Wednesday, 09 August 2023 10:03

How We Used Machine Learning to Investigate Where Ebola May Strike

Series: Roots of an Outbreak: How the Next Pandemic Could Start

The next pandemic is just a forest clearing away. We’re not doing enough to prevent viruses from spilling over from wildlife to humans.

Part Five
Southwest Nigeria

We’re investigating the cause of viruses spilling over from animals to humans — and what can be done to stop it. Read more in the series →

Part Five
Southwest Nigeria

We’re investigating the cause of viruses spilling over from animals to humans — and what can be done to stop it. Read more in the series →

The bright spots on the map struck us like a lightning bolt.

We had spent months teaching a computer about the Ebola virus –– feeding it information about the landscapes and populations in places where the disease had previously emerged, showing it how to analyze those outbreaks for patterns, and then instructing it to flag other areas that looked similarly perilous.

Some of the highlighted spots were predictable; the virus had repeatedly ravaged one of those countries.

But we didn’t expect our model to light up Nigeria, the most populous country in Africa. The West African nation and international travel hub has never seeded an Ebola outbreak, but just a year ago, it served as the springboard for another virus to travel into Europe and the Americas and spread across the globe. However that virus, mpox, originally known as monkeypox, is rarely fatal.

What if it had been Ebola, which kills about half of the people it infects?

We asked Nigerian public health officials whether they were concerned.

“Ebola is not part of our top concerns any more,” said Oyeladun Okunromade, the director of surveillance and epidemiology at the Nigeria Centre for Disease Control.

In the aftermath of the 2014 West African Ebola epidemic, the worst on record, Nigerian officials were on high alert. But last year, they took the virus off the list of the top infectious diseases the country needed to prepare for, downgrading Ebola in relation to threats like mpox, which Nigeria was actively fighting.

The disjoint between how our model sees Nigeria’s risk and how the nation’s health officials view it reveals a weakness in the way that governments and public health experts are preparing for future pandemics. The methods many countries use to rank threats focus mainly on factors that occur after an outbreak has already begun, such as the potential economic impact of an epidemic. Or they rely on past cases, looking at where a pathogen has previously struck.

Neither approach considers the root causes.

We’ve spent more than a year digging into the question of what causes outbreaks and what the world can do to prevent them. And we’ve learned that while science has advanced so we’re starting to understand the complex factors that trigger an outbreak, the world is not doing nearly enough to try to head off the next big one.

Most emerging infectious diseases come from wildlife. Those outbreaks require two essential elements: animals that carry a virus and opportunities for those animals to infect people.

Many of these fateful jumps, known as spillovers, have happened in forested, but populated, areas where trees have been cut down. Researchers have found that when people cut trees in patches, leaving the landscape dotted with holes like Swiss cheese, that creates more pockets and edges where humans and infected animals can collide. That world-shaking Ebola outbreak in 2014, for example, started in a Guinean village surrounded by a ring of forest.

Models that incorporate these environmental drivers could help countries look forward instead of backward as they determine how to allocate resources. Solomon Chieloka Okoli, an epidemiologist who works for Nigeria’s field epidemiology and laboratory training network, said his country, like many others, tends to react to outbreaks after they’ve started instead of trying to prevent them. That isn’t enough, Okoli said. “Being proactive is the best line of defense — if you wait, a lot of people will have died before you can get yourself together.”

Our model, created in consultation with scientists, was able to identify ecological factors that were common to past Ebola spillovers. The resulting risk map should be enough to prompt action, according to Christina Faust, a fellow at the University of Glasgow, Scotland, whose research focuses on how human activities like deforestation affect disease transmission.

Ebola often starts with a fever, so governments should invest in surveillance systems that help health authorities track patients with fevers, she said. “We should be watching these areas.”

Training Computers to Learn How Outbreaks Work

Models are not crystal balls; they can’t say exactly when or even whether a place will be hit with an outbreak. But they are great for understanding risk — where it is growing and where it may be shifting to.

“I love these as advocacy tools, because they’re meant for action,” said Dr. Maria Van Kerkhove, an infectious disease epidemiologist at the World Health Organization. “We just want these types of maps to inform and say: Make sure you’ve considered what might be circulating that you haven’t yet detected.”

We were curious to see where risky deforestation patterns are happening today. So we turned to a machine learning technique called “random forests” (no relation to actual tree-filled forests!) that can be used to spot patterns that might explain how some previous Ebola outbreaks happened. We limited our analysis to the geographic area where wildlife that can transmit Ebola is most likely to be found. This area covers 27 African countries from Guinea to Uganda.

We started with seven locations of past Ebola outbreaks that researchers have linked to forest loss. Then we selected 23 parameters, including demographic characteristics like the change in population from 2019 to 2021 (the most recent available data), as well as forest characteristics like the amount of tree loss and the patchiness of the surrounding forests.

We pulled data from satellite imagery and online population databases, fed it to the model and asked the computer to examine these factors across the seven known Ebola outbreaks. The model digested all this information and determined the relative importance of each parameter.

We also asked it to compare the outbreak sites to a set of places that were in the area where Ebola-carrying animals could live but had not seen an Ebola spillover.

Then we gave it a list of 1,000 candidate villages that had at least the same population size as previous Ebola spillover sites. (The 1,000 candidates were a random sample of all the villages that met our criteria; we weren’t able to run our model on the full set because of the amount of time and computing power that would have been required.) We asked the computer: Are there places that look very similar to past outbreak sites?

The model identified 51 locations with patterns of tree loss very similar to the seven previous Ebola outbreaks. The Democratic Republic of Congo had 16, which made sense; the country has recorded more than 10 Ebola outbreaks since the 1970s. The model highlighted additional spots in Ghana, Burundi and Benin.

More than half of the locations of concern, 27, were concentrated in Nigeria.

Out of the around 1,000 locations ProPublica tested, the model flagged 51 , mostly in Nigeria and the Democratic Republic of Congo.

Nigeria

Flagged locations

Forest loss,

2019-2021

Ibadan

Abeokuta

Ijebu Ode

Lagos

Gulf of Guinea

The model looks at factors like how patchy the area had become in the previous two years and how much “edge” has been created.

New patches of forest loss

Edge regions

Out of the around 1,000 locations ProPublica tested, the model flagged 51 , mostly in Nigeria and the Democratic Republic of Congo.

Nigeria

Flagged locations

Forest loss,

2019-2021

Ibadan

Abeokuta

Ijebu Ode

Lagos

Gulf of Guinea

The model looks at factors like how patchy the area had become in the previous two years and how much “edge” has been created.

New patches of forest loss

Edge regions

Out of the around 1,000 locations ProPublica tested, the model flagged 51 , mostly in Nigeria and the Democratic Republic of Congo.

Nigeria

Flagged locations

Forest loss,

2019-2021

Ibadan

Abeokuta

Ijebu Ode

Lagos

Gulf of Guinea

The model looks at factors like how patchy the area had become in the previous two years and how much “edge” has been created.

New patches of forest loss

Edge regions

Out of the around 1,000 locations ProPublica tested, the model flagged 51 , mostly in Nigeria and the Democratic Republic of Congo.

Nigeria

Flagged locations

Forest loss,

2019-2021

Ibadan

Abeokuta

Ijebu Ode

Lagos

Gulf of Guinea

The model looks at factors like how patchy the area had become in the previous two years and how much “edge” has been created.

New patches of forest loss

Edge regions

(If you — like us — are a nerd and want to read about our model in more detail, here is a comprehensive methodology.)

Why Nigeria’s Deforestation May Increase Its Risk

We were initially surprised to see the cluster of flagged locations in the southwest region of Nigeria, since the nation has never been the starting point for an Ebola outbreak. (The country has dealt with Ebola patients before, after an infected traveler flew to Lagos from Liberia during the West Africa outbreak in 2014.)

But we came to learn that Nigeria has experienced rapid deforestation over the past two decades. According to Global Forest Watch, the country has lost over 3,800 square miles of forest since 2001, and the rate of that loss has been accelerating. Nigeria has cleared the equivalent of nearly 170,000 football fields every year since 2017.

This is in part because energy prices have risen, making conventional fuel sources like kerosene unaffordable for many families, said NwaJesus Anthony Onyekuru, a professor of resource and environmental economics at the University of Nigeria. “They don’t want to use kerosene to cook, so they use wood,” he said.

Our model showed that this rapid forest clearing has happened in the dangerous, patchy pattern that researchers say leads to more interactions between humans and wildlife, and therefore increases the chances of spillover.

Scientists have found that bats can shed more virus when they’re stressed, such as by losing their habitats. That means that hunters may now encounter wildlife that is more likely to transmit a pathogen. Some Nigerians eat bats. Hunger has driven other residents to hunt for monkeys and rats in the forests, according to the epidemiologist Okoli. He said that consumption of large rats in the country’s southern region may have spurred the recent mpox outbreak.

Local deforestation has contributed to an increase in Lassa fever cases, said Dr. Charles Akataobi Michael, a senior technical officer at the Africa Centres for Disease Control and Prevention. Lassa fever can cause bleeding from the mouth, nose and gastrointestinal tract in severe cases, as well as neurological symptoms like hearing loss. The virus is carried by rodents, and people can be infected when food or household items are contaminated with the rodents’ urine or droppings.

The virus has been circulating in areas where people burn trees to create farmland, said Michael, destroying the rodents’ habitat. “They go to human habitats as a result of bush burning and deforestation to find food,” he said. “As we continue to alter the environment, the risk of disease outbreaks are increasing significantly.”

As the country’s population continues to grow rapidly, residents are chipping away at the forests to make room for farms. This land-use change is another way that risk may be increasing: Many outbreaks around the world have started when a virus jumped first from wildlife to a farm animal and then made another leap to humans. That includes deadly forms of bird flu and the brain-inflaming Nipah virus, which was immortalized in the movie “Contagion.”

Though we were initially surprised, we’ve since learned that Nigeria has appeared in other academic models as a potential Ebola hot spot. A 2019 analysis, published in the journal Nature Communications, identified Nigeria as a country at risk for an Ebola outbreak based on both current conditions and future climate and socioeconomic drivers.

In 2014, a different group of scientists used human and animal data to map locations most at risk of an Ebola outbreak. Among countries that had never reported an Ebola spillover before, Nigeria was at the top of their list. We know that Ebola isn’t constrained to country borders — after all, the worst Ebola outbreak to date started in Guinea, where the virus hadn’t previously been thought to be a threat. And this year, Marburg, Ebola’s cousin, has spread in two countries that had never before recorded an outbreak.

David Pigott, who led the 2014 analysis, said looking at prior cases isn’t the best way to evaluate risk: “The conversation of preparedness should not just be a function of what happened in the past.”

But that, we learned, is exactly what Nigeria is doing.

The Gap Between Knowledge and Action

The Nigerian experts we interviewed all acknowledged the importance of environmental factors in increasing outbreak risk. But many said that not much has been done to try and mitigate dangerous deforestation.

Okunromade, from the Nigeria CDC, helped create its One Health Strategic Plan — a national action plan based on the “one health” principle that the well-being of the environment, animals and humans are deeply interconnected. She said the government has brought together experts on human and animal diseases so that they can share information about pathogens such as mpox, Lassa fever and bird flu.

Yet when we asked what the country was doing to address environmental risks, she wasn’t aware of any initiatives, though she said it may be possible that other agencies were telling the public about the dangers of deforestation.

Okunromade said that experts used a tool developed by the U.S. Centers for Disease Control and Prevention to assess the risks of dozens of diseases that come from animals. The process has local experts select five criteria, commonly including epidemic potential or a country’s diagnostic capacity, and answer questions about different diseases for each criteria. Based on the answers, the diseases get scored as having a higher or lower priority.

When Nigerian officials ran this exercise in 2017, the devastating Ebola epidemic was fresh in their memories, and Ebola made the top five. “Looking at West Africa, at the countries surrounding us, looking at Sierra Leone, looking at Liberia, they were the worst hit. So that was why it made the list,” she said.

Ebola is a disease that would typically rank highly using the U.S. CDC’s tool because it gives more points to pathogens with a higher fatality rate. In 2022, Nigerian officials re-did the ranking exercise and initially, Ebola was still in the top five, but the officials felt it was more important to look at recent cases. Since there hasn’t been an Ebola outbreak in neighboring countries in recent years, the disease fell off their priority list, according to Michael, from the Africa CDC, who participated in the ranking process.

The CDC’s tool, which has been used by more than two dozen countries, does not require consideration of environmental causes like deforestation when ranking threats. Dr. Casey Barton Behravesh, the director of the U.S. CDC’s One Health Office, said that the process does not mandate which criteria should be considered and “it’s up to the country or region to decide on the criteria of greatest importance to them.” In examples she provided, two workshops, conducted in Alaska and the Economic Community of West African States, included a question about whether climate change would impact a disease. Some other countries considered the environmental impact of a potential outbreak, but they did not look at environmental factors that could increase the chance of a spillover. None of the examples included a question about deforestation.

There’s hope that new tools will evolve. The WHO is currently working with Pigott, who is an assistant professor of health metric sciences at the University of Washington, and other academics to develop risk maps for 16 different pathogens. Their model will incorporate data on environmental drivers of outbreaks. They aim to publish their work in a journal in future months, according to Pigott.

Pigott acknowledged that it can be hard for governments to prioritize a rare event like an Ebola outbreak. Still, he said, preparing for a disease like Ebola can be incorporated into plans for other pathogens. A malaria test may be the most logical place to start in a patient with a fever; if that is negative, health workers should be ready to test for Ebola, he said. But that only works if they are aware of the potential threat.

Ultimately, putting a disease on a priority list is only the first step. True prevention will need to address people’s lives, said Okoli, the Nigerian field epidemiologist: “If you say, ‘Don’t cut the bush to make charcoal,’ then you need to provide gas. If people are saying, ‘When I’m hungry, I get wild game,’ then you need to make it easier to get meat from the shops. You need to provide an alternative.”

Preventing the next outbreak from starting, Okoli said, should not be that hard. “It’s just about the political will and the willingness of the government to do something.”