about technology? And is that conservatism, cynicism or realism?

      A little of Column A. A little of Column B.

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Other stories you might like