Can someone become a researcher?

MIT researcher Brian Subirana: How AI can help fight corona

WirtschaftsWoche: Mr. Subirana, you have developed an artificial intelligence that recognizes Covid-19 infected people by the sound of coughing. How does this work?
Mr. Subirana: The algorithm essentially uses four biomarkers. Each individual was trained independently and with a different database. Then we collected cough recordings from Covid-infected people on the one hand, and healthy test subjects on the other. With the properties of the biomarkers, we have trained a network that can identify whether someone has corona or not.

So could AI be a substitute for today's tests?
We think it could be more of a complement of some sort. The nice thing about our test is that it can be used anytime, free of charge and non-invasively. You no longer need medical intervention. Our tool can easily be made available on an app server so that everyone worldwide is able to download and use it independently. We also do longitudinal analyzes. These are many advantages compared to regular PCR tests. We will only see whether the tool could also be used as a substitute for this once we have actually implemented it in a real context.

What disadvantages and challenges do you see for your technology?
One of the greatest challenges is political regulation. Politicians would need to quickly approve the use of the tool so that we can work together on that direction. If you look at the field of artificial intelligence in general, it quickly becomes apparent that, despite much advances in research, it is not really widespread to work with. This is largely due to a lack of regulation and a lack of business models. I see that as the biggest hurdle.

Face recognition has become an indispensable part of everyday life. However, wearing mouth and nose protection irritates the systems. For the time being.

Are you already talking to governments about the use of Corona AI?
We are currently debating this with four countries and we will have to wait and see what will happen next. But I think we haven't made the countries public yet. But I can tell you that so far we have actively collected samples of coughing noises in Italy, Spain, the USA and Mexico.

For which areas of application could the tool be used now?
The technology is ready for the market today. We only have the problem of regulation. In general, I see a lot of different areas of application. Two big things are particularly easy to identify: On the one hand, it can serve as a pre-screening to better manage the administration of corona tests and also to prioritize tests. On the other hand, it can of course be used for group activities and other leisure activities, for example before entering a restaurant, going to work or getting on a plane.

An app for everyone could change life in the pandemic from the ground up. How long will it be before it happens?
The idea now is to offer an open source tool with the right political approvals. This can be integrated into corona warning apps, for example. As mentioned, the technology is mature and ready to use. We could also meet other requirements, especially with regard to the language of the AI. We found out that people cough in their mother tongue, so Germans may cough differently. Then we would have to adjust the algorithm accordingly, but that can be done quickly.

Is that a possible “algorithmic bias”?
Good question, we don't know if there might be a bias for culture. This is still an open topic for future research. There could also be a bias for certain symptoms in symptomatic corona patients. But we're talking about smaller things that could be corrected in the algorithm. This is really nothing that would justify politicians not implementing the tool and making it available for use.

The accuracy of the test is 98.5 percent. How reliable are these numbers just to fight the pandemic?
It is definitely reliable enough for pre-screening. You don't need more. However, with a larger database, we could develop additional functions for the app. That is why we also accompany the complete evolution of the patient by carrying out the same study at several points in time. But this tool alone cannot stop the pandemic, any more than any other test can stop it. People would also have to consent to use the app and there are still 1.5 percent of the time it doesn't work. Ultimately, that could be many people.

And how does the AI ​​differentiate between normal pneumonia and corona infection?
If you compare your own car with a friend's Mercedes, you will find that you can only tell the difference between them by the noise they make when they drive by. We tend to think too quickly that it makes more sense to work with the visual because we can see it and discuss it. Nonetheless, it is easier to identify asymptomatic infections than to distinguish between regular pneumonia and Covid infection. We just need more data. This is also the reason why the AI ​​still has a few errors in identifying the symptomatic cases.

More on the subject: Digital corporations have a reputation for only filling their workforce with programmers. Now they are working with humanities scholars. Can that go well?

© Handelsblatt GmbH - All rights reserved. Acquire usage rights?