When a mysterious illness first pops up, it can be difficult for governments and public health officials to gather information quickly and coordinate a response. But new artificial intelligence technology can automatically mine through news reports and online content from around the world, helping experts recognize anomalies that could lead to a potential epidemic or, worse, a pandemic. In other words, our new AI overlords might actually help us survive the next plague.
These new AI capabilities are on full display with the recent coronavirus outbreak, which was identified early by a Canadian firm called BlueDot, which is one of a number of companies that use data to evaluate public health risks. The company, which says it conducts “automated infectious disease surveillance,” notified its customers about the new form of coronavirus at the end of December, days before both the US Centers for Disease Control and Prevention (CDC) and the World Health Organization (WHO) sent out official notices, as reported by Wired. Now nearing the end of January, the respiratory virus that’s been linked to the city of Wuhan in China has already claimed the lives of more than 100 people. Cases have also popped up in several other countries, including the United States, and the CDC is warning Americans to avoid non-essential travel to China.
Kamran Khan, an infectious disease physician and BlueDot’s founder and CEO, explained in an interview how the company’s early-warning system uses artificial intelligence, including natural-language processing and machine learning, to track over 100 infectious diseases by analyzing about 100,000 articles in 65 languages every day. That data helps the company know when to notify its clients about the potential presence and spread of an infectious disease.
Other data, like traveler itinerary information and flight paths, can help give the company additional hints about how a disease will likely spread. For instance, earlier this month, BlueDot researchers predicted other cities in Asia where the coronavirus would show up after it appeared in mainland China.
The idea behind BlueDot’s model (whose final results are subsequently analyzed by human researchers) is to get information to health care workers as quickly as possible, with the hope that they can diagnose — and, if needed, isolate — infected and potentially contagious people early on.
“The official information isn’t always timely,” Khan told Recode. “The difference between one case in a traveler and an outbreak depends upon your frontline health care worker recognizing that there is a particular disease. It could be the difference in preventing an outbreak from actually occurring.”
Khan added that his system can also use an array of other data — such as information about an area’s climate, temperature, or even local livestock — to predict whether someone infected with a disease is likely to cause an outbreak in that area. He points out that, back in 2016, BlueDot was able to predict the appearance of the Zika virus in Florida six months before it actually showed up there.
Similarly, the epidemic-monitoring company Metabiota determined that Thailand, South Korea, Japan, and Taiwan had the highest risk of seeing the virus show up more than a week before cases in those countries were actually reported, partially by looking to flight data. Metabiota, like BlueDot, uses natural-language processing to evaluate online reports about a potential disease, and it’s also working on developing the same technology for social media data.
Mark Gallivan, Metabiota’s data science director, explains that online platforms and forums can also give an indication that there’s a risk of an epidemic. Metabiota also claims it can estimate the risk of a disease’s spread causing social and political disruption, based on information like an illness’s symptoms, mortality rate, and the availability of treatment. For instance, at the time of this article’s publication, Metabiota rated the risk of the novel coronavirus causing public anxiety as “high” in the US and China, but it rated this risk for the monkeypox virus in the Democratic Republic of the Congo (where there have been reported cases of that virus) as “medium.”
It’s hard to know just how accurate this rating system or the platform itself can be, but Gallivan says the company is working with the US intelligence community and the Defense Department on issues related to the coronavirus. This is part of Metabiota’s work with In-Q-Tel, the nonprofit venture firm associated with the Central Intelligence Agency. But government agencies aren’t the only potential clients of these systems. Metabiota also advertises its platform to reinsurance companies — reinsurance is essentially insurance for insurance companies — that might want to manage the financial risks associated with a disease’s potential spread.
But artificial intelligence can be far more useful than just keeping epidemiologists and officials informed as a disease pops up. Researchers have built AI-based models that can predict outbreaks of the Zika virus in real time, which can inform how doctors respond to potential crises. Artificial intelligence could also be used to guide how public health officials distribute resources during a crisis. In effect, AI stands to be a new first line of defense against disease.
More broadly, AI is already assisting with researching new drugs, tackling rare diseases, and detecting breast cancer. AI was even used to identify insects that spread Chagas, an incurable and potentially deadly disease that has infected an estimated 8 million people in Mexico and Central and South America. There’s also increasing interest in using non-health data — like social media posts — to help health policymakers and drug companies understand the breadth of a health crisis. For instance, AI that can mine social media posts to track illegal opioid sales, and keep public health officials informed about these controlled substances’ spread.
These systems, including Metabiota’s and BlueDot’s, are only as good as the data they’re evaluating. And AI — generally — has a problem with bias, which can reflect both the engineers of a system and the data it’s trained on. And AI that’s used within health care is by no means immune to that problem.
Still, all of these advancements represent a more optimistic outlook for what AI can do. Typically, news of AI robots sifting through large swathes of data doesn’t sit so well. Think of law enforcement using facial recognition databases built on images mined from across the web. Or hiring managers who can now use AI to predict how you’ll behave at work, based on your social media posts. The idea of AI battling deadly disease offers a case where we might feel slightly less uneasy, if not altogether hopeful. Perhaps this technology — if developed and used properly — could actually help save some lives.