Rift Valley fever and the problem of forecasting

The notion that prediction is difficult, especially about the future, is absolutely true in the domain of infectious disease. Despite the difficulty we must try, and I think there's reason for hope, given recent studies utilizing powerful machine learning techniques and diverse data. There's much innovation being brought to the issue.

Recently, a topic I've written about in the past has appeared in the news: Forecasting of Rift Valley fever (RVF) in East Central Africa. Bernard Bett has written a thoughtful piece on the PLoS Translational Global Health blog, which I recommend reading, examining the gamut of public health tools and responses needed in order to combat RVF. He begins by framing the issue succinctly, 

Recent climate predictions suggest East Africa may be in line for an epidemic of Rift Valley fever -- an infectious disease which can hit people, their livestock and livelihoods, and national economies hard. Data from the Climate Prediction Centre and the International Research Institute for Climate and Society suggest there is a 99.9% chance there will be an El Niño occurrence this year, with a 90% chance it will last until March/April 2016. At least two of the most recent Rift Valley fever epidemics in East Africa -- those in 1997/98 and 2006/2007 -- were associated with El Niño weather patterns, with Kenya suffering losses amounting to US$32 million in the most recent. Given the strong predictions of an El Niño occurrence, and the established association between El Niño and Rift Valley fever risk, countries in the Horn of Africa need to start laying out measures to manage the developing risk. . . 

The health, economic, and social costs of this disease are well known and there is a wealth of research establishing both RVF epidemiology and its strong ties to climate (including El Niño) and the environment. Nonetheless, there was little early response undertaken given remotely sensed (i.e., satellite-based) RVF forecasting in 2006-07. Peter Roeder described the situation in a 2007 ProMED post (archive 20070112.0164),

It is interesting, if rather disheartening, to watch another RVF epizootic emerge and evolve in eastern Africa and to note that it is such a close recapitulation of events that occurred in 1997/8 and decades before. It is a recapitulation not only with respect to disease evolution but also in terms of national and international preparedness—or lack of it. Those who followed ProMED in those days will be aware that the epizootic attracted intense international attention and was closely reported in postings, which contain much useful information. Despite seminal work on developing early warning systems based on remote sensing . . . it seems that the capacity to respond has not improved greatly in the high-risk countries in Africa. 

We are presently seeing the emergence of a very powerful El Niño, possibly one of the strongest in the historical record, and this was forecast in mid-August of this year. While such a climate forecast, especially when combined with other data, could reasonably be interpreted as a 2-3 month warning of the potential for RVF in parts of Africa, it's important to appreciate the complexity of acting on such information. As I wrote in 2012,

A recent, comprehensive set of case studies of the 2006–2007 outbreak in East Central Africa was published in the American Journal of Tropical Medicine and Hygiene (August 2010), and many of the nuances are described there. For example, current preparations of the Smithburn vaccine have a shelf life of approximately 4 years. Outbreaks in the Horn of Africa region occur aperiodically, with a mean of near 10 years between outbreaks. Veterinary health authorities cannot spend scarce resources on continually replenishing a stock of RVF vaccine when other needs are present continuously. Nor can manufactures maintain large stocks that are likely to expire before sale. Thus, vaccine may not be available at any given time. Nonetheless, waiting until there is a need to manufacture vaccine is problematic.

In other words, although vaccination is a powerful strategy for protecting against RVF virus transmission, maintaining vaccine stocks isn't straightforward. Moreover, simply having vaccine available isn't enough: Effective and safe administration triggered by any early warning, such as the one described in the impressive study of Anyamba et al in 2009, is complicated. In the case of the 2006-07 outbreak, for example, by the time a warning was issued, early outbreak areas were already inundated by rains, making travel and delivery of supplies difficult. In fact, in some scenarios it may take up to 150 days from a RVF vaccine order until the successful acquisition of vaccine-associated herd immunity -- much greater than the few weeks of advanced warning the state of the art can current supply. (Note: There's a distinction between a statistical forecast for a specific disease and simply noting that the strongest El Nino in decades is going to mess with everything.)

If a disease forecast is to have impact, many factors must come into alignment, including the forecast supplying sufficient lead time, decisionmakers having enough confidence in the forecast to act, and the existence of a public health infrastructure capable of supporting an effective (and potentially complex) response. These are important issues to keep in mind when thinking about surveillance and early warning, regardless of the disease and setting.

(image source: Wikipedia)

First US MERS case announced

Yesterday the first case of Middle East respiratory syndrome (MERS) observed in the US was announced. The case is a man who flew to Chicago, Illinois from Riyadh, Saudi Arabia by way of London, England. After landing in Chicago, he took a bus to Indiana. He arrived in the US on April 24 and began experiencing shortness of breath, coughing, and fever on April 27. He presented to a community hospital in Munster, Indiana, on April 28; because of his symptoms and travel history, the man was tested for the MERS coronavirus.

Understandably, the story is being covered widely in the press. One can learn more about MERS from several sources. The ECDC MERS-CoV epidemiological update as of April 30 is an excellent summary of MERS globally, and additional resources can be found at the CDC MERS Website. UPMC has also published a nice summary of the global situation. 

It would be interesting to utilize mathematical modeling to estimate where, assuming the patient was infectious while on the bus and planes, new cases might be likely to develop, geographically speaking. Such analysis could be helpful for instituting intensified surveillance in the places most likely to have such cases; using models, it may be possible to forecast the most likely cities. A rich set of methods, spanning the 1960s to the present, has been developed to understand the spatio-temporal spread of influenza and those could probably be adapted for such purposes if data on domestic bus, rail, and air travel since the plane landed is available. And while it's true that significant uncertainties in latency and incubation periods, transmission rates, and other epidemiologic parameters exist, there are techniques that allow us to estimate the effects of such uncertainties on model results.

Gathering the data, especially the transportation data, may be difficult. A NYT article this morning noted

The typical incubation period for MERS is five days, and the patient is not known to have infected anyone else. Airline passenger lists will be used to contact everyone who sat near him.

But because bus companies often do not know who bought tickets or who sat where, “that bus ride may be a challenge,” said Tom Skinner, a C.D.C. spokesman.

Information on rail passengers may be similarly difficult.

The time to develop adaptable modeling capabilities and gather such data is before new diseases emerge and spread. Of course, it's not possible to anticipate specific details about a new disease prior to emergence, but we do have some data on MERS already, plus lots of insight into how respiratory diseases spread in general. 

Regardless of models, it will be interesting to watch the MERS situation develop globally in the coming months.

(image source: CDC)

Disease prediction versus forecasting: The importance of words

Knowledge is power. With knowledge, it is sometimes possible to affect outcomes. However, in order to affect an outcome, one must know early enough to undertake some action. The problem is timeliness; how to know early enough is a very important question.

In order to partially solve this problem, we often make estimates of what we think we know -- in colloquial English, we predict or forecast -- before we are certain. In this way, we hope to give ourselves time to act.

Prediction and forecast may seem like they mean the same thing, but a quick spin around the Internet suggests that there is little or no consensus on what it is. Some fields use the terms differently; the climatologists, for example, have a nice set of working definitions that state clearly what is meant by the two terms.

But, for the case of making prospective predictions about outbreaks of infectious disease, especially if the predictions are made with the use of mathematical models, I think we should consider using the term "forecasting" for a couple of reasons.

First, whereas "prediction" tends to specify the time, location, and magnitude of an event (e.g., "We predict that a large outbreak of disease X will occur in region Y around date Z"), "forecasting" connotes a statistical statement (e.g., "We forecast that there is a 70% chance that a significant outbreak of disease X will occur in region Y around date Z"). In other words, "prediction" is a categorical statement that is either proven right or wrong (like predictions from a crystal ball), but "forecast" implies a probabilistic statement. "Forecast" conveys a degree of confidence. On the other hand, if we predict disease risk (see, e.g., Hay et al 2013), which is thought of as the probability of disease occurrence, then I think we're on solid footing.

Second, in some fields of science (e.g., physics), natural "laws" are known to incredible degrees of confidence. Given initial conditions to a specified accuracy, predictions can be made that are very reasonably believed. Unfortunately, infectious disease epidemiology is not one of those fields. Moreover, as George Box observed, "Essentially all models are wrong, but some are useful". Models tell us things about problems and allow us to do analysis in a systematic and (hopefully) objective way, but models are only models. They are not laws.

As Neils Bohr and others have noted, "Prediction is very difficult, especially about the future". We need to convey the finite confidence, whether high or low, in models to those using them. 

(image source: ABE Books)