back to article Y'know how you might look at someone and can't help but wonder if they have a genetic disorder? We've taught AI to do the same

Artificial intelligence can potentially identify someone's genetic disorders by inspecting a picture of their face, according to a paper published in Nature Medicine this week. The tech relies on the fact some genetic conditions impact not just a person’s health, mental function, and behaviour, but sometimes are accompanied …

  1. Anonymous Coward
    Anonymous Coward

    Ethics

    From the article:

    > it's tricky to get hold of actual patients' faces ethically, making datasets scarce

    That, however, will not be a problem for Facebook. Uploaded pictures with text such as "our adorable baby with NN's Syndrome" will be available in spades.

  2. malle-herbert Silver badge
    Joke

    Why not test the system on politicians ?

    A lot of them are clearly suffering from anal-cranial inversion syndrome ....

    1. Chris G Silver badge

      Re: Why not test the system on politicians ?

      Yeah! The Japanese flag in the centre of their face is often a strong clue.

      My old aunt had this tech sussed decades ago; eyes too close together shifty expression, thin lips, all indicators of deficiency. She was a racist too.

    2. LoPath

      Re: Why not test the system on politicians ?

      Convolutional Neural Network (CNN)? That's fake news!

  3. Teiwaz Silver badge

    aaaand another step to enable systemic eugenics

    Things like this increase my perception we're still on track to repeat the 20th century

    1. Voyna i Mor Silver badge

      Re: aaaand another step to enable systemic eugenics

      From TFA:

      "Both training and testing datasets are largely made up of faces from Caucasian backgrounds, and the algorithm's performance may differ for people of other races"

      Right there in TFA. "Caucasians"? "Other races"? Pure USA bullshit racism in one sentence. The USA loves its racism and doesn't want to be parted from it by mere biology and biochemistry, hence the Wikipedia article which is a mishmash of racist statements ("the neutrality of this article is disputed") and some extremely careful putting in the boot in a form that the racists won't notice is criticism.

      How often does it have to be repeated that the "Caucasian" thing is pure pseudo-scientific garbage* and there's only been one human race since either the Neanderthals or the Hobbits went extinct?

      I hope that sentence was a comment and not in the original research.

      *It has its origins in an 18th century German trying to "prove" the superiority of "caucasians" over "mongoloids" - the German idea which, ultimately, led to WW2 in which the "mongoloids" got their chance to make their case.

      1. FrankieA

        Re: aaaand another step to enable systemic eugenics

        "Both training and testing datasets are largely made up of faces from Caucasian backgrounds, and the algorithm's performance may differ for people of other races. "It is possible to extend this system to a wide range of ethnic backgrounds. We are extending our system to be able to support this functionality," Gurovich said."

        This is the only reference to race in the article I could find. It is saying that because the training and test datasets are largely made up of Caucasian faces, they don't know how it will perform on faces of people of other ethnicities. I don't see what the problem is? Are you telling me you don't see any physical differences between the face of a Chinese man and a black man?

      2. Anonymous Coward
        Anonymous Coward

        Re: aaaand another step to enable systemic eugenics

        >How often does it have to be repeated that the "Caucasian" thing is pure pseudo-scientific garbage* and there's only been one human race since either the Neanderthals or the Hobbits went extinct?

        Given that cranial shapes are used in forensic sciences to determine the broad race of the deceased, I guess you can repeat this forever. The supraorbital foramina is one such distinguishing feature used by identifications for decades.

        >*It has its origins in an 18th century German trying to "prove" the superiority of "caucasians" over "mongoloids" - the German idea which, ultimately, led to WW2 in which the "mongoloids" got their chance to make their case.

        If you really believe that led to WWII I suggest you read more history. This is outrageous pseudo analysis of something that lead to 80 million dead.

  4. Justin Case
    Devil

    This can only end badly

    Physical appearance might be the result of genetics. It also might not be. Too much of this AI malarkey seems to be devoted to inferring one thing from another thing and then representing that as the truth rather than what it is, an educated guess.

    1. Anonymous Coward
      Anonymous Coward

      Re: Yep.

      It is an indicator, but not necessarily a cause -> effect.

      That is, some genetic defects have multiple resulting deformities. One of those may be physical development.

      Other times it's not a guarantee.

      But we already have the "facebook is down, must be the phase of the moon" problem with 99% of people, so I don't hold out any hope.

    2. FrankieA

      Re: This can only end badly

      What else would cause your face to look a certain way besides genetics? I'm not talking about a scar from falling off your bike or a red nose from a cold.

  5. James 51 Silver badge

    If these conditions are the results of genetics, surely a test that looks for the genes which cause the condition would be a better diagnostic tool?

    1. Anonymous Coward
      Anonymous Coward

      Just because we know a syndrome and know what to look for, do not mean we know what genes are causing this. In a related example neural nets can analyse people's sexual preferences, supposedly with great accuracy. We still don't know what genes are at the basis of this, or even if there is any.

    2. I ain't Spartacus Gold badge

      James 51,

      Blood tests take time and money.

      When a child first presents with symptoms, parents will got to their GP. Who's not an expert in all these esoteric conditions.

      As my Mum's GP said to her when I was a few weeks old - "your son can't have [insert genetic condition], it's extremely rare, and I've never seen a case of it".

      And at least in the UK you have to get a GP referral to see the actual specialist who can diagnose that yes I do have that rare condition. And the reason my GP had never seen it before was because it was, of course, rare.

      So in this case a simple but not very reliable tool to help narrow down possibilities might be useful. As long as it's not giving out so many false positives that the specialists can't get any work done.

      The other downside of having a rare genetic condition is that there aren't many qualified doctors you can see. So the last thing I want is all my possible appointments used up by some AI causing GPs to send half their patients in for checking. To be fair, with pointers in the right direction GPs can usually look up what tests they need to do, their problem is just that they're generalists who don't see enough of the odd cases to be able to recognise them without a hint.

      For example I ended up in A&E having an X-Ray on a wrist injury a few years ago. And had to wait an extra ten minutes before finding out if it was broken so the doc could use me as a quick exam question for his students. So there's 5 doctors who should be slightly better equipped to do that diagnosis in future.

      Not that I'm convinced this AI thing will work, but it could be capable of doing just enough to help GPs a bit for very little cost - so I can see it being quite successful as an approach. Maybe. If they can get it working reasonably well and still persuade users that it's only an indication not absolute truth. The second maybe being a harder job than the first?

      1. James 51 Silver badge

        The AI is only cheaper if it works. With such a small training set no one else can see, I temain to be convinced.

        1. vir Silver badge

          It's fine; we just need to expand the training set. Once we get all 7 (8?) billion people's faces in the model, it will be 100% accurate.

          1. James 51 Silver badge

            Only if the dignosis rate among those people is 100% accurate.

  6. just another employee

    Get your Callipers out

    1930's Germany here we come.

    AI obviously makes things (now) acceptable.

    1. Voyna i Mor Silver badge

      Re: Get your Callipers out

      Why look to 1930s Germany?

      In the US, Indian Indians were banned from becoming citizens because they were not "caucasians". Then someone pointed out that according to the best racist science, Indian Indians were Caucasians. So it was then decided that they still couldn't be citizens because they were not "white".

      The rot began in Germany in the 18th century with "racial science" but was enthusiastically adopted in the US. Had it not been, perhaps US industrialists would not have been so keen on Hitler and WW2 might not even have taken place in the West.

      Hitler needed IBM and Ford.

    2. FrankieA

      Re: Get your Callipers out

      What does this have to do with 1930's Germany? Determining the viability, virility, and success of someone's genes has been something humans and ever other species of animal have been doing since life has existed on this planet. Could you imagine how long we'd have made it as a species if we couldn't tell when someone had bad genes before mating with them?

  7. Anonymous Coward
    Anonymous Coward

    Reading the learning sytem's mind...

    “DeepGestalt, like many artificial intelligence(sic) systems, cannot explicitly explain its predictions and provides no information about which facial features drove the classification,”

    Why? I keep reading this about machine learning, and I have great difficulty believing it's impossible to log what the learning system (it's not an AI) is actually doing. It might be really hard, and create a huge amount of data to trawl through, but understanding this will make the system a lot more useful.

    1. imanidiot Silver badge

      Re: Reading the learning sytem's mind...

      Not really, due to the way most machine learning (what this most likely is) works what is in between the input and the output is just a bunch of weighted tensors. You can get the weights and the methods used to combine the inputs to get the outputs, but those don't really give you any more info than the outputs themselves.

      1. Anonymous Coward
        Anonymous Coward

        Re: Reading the learning sytem's mind...

        After re-reading the article, I've noticed that someone at least *does* know what the learning system does, even if it's not the person writing the report. From the text next to the accompanying image: " First, the input image is analysed using landmarks and sectioned into different regions before the system spits out its top 10 predictions."

        Assuming that's true, it's using some form of feature recognition algorithm, then using conformal mapping to break down facial features and normalise the extracted shapes, and *then* the learning algorithm kicks in and does the job of comparing the features to weighted mean for a given genetic disorder in the known data set.

    2. Anonymous Coward
      Anonymous Coward

      Re: Reading the learning sytem's mind...

      Because it just makes a matrix of numbers, and picks a number from it.

      We don't know what that pixel/number means, we just know it is there.

      A difference between getting a slap in the face and getting a slap in the face and someone saying "because you said I was fat". ;)

      1. Robert Helpmann?? Silver badge
        Joke

        Re: Reading the learning sytem's mind...

        Perhaps we could get another machine learning system to monitor the first and learn to provide explanations as to what basis it had for its output... and then another one to validate the results from the secondary system... I suspect if we repeat this enough times, we will find the underlying algorithm amounts to "Because I said so."

  8. InNY
    Trollface

    Is this the

    "new" and "improved" version of phrenology?

    1. holmegm

      Re: Is this the

      Is this the "new" and "improved" version of phrenology?

      Er, no ... facial abnormalities are literally used in medicine to screen for and identify genetic disorders.

      The idea itself is science, not pseudo-science. Can't speak to the execution here though.

      1. InNY

        Re: Is this the

        Whoosh?

      2. gnarlwood

        Re: Is this the

        Actually, it should take AI to fully discredit phrenology once and for all. Or revive it. Just wait, somebody will have a crack at it. Remember that facial features are not uninfluenced by the bone beneath.

    2. I ain't Spartacus Gold badge
      Happy

      Re: Is this the

      No. The improved version of phrenology is when you have the "correct" bumps engineered into your head by a guy with a hammer so that you get the personality traits you want.

      [With thanks to the great Terry Pratchett. And where's our TP icon?]

      1. Teiwaz Silver badge

        Re: Is this the

        The improved version of phrenology is when you have the "correct" bumps engineered into your head by a guy with a hammer so that you get the personality traits you want.

        Ah, you mean Retrophrenology

        You can go into a shop and order an artistic temperament with a tendency to introspection and a side order of hysteria. What you actually get is hit on the head with a selection of different size mallets

        - Men at Arms by Terry Pratchett

  9. OhHelloThere

    Man, can't wait to find out if I have down syndrome!

  10. StuntMisanthrope Bronze badge

    Politico treason, pet.

    It’s a complex napoleonic number and accedes to hard labour. #zchromosome #fawkesmodal

  11. Anonymous Coward
    Anonymous Coward

    The 600 series had rubber skin. We spotted them easy

    https://i.redd.it/pk91v223t7r01.jpg

  12. find users who cut cat tail

    ...the software shouldn't really be used to make a final decision

    But it will be. That's how all other software is used.

  13. Anonymous Coward
    Anonymous Coward

    Wake me up the first time it finds a previously undiagnosed sufferer whom a clinician missed.

  14. fredj

    In general I believe we all do this already. It starts at school when you are surrounded by pears who are being marked and tested every day and you are living close together for many hours at a time. You become "trained" by constant reference to the data set performance in which you live and where you exist within it. When I was among young people I could quickly estimate their intelligence. Where I fall down is with levels of intelligence vastly more so than mine. That is outside my data set . Old people give me the same problem but that is probably because I am heading for my four score years and live in a village with a fair number of professors down to dole scroungers.

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

Biting the hand that feeds IT © 1998–2019