back to article Boffins' neural network can work out from your speech whether you'll develop psychosis

Machine-learning algorithms can help psychologists predict, with 90 per cent accuracy, the onset of psychosis by analyzing a patient's conversations, according to this research here. Psychosis can be a symptom of psychiatric disorders such as schizophrenia or bipolar disorder. People experiencing psychosis find it difficult to …

  1. Anonymous Coward
    Anonymous Coward

    What could possibly go wrong

    So, based on an AI trained on 40 people, in a few years time we're likely to find people being "sectioned" by computer analysis?


    1. Khaptain Silver badge

      Re: What could possibly go wrong

      Scary doesn't even come close to describe what's going on here.

      How many previous and current world leaders would love to "use for the greater benefit of all", these kinds of algorithms.

      1. m0rt Silver badge

        Re: What could possibly go wrong

        "How many previous and current world leaders would be sectioned by these kinds of algorithms."


        1. Christoph Silver badge

          Re: What could possibly go wrong

          You mean someone who used phrases of low semantic density similar to the example of ‘Sometimes things are things’?

          Such as "Brexit means Brexit"?

          1. Rich 11 Silver badge

            Re: What could possibly go wrong

            I had a few other examples in mind, given the coherency and delusion aspects.

            "A lot of the wall was, you know, I'm good at this stuff. That's what I do. We renovated a lot."

            "If you have a windmill anywhere near your house, congratulations, your house just went down 75 percent in value. And they say the noise causes cancer."

          2. nematoad Silver badge

            Re: What could possibly go wrong

            "Such as "Brexit means Brexit"?

            Yes. This makes the quote "People experiencing psychosis find it difficult to tell what’s real or not... and are led to believe delusional thoughts." a pretty good summation of the current Tory party. The dangerous thing is one of these numptys is going to be the next PM.

            Run away!

      2. Mage Silver badge

        Re: What could possibly go wrong

        It's totally stupid.

        It's not AI

        It's not a big enough or accurate enough dataset even for the pattern matching engines mis-sold as AI.

      3. Michael Wojcik Silver badge

        Re: What could possibly go wrong

        How many previous and current world leaders would love to "use for the greater benefit of all", these kinds of algorithms.

        What use would this be to a repressive government? Autocrats aren't short of excuses for oppressing the populace, and a diagnosis of pre-psychosis says nothing about whether the subject is likely to support or oppose the status quo.

    2. Paul Kinsler

      Re: What could possibly go wrong ... "sectioned" by computer analysis?

      The published article itself (as linked by elreg) is open access, so we can read it.

      Here are two quotes from the discussion section:

      ** "This work is a proof of concept study... "

      ** "This study had a relatively small number of participants. A well-known problem in these studies is overfitting and consequently poor generalization.38 In the current study, we took steps to guard against overfitting by limiting the number of predictive variables ..."

      This article seems to be of a fairly typical type, i.e. "We tested out this idea on a small dataset, and it seemed to give promising results. Because of this, it is probably worth us doing a more comprehensive study in the future; but we are telling you all now because you might find it interesting".

      I think any appendix, potentially stating something along the lines of "Aieee! Run for the hills, they're going to get you!" would probably have been deleted on the advice of the referees as being not in the correct scientific style. :-)

      1. Teiwaz Silver badge

        Re: What could possibly go wrong ... "sectioned" by computer analysis?

        Here are two quotes from the discussion section:

        ** "This work is a proof of concept study... "

        All very cogent, but then then we have a Police service trialling an unproven system of equally dodgy provenance - which probably had exactly this disclaimer but since it said 'proof' they thought they had it 'bang to rights' and assumed they could convict on it.

        1. Michael Wojcik Silver badge

          Re: What could possibly go wrong ... "sectioned" by computer analysis?

          A repressive authority does not need genuine scientific evidence to back up its policies. That's amply demonstrated by history.

      2. Mage Silver badge

        Re: This article seems to be of a fairly typical type

        i.e. We got interesting results.We need a bigger study (Really means: "Now fund our wages for the next five years").

    3. holmegm

      Re: What could possibly go wrong

      No, that isn't at all likely.

      There aren't even enough beds in inpatient units for those who clearly need them. And people are discharged way too early.

      And in the US, at any rate, you need two psychiatrists and a judge to sign off on an involuntary admission. (At least in every state I've ever heard of.)

      So no, that isn't at all likely. At most this *might* be a tool, to help identify if someone needs further evaluation.

    4. Michael Wojcik Silver badge

      Re: What could possibly go wrong

      How reassuring that two dozen Reg commentators believe that we should avoid primary research on the grounds that it might someday be used for undesirable purposes.

      The anti-intellectualism certain portions of the Reg readership never fails to impress.

  2. Pascal Monett Silver badge

    When did the study take place ?

    I can't find a date for when they started this study. If they had time to follow through and conclude that some got psychosis and some didn't, then it means that the study had to be done at least ten years ago. AI was not so prevalent a decade ago, so how did they use AI to bolster their study ?

    1. Anonymous Coward
      Anonymous Coward

      Re: When did the study take place ?

      Based on old data sets? Can be dangerous cherry picking a model that "fits" the data set. You get geocentric theory that way. Yes, you can get a correct model, but you can also get an incorrect one that only fits one data set.

    2. Paul Kinsler

      Re: When did the study take place ?

      The article states a two-year follow up period - presumably ten years would be a bit long to wait for a basic proof-of-concept study like this one.

    3. MrMerrymaker Bronze badge

      Re: When did the study take place ?

      Maybe one of them developed multiple personality disorder - with one personality being perfectly fine so it counts as half, or ... oh I don't know

      It's a scam, basically

    4. Mike Moyle Silver badge

      Re: When did the study take place ?

      The way that I read it was that they used filmed/taped interviews from however many years ago -- for which they already knew who had developed psychoses and who hadn't. Thirty of the subjects were used for training. They used the last ten to test whether the system would correctly identify which of them had developed psychoses and which hadn't. It apparently was correct in nine out of the ten cases. So, it wasn't a case of "run the software then wait ten years to see how it did"; they could check the system's conclusions against data that they already had.

      At least, that's how I read it.

  3. BebopWeBop Silver badge


    Someone has a sense of humour, but does this training set mean that there will be a bias against recognition of certain types/degrees of psychosis?

    1. stiine Bronze badge

      Re: Reddit

      Only the made-up kinds.

      Now if they really wanted to study psychoses, they should have monitored 4chan for, oh, just a few minutes.

      1. heyrick Silver badge

        Re: Reddit

        Or the current pretenders to be Prime Minister?

        1. amanfromMars 1 Silver badge

          Re: Real Live AIBetaTesting of Readily Available Future Facilities/Abilities/Utilities

          Or the current pretenders to be Prime Minister?.... heyrick

          One surely hopes for a blending of genius rather than a surfeit of psychosis be rooted and anchoring to foundational ground zero bases for any SMARTR Governance Machine, heyrick.

          In these rapidly changing days with exponentially expanding ways to share formerly held captive and secret and sensitive information about that which is Vitally Important and EMPowering, is there a leader amongst them brave and bold enough to recognise and embrace help whenever it is shared, and of which they would/should be made fully aware of. :-) Easily achieved with simple sight of Real Live AIBetaTesting Narrative.

          How to Deal with the Future, is not a trick question. IT expects Answers ..... for IT is a Clear AI and Presenting NEUKlearer HyperRadioProACTive Danger to Fake News with Corrupted Mass Media Systems with an almost Universal Command and Control of Virtual Future Realisations with Sublime Autonomous Services for Servering to Almighty Futured Seeds and Feeds and Needs for such is surely Stealth InterNetWorking.

          :-) Surely you weren't thinking Bletchley Park type Boffinry was just a Jolly Roger of an Experiment and/or a One Off Experience ..... whenever Master Moulds were Crafted there for Global Export/Import.

          Casting Excellent Minds over Deep and Dark Matters is de Rigeur Default in Such, a Vast Advanced Proprietary Intellectual Properties Field. Another not a trick question is.... Is it a human confection or alien treat?

          Methinks that sort of puts Brexit in something of a Harry Lime Light? :-)

          So what's it to be? A Genius Blend or an Over Sufficiency of Psychosis for here lie Abiding Fields of Battle. Trick or Treat? Honest Truth or False Promise? What leads to where you have been and/or where you be thinking you might like to go to?

        2. quxinot Silver badge

          Re: Reddit

          >Or the current pretenders to be Prime Minister?

          Congress would be a much better data set. Also, Parliment.

          Remember, you need the biggest n value you can for these things.

  4. Chris G Silver badge

    We need to publish something

    What can we do to get an interesting paper out easily?

    That seems to be what this paper is about; a small sample of people overly weighted in one direction, no passable control sample and a questionable source of 'normal conversation' that isn't even conversation.

    While the aims of the research are laudable the methods are laughable.

    Well, that what the voices tell me.

    1. el kabong Silver badge

      Trumpification of science

      Just keep the fog machine rolling along, do whatever it takes to keep it rolling, keep publishing, always, everywhere, use twitter.

    2. b0llchit

      Re: We need to publish something

      Easy, just use It will make you the paper you always dreamed about.

  5. el kabong Silver badge

    So called "boffins" take the easy route...

    produce a few click bait articles, take more funding, repeat.

    These people give boffinry a bad name.

    1. 96percentchimp

      Re: So called "boffins" take the easy route...

      Imagine you're a boffin with a hypothesis. You can obtain enough funding for a pilot study with a small sample running over a short period of time.

      Do you (a) abandon the hypothesis because your study isn't perfect, or (b) perform the pilot study and hope it generates enough interest and enhances your seniority in the community enough to win greater funding to take your research further?

      Don't blame the boffins for the system they have to work in.

  6. Anonymous Coward
    Anonymous Coward

    " Speech from 30 of the participants was used to teach the software what potential psychosis sufferers talk like"

    This could be HUUUUGE!

    1. Benson's Cycle

      Sample size depends on what is being looked for.

      It may or may not. An awful lot depends on the level of correlation.

      For instance, suppose you do a long term study because you suspect that a certain cell type in the liver is pre-cancerous. You already know the prevalence of liver cancer in the general population. You find 30 people with these cells, you prepare specimens and keep in touch. After 10 years, suppose 10 of them have developed the same kind of liver cancer and have no exposure to hepatitis, known carcinogens or high levels of drinking. That's a pretty solid result.

      On the other hand, you announce that an in-depth study of eating habits shows that eating more than 70g of processed meat a day causes a 20% elevated risk of bowel cancer. That sounds pretty convincing...but it's likely that intake was self reported, and perhaps there wasn't much analysis around exactly what kind of red meat and processed food was being eaten and what other lifestyle factors might be involved. So it might be a useful statistical guide to populations, but not much use for individual risk assessment.

      Small samples were discredited in many people's minds by Wakefield and MMR, but the point about his methodology was that, as in Alice in Wonderland, it was verdict first trial afterwards - identify a group of autistic children and find out if they had had the MMR vaccine. It showed a truly shocking lack of understanding of scientific method and he was duly struck off.

      1. Michael Wojcik Silver badge

        Re: Sample size depends on what is being looked for.

        It's almost like the people who come here to make the same sophomoric complaints about every science article in the Reg have no understanding of research methodology, and are just airing their uninformed opinions.

  7. Anonymous Coward
    Anonymous Coward

    I'm missing something

    Of the 10 'test' subjects half went on to develop psychosis (so that's 5) but they claimed that the software was 90% accurate in detecting who would get pyschosis.

    How do you get 90% of 5? Did one get better? If it predicted 4 instead of 5 then that's 80% accurate. If it predicted 6 but only 5 actually became pyschotic then that's 83% surely?

    1. Anonymous Coward
      Anonymous Coward

      Re: I'm missing something

      There were 10 test subjects. Apparently it did pretty good at predicting which five didn't go on to develop psychosis as well.

    2. el kabong Silver badge

      You're missing something? Don't fret!

      It's not worth your time, you're doing OK, you really don't need to understand that garbage, no one needs, there's nothing in there that needs understanding. REALLY!

      Why start your day wasting time in an effort to take that garbage seriously? There are more productive ways to start your day.

  8. Anonymous Coward
    Anonymous Coward


    It can work out if someone might develop psychosis. Interesting.

    As long as it doesn't work out where you've buried the lusers and manglers bodies you're still good though right?

    1. quxinot Silver badge

      Re: So...

      Is it really symptomatic of a psychosis if it's necessary to retain your sanity and well-being?

  9. Amentheist


    English language, and what if the person is of such background where they'd naturally have a smaller vocabulary for example.

    1. Swarthy Silver badge

      Re: Language?

      I think it is less about the vocabulary and more about the density of meaning.

      An expanded vocabulary empowers you to leverage synergies and blue-sky to obfuscate your dialog, putting lipstick on the porcine and metaforming the narrative. Or you can speak clearly and get your message across.

      1. heyrick Silver badge

        Re: Language?

        "Or you can speak clearly and get your message across."

        Brexit means Brexit.

        Very short. Very concise. Very devoid of meaning.

        1. Benson's Cycle

          Re: Language?

          And she was allowed to get away with it. I'm not sure what conclusions to draw from that.

          1. heyrick Silver badge

            Re: Language?

            Obvious conclusion: both sides of government are not fit for purpose.

        2. YARR

          * is *

          A tautological identity statement as a riposte indicates that a prior statement about the same subject is not entirely true.

          1. YARR

            OK, so I wasn't entirely right there.

            A tautological identity statement indicates a tautological identity statement.

            ... and Brexit is the new Remain.

            1. Anonymous Coward
              Anonymous Coward

              Brexit is Brexit!

      2. Alister Silver badge

        Re: Language?

        So... Anyone who's regular vocabulary includes a high preponderance of buzzword bingo and marketing speak is likely to be psychotic?

        Hmm... <glances across office at the sales team./>

        1. Anonymous Coward
          Anonymous Coward

          Re: Hmm... glances across office at the sales team.

          Well I am looking at HR as a fountain of pre-psychotic babbledegook.

          Checks 'post anonymously' ticked

          1. Trev 2

            Re: Hmm... glances across office at the sales team.

            Yes, but "they" still know who you are and might tell HR so you really need to get rid of HR...then kidnap Lassie before she tells the police you lured all the HR people to an abandoned mine shaft.

        2. Trev 2

          Re: Language?

          There have been plenty of studies suggesting upper management types are generally delusional and psychotic, so it might be them you need to be worried about not the sales team.

      3. Benson's Cycle

        Re: Language?

        I really think we should run that one up the flagpole and see if anybody sings "Hail to the Chief."

        (I'm starting to wonder if those lovely people I used to work with were just trying to give themselves thinking time or were pre-psychotic.)

  10. Eddy Ito Silver badge

    ... by analyzing a patient's conversations...

    Conversations with whom? I would think people change how they speak depending on the other participants, assuming there are any. For instance I find it helps to dumb it down when talking to political types but not my peers. Likewise, I would expect one to be more guarded, evasive perhaps, when talking to police than with a drinking buddy although semantic density may be low in both cases but for different reasons.

    1. Michael Wojcik Silver badge

      Conversations with whom?

      As the article states, conversational data from NAPLS was compared to data from Reddit. For the latter, obviously, the interlocutors are other Reddit participants.1 For NAPLS, there's a link to the NAPLS site in the article, but I'll save you five minutes of reading and refer you to the NAPLS 2 symptoms paper, which shows that the conversational data available in NAPLS comes from researchers conducting clinical interviews per DSM-IV SCID.

      NAPLS is a longitudinal study, so there's reason to presuppose its data is fairly varied and represents a decent distribution across the subject population.

      While it's certainly true that the speech occasion and community have a strong effect on sociolinguistic and pragmatic aspects of speech, the strong correlation demonstrated by this - very preliminary! - study suggest that an SCID-type speech situation does not greatly distort the features that the model is correctly identifying as psychosis prodromes.

      1Unless the subject is fully psychotic and having delusional conversations, as we sometimes observe here on the Reg forums.

  11. Michael H.F. Wilkinson Silver badge

    Maybe feed the AI some Vogon Poetry

    with interesting rhythmic devices that counterpoint the surrealism of the underlying metaphor

    I'd better get me coat. Doffs hat (grey Tilley today) to the late, great Douglas Adams

  12. MrMerrymaker Bronze badge

    I wish I had contributed to the dataset

    I am strong and stable! I believe in blue sky thinking! No deal is better than a bad deal!

  13. trindflo

    Identifying potential job candidates?

    I can think of some jobs where these algorithms would be detecting desirable traits. Some of these groups even have access to huge datasets for analysis.

    1. Yet Another Anonymous coward Silver badge

      Re: Identifying potential job candidates?

      You don't have to be psychotic to work here - but it helps pass the interview screening ?

  14. Anonymous Coward
    Anonymous Coward

    I wonder what they would make of amanfromMars 1?

    1. Kevin McMurtrie Silver badge

      It's about prediction so the test would be running a bit late.

  15. LeahroyNake Bronze badge

    low semantic density

    That's when a customer phones up with a fault to report and they haven't got a clue how the thing works... Only their idea of how it should work as explained to them by sales.

    High semantic density, engineers or specialists discussing the problem.

    Low semantic density, management's understanding of the issue.

    Makes me want to go postal most weeks.

  16. John Savard Silver badge

    False Negatives

    By using Reddit for their 'normal' database, they're doing the right thing! Better a false negative than a false positive.

    If they had taken their 'normal' database from a more reliable source of sane people, then it might have detected upcoming psychosis in people in between the two samples that actually were in no danger.

  17. ocflyfish


    The voices in my head say this is bogus, and to stop reading the article.

  18. Anonymous Coward
    Anonymous Coward


    I can't wait until every single word the passes over the internet that has been captured in unrestricted dragnets is analysed by AIs to generate a personality profile for every single endpoint and given scores for dissent and likelihood of subversion! That will be such a wonderful future to live in :D

    1. Hans 1 Silver badge

      Re: Wow!

      They already have that in RPC. No joke icon needed, sadly ....

    2. Michael Wojcik Silver badge

      Re: Wow!

      Plenty of researchers, academic and commercial, are routinely running vast corpora of linguistic data gathered from public online sources through various models. In some academic settings, the use of such data is restricted in certain contexts (via human subjects research rules enforced by IRBs and similar institutional controls). In corporations, it's a free-for-all.

      You don't really need to wait. The difference between "most online language use is analyzed" and "all of it is" is minimal.

  19. Hans 1 Silver badge

    I will develop psychosis if they start bothering me with their mic!

  20. jtaylor

    Seems plausible

    This is interesting stuff.

    People with schizophrenia tend to make "loose associations," relating things which the average person would not connect. They see cause and effect where others see disparate events. They also tend to take events personally. As a real-life example, "I saw an ad for diabetes treatment" becomes "I have diabetes. They told me by showing me a TV ad about it."

    It sounds like this research is looking for behavior patterns that might indicate loose associations.

    This does not at all sound like it could lead to a diagnostic test (we already have diagnostic criteria), but more something that "could indicate elevated risk." This might enable families to become more educated about mental illnesses and begin to form support networks before there's a crisis. Given that symptoms often become problematic during puberty, there's the potential to help people through critical developmental stages, with huge improvements to quality of life.

    1. Michael Wojcik Silver badge

      Re: Seems plausible

      This does not at all sound like it could lead to a diagnostic test (we already have diagnostic criteria), but more something that "could indicate elevated risk."

      That's the whole point of NAPLS. They're looking for psychosis prodromes - characteristics that indicate higher-than-normal risk. NAPLS research publications include investigations into endocrinological and brain-imaging (cortical thickness) markers. I haven't looked at the results.

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