先人一步，人工智能助力抗疫

On the last day of 2019, an artificial intelligence warning system run by Toronto startup BlueDot flagged a news report from China about a mysterious pneumonia strain in the city of Wuhan. The system, which sifts through 100,000 articles and online posts daily in 65 languages, alerted BlueDot’s human employees, who immediately saw parallels to the deadly SARS outbreak in 2003.

After switching to a system based on data from billions of airline passenger itineraries, BlueDot was able to determine almost instantaneously which cities worldwide were most at risk if the mystery illness spread. The company quickly sent out warnings to health authorities and other clients about what would come to be called the coronavirus outbreak, which has so far infected almost 100,000 people and killed more than 3,000 as of early March.

“Outbreaks don’t care whether it’s New Year’s Eve or not,” says Dr. Kamran Khan, CEO at BlueDot and a medical professor at the University of Toronto. “In order to get in front of these diseases and threats, we have to move even faster than they do.”

It’s a far cry from when Khan started BlueDot about seven years ago. Back then, mapping the potential spread of a virus and alerting authorities could take several weeks. And reluctant governments would sometimes sit on the data for weeks or months after that.

But the era of A.I. and big data has revolutionized tracking and forecasting the path of infectious disease outbreaks like that of the coronavirus. Fueled by algorithms that can translate languages and distinguish between different meanings—Anthrax, the heavy metal band, versus anthrax, the infectious disease—BlueDot and its rivals suck up all the data they can to uncover potential epidemics.

The earlier and more detailed their warnings are, the better health authorities can tell where to screen for infected people and allocate resources. A brief head start can save thousands of lives.

With the coronavirus, A.I.-based alerts helped the World Health Organization and China’s officials react more quickly than they did during previous outbreaks like that of SARS. Still, early warnings can do only so much: China’s government has been criticized for moving too slowly, while the U.S. stumbled over a lack of test kits.

The systems created by the startups feed off information generated by an ever more interconnected and mobile world, using everything from search keyword data to the location of people clicking on Wikipedia pages.

Much of the data comes from the world’s largest Internet companies, including Google, which supplies search keyword and location data to some pandemic-detection startups. Meanwhile, Facebook has shared aggregated data about users’ movements as well as posts mentioning the coronavirus from Facebook Groups and Instagram. Anonymized data from Twitter, China’s Tencent, and others also fuels the algorithms, which typically run not on the monitoring firms’ own computers but on servers managed by Amazon, Microsoft, and Google that use chips specifically designed for A.I.

To be sure, pumping huge amounts of information into A.I. and machine-learning systems is no guarantee of success. For example, Google shuttered a project that forecast the severity of seasonal flu outbreaks after it wildly overestimated the 2013 cycle. One problem was that Google’s own efforts to help people search for health care information fooled the system into forecasting that more people were getting sick.

The challenge for companies developing pandemic-detection systems is to ensure that they focus only on relevant bits of information, without getting misled by hysteria that’s unrelated to actual illnesses. That’s why all of the systems still rely on humans to look deeper into each case and why they frequently adjust the sources of information that their technology relies on. “You have to recognize that data is constantly changing based on what people are doing online and always have to retune your algorithms for that,” says John Brownstein, chief innovation officer at Boston Children’s Hospital and cocreator of another A.I. alert system, HealthMap, which warned about the coronavirus a day before BlueDot.

HealthMap’s A.I.-generated warning about the coronavirus was backed up by intel from local physicians in Wuhan who were sharing their concerns in an online forum called ProMed. Such posts are the “early canaries in a coal mine that can provide data pointing to do a deeper dive,” Brownstein says.

Using fresh data is also important. Initial simulations of how the coronavirus may spread relied on past air travel itineraries. But once the outbreak became known and governments began banning movement in certain regions of China, travel patterns changed, notes Mark Gallivan, director of data science at Metabiota, another startup using A.I. to detect pandemics. As a result, the San Francisco company updated its library of historical passenger information with real-time location data from millions of mobile phones. “The first four countries that showed the highest importation risk on Jan. 14 were actually the first four that ended up receiving cases,” he says.

Another approach is to eschew all the online chatter and news reports and instead use actual medical data. San Francisco startup Kinsa sells smart thermometers that work with an app to help people decide when to see a doctor. With about 1 million households and more than 1,000 schools using Kinsa gear, those thermometers provide clues about the spread of the seasonal flu in the U.S. The eight-year-old company claims to have exceeded the accuracy of the Centers for Disease Control’s flu forecast for some years and hopes to develop a system that could predict flu outbreaks in local areas up to three months in advance.

“The difference is the quality of the data,” Kinsa CEO Inder Singh explains.

Of course, the Kinsa method works only where people use its devices. In the U.S., that means most cities but not so much in rural areas. And the company has yet to expand to other countries, where even a $20 smart thermometer may be too pricey for most people.

Ultimately, though, more medical devices reporting directly to A.I. systems could make for the quickest and most accurate early-warning system, says Metabiota’s Gallivan: “For earlier detection, it’s about creating a much smarter public health and medical system.”

The data fueling A.I. pandemic predictions

Smart, connected medical devices

Millions of patients are treated with thermometers and other devices that send data to an app. The aggregate information can provide early warning of a cluster of patients with fever, for example.

Search keywords and locations

The questions people want answered at a particular time and place can signal an outbreak. But the data must be filtered carefully, as search queries can reflect hysteria as much as a real epidemic.

Local news articles

Reporters on the ground often write stories about unusual medical problems or virus outbreaks. The articles can be translated and analyzed using natural-language processing.

Air travel patterns

Airlines generate about 4 billion travel itineraries annually. That historical data can be used to predict how an outbreak may spread to other cities based on the most popular destinations from the source city.