Improving epidemic forecasting helps response teams target aid more effectively
In 2014, the first signs that an especially deadly Ebola epidemic might be brewing came on March 22, with reports of 49 cases in Guinea. By the end of August, the disease had spread to Sierra Leone and Liberia, bringing the total of likely, confirmed and suspected cases in West Africa to 3,685.
On September 23, 2014, the U.S. Centers for Disease Control and Prevention (CDC) announced a new modeling tool, called Ebola Response, to estimate the number and locations of future cases, and to construct aid scenarios that would eventually stop the epidemic. Without interventions, the number of infected people would double every 20 days and reach “extraordinary levels,” according to the Ebola Response forecast detailed in an accompanying academic paper. The model forecast between 550,000 and 1.4 million cases by January 20, 2015, without a concerted effort to stop the spread.
The authors of the CDC study, published barely six months after news of the first outbreak broke, noted several limitations to the Ebola Response tool. The first: Extrapolating (then) current trends to forecast all future cases might not be appropriate, the authors said, because spontaneous changes in travel or burial practices could alter future spread. Limiting model projections to shorter durations, such as three months, might be a better way of forecasting, the study noted.
With West Africa in full-blown crisis, aid directors acted according to Ebola Response recommendations. Its model scenarios suggested that isolating at least 70 percent of victims would be the most effective way to halt the virus’ spread. To this end, relief efforts focused on building Ebola treatment units in areas the model predicted would need them later.
But as the epidemic continued into roughly its 20th week and beyond, the virus did not spread with the same pattern or zeal that it did in the early months that were used to calibrate the Ebola Response models. What changed? Families canceled gatherings, transportation shut down and schools and workplaces closed, effectively narrowing or closing off entirely the avenues for viral spread. The number of cases in hot spots peaked quicker than expected.
The U.S. alone spent some $1.4 billion on relief efforts during the epidemic. Owing to inaccuracies in forecasts, much of it went to setting up Ebola treatment units that did not open until after the number of cases in those locations began to drop. Nine of the 11 units built in Liberia with U.S. aid never saw a single Ebola case during the epidemic, according to news reports.
By the end of the epidemic in early 2016, a total of 28,600 people had contracted Ebola in West Africa. Of those infected, 11,300 died. Preventive measures taken in communities, often without official aid, kept the victim numbers mercifully short of the 1.4 million worst-case projection.
A Better Forecast, a Quicker Resolution
When an infectious disease outbreak threatens mass sickness or death, accurate forecasts of contagion spread are key to keeping isolated occurrences from turning into widespread epidemic. Without them, resources are sent to the wrong locations and wasted.
A study forthcoming in the journal Manufacturing & Service Operations Management describes a new model that appears to provide much more accurate forecasts of contagion spread, as well as better guidance for effective aid deployment. In part by addressing some of the limitations noted in Ebola Response, the new model stands to greatly improve chances of bringing contagion under control in the early phases of disease outbreak. By iteratively pairing short-term spread forecasts with an allocation plan that adapts to the changing predictions, the model allows aid workers to see, as the crisis evolves, where medical personnel, equipment and treatment supplies should be distributed to offer the best chances of cutting off the epidemic.
The study’s authors, UCLA’s Anderson’s Elisa Long, and Eike Nohdurft and Stefan Spinler from the Kühne Institute of Logistics Management at WHU in Germany, test and demonstrate the model’s capabilities using data available in the early weeks of the 2014 Ebola outbreak.
The model produces forecasts very close to eventual reality in both the number of cases and the geographic direction of the contagion’s spread, according to the study. Rather than forecast months into the future, where predictions grow less and less accurate, the model design allows for quick and constant updates to short-term forecasts. Its forecasts improve over time as more data are collected.
Using four weeks of data from September 2014, for example, the model produced epidemic projections for the following four weeks with an aggregate error of only 4 percent, relative to the actual number of cases that occurred over this period. With 16 weeks of training data, the aggregate error decreased to less than 1 percent.
The researchers also offer a more effective plan for deploying resources based on the early forecasts, devising an allocation model that balances the needs in regions already in full-blown crises with those in yet-to-be-hit locations that are particularly vulnerable to infection. Using beds in Ebola treatment units as a quantifiable measure (beds are deployed with medical personnel and other resources), the study finds that aid allocated and reassigned frequently with updated forecasts can avert significant numbers of cases. The model’s allocations bring the contagion under control more quickly than the system used in the 2014 outbreak, even when data and resources are rather limited.
In one scenario that mimics the first 28 days of outbreak, assigning only 20 new beds per week to specified regions averts 23 percent of new cases. Long calls this a great result, especially since it relies only on the limited number of beds that were available at the time. “By allocating beds in a smarter fashion, a lot of cases could have been prevented, even using the imperfect data we had available at that time,” she said in an interview. “Ideally, we would like to increase the number of available resources to maximize the net impact.”
With 16 weeks of outbreak data, the same number of bed allocations can avert 86 percent of new cases, according to the findings. The giant leap in success of intervention at 16 weeks comes from the additional data collected over time, Long explained. For example, the data at four weeks into the epidemic did not predict the late outbreak of disease in Kono, Sierra Leone. Based on that early forecast, no resources would go there. But the model does pick up on a rapid increase in cases at 16 weeks, and it would have identified the need for more beds there earlier than the traditional model did.
The model stands to improve aid responses during an emerging outbreak of any rapidly spreading infectious disease about which data on cases are periodically reported, including Zika and influenza. With parameters that evolve over time to meet changing behavior on the ground, short forecast windows to reduce error rates, and an aid deployment strategy selected for optimal efficacy, the model offers a much-needed tool for predicting and getting ahead of disease.
People Adapt, So the Model Does, Too
The forecasting improvements in the new model result from the inclusion of data related to two variables that the CDC models did not effectively incorporate in 2014: changes in human behavior and geographic distance from other highly infected populations. Both factors are known to greatly affect disease spread but have been tricky to build into epidemiology models.
Behavioral changes by the people living in or near outbreaks — precautions that vary from changing burial rituals to families isolating themselves in their own homes — were key reasons that the 2014 epidemic did not become a worst-case scenario, according to follow-up studies. The new model accounts for people’s reduction in travel and their taking steps to avoid infection as the threat of disease near their own homes grows.
Logically, distance from outbreaks is a great predictor of where the disease will spread. But traditional epidemiology models often fail to adjust infectivity patterns based on geographic distance. They typically rely on data collected at the country level, which can lead to similar epidemic forecasts for a region hundreds of miles away from an outbreak and one right next door. The shortcomings of this approach have been highlighted in several studies about what went wrong in the 2014 response.
The new model fine tunes this risk factor by breaking down the country data into smaller regions, and then applying data from a third nearby region, Long explained. “For example,” she said, “if we know the number of past cases in Port Loko and Moyamba, then we can better predict the number of future cases in both of these regions and in nearby Tonkolili.” Tonkolili sits between Port Loko and Moyamba, two regions of Sierra Leone that were very hard hit by the 2014 epidemic.
The researchers discovered that incorporating either behavior or distance factors (but not both) in the model using the early 2014 data did not always improve predictions about where or how hard Ebola would hit next. In fact, occasionally error rates (the difference between the model’s forecast number of cases and the actual number of cases that occurred) were lower in specific regions if either factor was ignored, as illustrated in this map. Hover over any location to see the forecast error rates without behavior calculations and, separately, without distance effects.