Skip to main content

Alibaba Claims AI Can Identify Coronavirus Patients With 96% Accuracy

(Image credit: Shutterstock)

Alibaba developed an artificial intelligence (AI) capable of detecting new Coronavirus infections with 96% accuracy, Nikkei Asian Review reported, and the diagnostics system based on this AI is expected to be rolled out to more than 100 hospitals.

The AI is said to be able to differentiate between COVID-19 cases and viral pneumonia with 96% accuracy when analyzing computerized tomography (CT) scans. It's much faster than humans, too, because it can make a diagnosis in 20 seconds.

Researchers trained the AI on more than 5,000 confirmed cases of novel Coronavirus 2019 (COVID-19). It also uses "the latest treatment guidelines and recently published research," Nikkei Asian Review said, in its efforts to diagnose COVID-19 infections.

The hope is that being able to diagnose COVID-19 cases faster will allow hospitals in China to deal with the influx of people who might be infected with the virus. That could in turn speed up treatment times and, potentially, help contain COVID-19.

That would be good news for the tech industry. Coronavirus has been blamed for various tech show cancellations, potential component shortages (as well as resulting PC supply issues) and weakened demand from Chinese consumers.

The AI was technically developed at the Academy for Discovery, Adventure, Momentum and Outlook (DAMO Academy) that Alibaba established in 2017 with an initial commitment of $15 billion to be spent over the following three years.

  • derekullo
    But can Alibaba see why kids love the taste of Cinnamon Toast Crunch?

    https://knowyourmeme.com/memes/why-do-kids-love-the-taste-of-cinnamon-toast-crunch
    Reply
  • Rdslw
    This metric is MATRIX BASED. true positive , false positive , false negative true negative.
    Are the 96% true positive then where goes the last 4 % ? If it lets go of infected people, its not good.
    Also another thing I've seen such companies to ignore false positives in such metrics.
    And it uses CT so throughput is to low for china...
    Reply
  • escksu
    96% is just not good enough. At 10,000 thats 400 pple..... if 100 of then is false negative, its gg.....
    Reply
  • Samat
    Rdslw said:
    This metric is MATRIX BASED. true positive , false positive , false negative true negative.
    Are the 96% true positive then where goes the last 4 % ? If it lets go of infected people, its not good.
    Also another thing I've seen such companies to ignore false positives in such metrics.
    And it uses CT so throughput is to low for china...

    Current testing methology is only 40% accurate with 30% chance for false positive or negative. So 96% accuracy would be a huge inprovement.
    But definitely some accurate method, without having to take a CT scan, is needed.
    Reply
  • Chung Leong
    escksu said:
    96% is just not good enough. At 10,000 thats 400 pple..... if 100 of then is false negative, its gg.....

    It's good enough as long as one undetected patient does not transmit the decease to more than 25 others. Reducing the transmission rate to below 1 is the key to halting the epidemic. Of course, from the point of view of individual patients, 96% accuracy is decidedly unsatisfactory.
    Reply
  • spiketheaardvark
    I'm curious how many of these patients are routinely getting CT scans. CT scan time isn't cheap and in the middle of an out break the machines are going to swamped. I also doubt that it will be useful for asymptomatic people. I can see the usefulness of running the analysis on people that are getting CT scans as standard of care but not much else. And then with more standard testing for follow up.
    Samat said:
    Current testing methology is only 40% accurate with 30% chance for false positive or negative. So 96% accuracy would be a huge inprovement.
    But definitely some accurate method, without having to take a CT scan, is needed.
    I really hope the RT-pcr test is better than that. I don't have to live up medical diagnose standards in a research lab, but if I had a pcr with that detection rate I'd consider it very broken. I would be very surprised if the CDC pushed out a test that poor. Where you take the sample from and how you handle it though are going to be very important, and it's possible the application of a good test is wanting. Our best bet probably a combination of PCR and serological test. Both are fast, relatively cheap, routinely done in clinical labs, and easy to do in very large numbers.
    Reply
  • bit_user
    Rdslw said:
    Are the 96% true positive then where goes the last 4 % ? If it lets go of infected people, its not good.
    They can usually bias the results one way or another. As you point out, you probably want to minimize false negatives, depending on the availability of quarantine facilities.

    Edit: I didn't find the research cited by the article, but here's a similar attempt by another group:

    https://arxiv.org/abs/2002.09334
    Rdslw said:
    Also another thing I've seen such companies to ignore false positives in such metrics.
    Those can be addressed by follow-up testing with another methodology.

    Rdslw said:
    And it uses CT so throughput is to low for china...
    Maybe, but maybe not. China is throwing a lot of resources at this, and it sounds like the test is only for those with significant symptoms.

    Edit: It seems the patients should've already been getting CT scans. From the article Tom's cited:
    The Chinese National Health Commission on Feb. 5 widened the criteria for the diagnosis of new infections, adding CT scan results on top of the previous nucleic acid test method, to ensure patients with clinical symptoms will receive standard treatment as soon as possible.
    Reply
  • bit_user
    Samat said:
    Current testing methology is only 40% accurate with 30% chance for false positive or negative.
    On the news, I heard someone (presumably in quarantine) tested negative 6 times and they finally got a true positive, on the 7th.

    There are also reports of a couple people testing negative, being released, and then testing positive again. It's unlikely they got re-infected, so that points to more accuracy problems with the test (or how it was administered).
    Reply
  • bit_user
    spiketheaardvark said:
    I also doubt that it will be useful for asymptomatic people.
    Maybe not for completely asymptomatic subjects, but here's an attempt at a cheap, unobtrusive method:

    https://arxiv.org/abs/2002.05534
    Reply
  • Chung Leong
    To add to my previous comment, 96% accuracy is not as good as it sounds when what you're trying to detect is relatively rare. Imagine we have 10000 patients: 100 are infected with the coronavirus, 9900 have ordinary pneumonia. The test would yield 96 true positives and 380 false positives. So only one in five diagnoses is correct. From the patients' point of view that's pretty bad.
    Reply