Alibaba develops AI that can identify coronavirus infections with 96% accuracy

midian182

Posts: 5,759   +46
Staff member

Nikkei Asian Review reports that the diagnoses algorithm was developed by Alibaba's research institute Damo Academy, which it established in 2017. It’s been trained on more than 5,000 confirmed coronavirus cases and can identify differences in CT scans between patients infected with COVID-19 and those with ordinary viral pneumonia with 96 percent accuracy.

The AI could be especially helpful because of the speed at which it can make a diagnosis. It can complete the identification process in 20 seconds, while a doctor will usually take between five and 20 minutes to analyze a CT scan, and they can sometimes include more than 300 images.

The algorithm’s creators said it also includes “the latest treatment guidelines and recently published research.”

Alibaba said more than 100 hospitals would adopt the system in the provinces of Hubei, Guangdong, and Anhui, all of which have seen a high number of infections.

Being able to diagnose accurately those with novel coronavirus at a faster pace could ease the pressure on hospitals in China, which are overwhelmed with patients. It could also allow staff to spend more time treating the infected.

In the most recent tech-related coronavirus news, we heard that Apple has been sending care packages (that include iPads) to employees stranded in China. The virus has also caused Galaxy S20 sales to be 50 percent lower than the S10 in South Korea.

Image credit: SamaraHeisz5 and pang_oasis via Shutterstock

Permalink to story.

 

Squid Surprise

Posts: 3,151   +2,056
Headline is a bit misleading...

"It’s been trained on more than 5,000 confirmed coronavirus cases and can identify differences in CT scans between patients infected with COVID-19 and those with ordinary viral pneumonia with 96 percent accuracy."

That doesn't mean it can diagnose COVID-19 with 96% accuracy - it means it can differentiate it from viral pneumonia!!

How does it do on "normal" patients displaying no symptoms? Can it differentiate between cases of Influenza and COVID-19?

This is still a great start, but it isn't nearly as useful as the article makes it seem.
 

Uncle Al

Posts: 7,078   +5,438
While the news is promising I would want an analysis and verification from NIOSH and CDC before endorsing it. I trust they are sound results but the unfortunate reputation of any Chinese goods would cause me to need verification and regular verification of each batch .....
 

AfraidOfTheWind

Posts: 12   +4
Accuracy isn't the right metric here, with medicine there is more than 1 way to be wrong and it really matters how you are wrong. Example, telling someone they are sick when they are not, or, telling someone they are not sick when they are. Simple accuracy doesn't capture the nuance. 96% accuracy means 1 in 20 are going misdiagnosed. The acceptable error is probably to treat patients who are not sick.
 

Bullwinkle M

Posts: 324   +206

It shouldn't be too difficult to modify this tech for use as a mobile scanner

Simply cross reference facial scans with the database of Hong Kong protesters who had their mobile devices infected with Gov't backdoors and send an encrypted PM to the mobile incinerators whenever another one is identified with 100% accuracy

What?
Too soon?


 

RobertW

Posts: 21   +14
Um, ok, Let's apply some Bayesian reasoning to this claim. If the test is 96% accurate and the prevalence of coronavirus is 30% (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1628163/), the probability of having the disease given you test positive using this method comes out to 91%. I guess that's pretty good. The point of parity is a prevalence rate of about 5%. So this would be an acceptable failure rate for this type of disease. Good job Alibaba.
 

p51d007

Posts: 2,417   +1,686
LOL, hope it works, but I don't know considering some of the crap they sell on Alibaba if it works or not.
 

brucek

Posts: 388   +433
The tech enthusiast in me thinks this is cool, but as far as a medical innovation, I'm not that impressed given that you still need an expensive CT Scan and you still need a radiologist to examine the scan for all issues other than detection of coronavirus.
 

seeprime

Posts: 460   +495
That's a nice find. But, here in the medical third world USA, if you don't have $1000, or great health insurance, you may not be able to afford the CT scan.
 

brucek

Posts: 388   +433
CT Scans being a finite resource applies everywhere in the world, not just in the US.

Yes, I know there's a large issue about the relative merits of paying for the CT machines, the radiologists, the techs, and the space to house them via direct customer pay, or via insurers, or via tax.

But no matter which way you go on that, there's no changing the fundamental point that there is a limited capacity to perform these scans, and some mechanism needs to gate who gets the limited number available.

There's also the fact that there is radiation involved (about 200 chest X-rays worth, or seven years of normal environmental radiation) and while serious problems are very rare, these machines have on occasion caused significant harm (the NY times had a horrifying article a while back on massive overdoses of radiation being delivered due to programming bugs.) You don't want any more of these than you really need.

Bottom line another, more scaleable, method of diagnosis would be much appreciated.
 
  • Like
Reactions: Black Paper

jbc029

Posts: 81   +149
Headline is a bit misleading...

"It’s been trained on more than 5,000 confirmed coronavirus cases and can identify differences in CT scans between patients infected with COVID-19 and those with ordinary viral pneumonia with 96 percent accuracy."

That doesn't mean it can diagnose COVID-19 with 96% accuracy - it means it can differentiate it from viral pneumonia!!

How does it do on "normal" patients displaying no symptoms? Can it differentiate between cases of Influenza and COVID-19?

This is still a great start, but it isn't nearly as useful as the article makes it seem.
While this is the actual point of the story, it still has purpose. This software *cannot* tell if someone has a normal flu bug or CV before it has progressed to seriously threatening status.

However, as volume mounts and numbers of people with severe respiratory distress build, this can provide an option to alleviate a bottleneck of people waiting for directed treatment due to limited staff.

Dialing back edge case criteria can raise the accuracy a bit further and allows routing of the most ambiguous cases to an actual person, while getting help for those that are obviously in one camp or the other.

It's a good thing. It's why AI exists. Not to replace people, but to augment them.
 

Yynxs

Posts: 544   +183
TechSpot Elite
There's an interview with a radiologist from Johns Hopkins late last week on Fox News. Blown off as usual by people that don't understand health and scientific announcements or how to explain them to their public. He stated the Coronavirus lung damage has a definite and discernible pattern and the information was being disseminated to radiologists across the US. There were pictures. Fox didn't show any closeups.

Likely the AI is seeing the same. Now the concept of running a CT on every man, woman, and (especially) child with a chest and breathing problem during the school year and cold/flu season simply boggles the mind.

Our doctors are using the mantra: Out of country? Fever? Chills? With fever being the code word. Corona presents with fever.

Also, my last calculated deathrate worldwide was 3.43%. Still climbing.
 

jbc029

Posts: 81   +149
Also, my last calculated deathrate worldwide was 3.43%. Still climbing.
Newsflash, regular influenza can also present with fever. The similarities are what has led to such a remarkably high mortality rate in the first 45 days. Because it was largely dismissed as normal flu until the alternate virus was found. Damage was already done, because the progression of the disease wasn't identified early enough.

Mortality in Wuhan, still the epicenter for the outbreak, for the month of February is .7%. That's cases diagnosed after Jan 20 and treated through the active infection period. Now that they know how to identify it a lot earlier, it is much much less lethal. Things like this AI help in getting that response time down.
 

Yynxs

Posts: 544   +183
TechSpot Elite
Newsflash, regular influenza can also present with fever. The similarities are what has led to such a remarkably high mortality rate in the first 45 days. Because it was largely dismissed as normal flu until the alternate virus was found. Damage was already done, because the progression of the disease wasn't identified early enough.

Mortality in Wuhan, still the epicenter for the outbreak, for the month of February is .7%. That's cases diagnosed after Jan 20 and treated through the active infection period. Now that they know how to identify it a lot earlier, it is much much less lethal. Things like this AI help in getting that response time down.
I am somewhat curious about propaganda values among various populations, especially, 'informed populatons'.

Which "...Mortality in Wuhan..." rate are you quoting?
1) The official Chinese Governments' released to Wuhan population?
2) The post-crematorium bodies acknowledged?
3) The "less than influenza" from the Canadian and British CDC?
4) just the number of Chinese medical personnel in full respirator gear?
5) The before 'diagnosed with CT scan' change?
6) 'only badly affects the elderly or those with "other health problems"?
7) 'diagnosed with CT scan' being dropped as a method?

As mentioned above, how is an AI interpretation of a highly limited and especially risky to infants and children (ignoring pregnant women) medical procedure useful?

If your 'news flash' about 'regular influenza' (the most inoculated annually disease in the first world) presenting with fever helped at all, I would suggest a modified FLIR app for your smart phone with a GT 98F/37C overlay filter to simply hold up and point at any other individual and even mass scanning crowds. Unfortunately, human variance being what it is, the word "fever" is not exact. (my personal normal body temp is 96.5 and at 98.6 resting, I'm feeling pretty sick as an adult).

Hold on to that "0.7% rate. We can use it later to compare to 'endemic' after the rest of the worlds' bodies are disposed of.