Every three to five years, on average, cities and regions experience a major dengue outbreak. This has a lot to do with the dynamics of infection and a certain herd protection — when a large part of the population is infected and generates antibodies against certain subtypes of the virus.
In addition, the number of Aedes mosquitoes, the number of breeding sites, and climate issues also influence this dynamic.
Despite all this epidemiological complexity, a study by Impa (Instituto de Matemática Pura e Aplicada) shows that it is possible to predict a large outbreak of dengue only from meteorological data. This greatly simplifies life, and can help many cities prepare for the upcoming problem — up to six months in advance.
It is not possible to say that the mathematical model in question is definitive, since the accuracy varied among the cities whose meteorological data were analyzed.
While for Rio, São Luís and Aracaju the hit rate was 100%, for Recife it was only 20% — that is, in this case 80% of the predictions were wrong. In the other cities tested, Belo Horizonte, Manaus and Salvador, the model was correct 80% of the time, the same percentage that represents the overall performance.
The machine learning algorithm was fed with daily rainfall and temperature data for the period between 2001 and 2012 for each of the cities. The test to measure accuracy was made for the years 2013 to 2017.
“Usually this is an analysis that requires many types of data, but information, for example, from the beginning of the last century. Then we have to adapt to this scarcity”, says Caio Souza, first author of the study.
If there is barely complete meteorological data, imagine others, such as the number of mosquitoes and the presence of antibodies in the population. It could even be interesting to consider these variables in the model, but, due to unavailability, rain and temperature had to do all the work, so to speak.
“The cool thing about machine learning is this: identifying patterns of things that are difficult to discover, like dengue outbreaks, from easy-to-obtain data, like weather,” says Souza, who is a computer engineer and doctoral student. at Impa.
That’s what machine learning is all about: identifying patterns of things that are hard to figure out, like dengue outbreaks, from easy-to-get data like weather.
The leap forward in the model proposed by Souza and colleagues, in addition to improving a previous code, rewriting it in a new programming language (Python), was to process the “signal” generated by the rainfall and temperature data sets and use this version modified when making the projections. “It’s like turning a book into a short summary, with just the information the model needs.”
If what goes into this black box is meteorological data, the output is of the binary type: “yes” or “no” for the occurrence of a large outbreak of dengue in the following months. Unlike other types of projection, Souza’s proposal does not aim to estimate the number of cases.
Another limitation of the model, the authors explain, is that it does not consider whether or not there were measures that directly attack the vectors of the disease, mosquitoes such as Aedes aegypti. An effective killing of insects, with the destruction of breeding sites and, along with them, eggs and larvae, can prevent a large outbreak in a region, despite a climatic propensity.
The work was recently published in Expert Systems With Applications, and is also signed by Pedro Maia, from the University of Texas, Lucas Stolerman, from the Harvard Medical School, as well as Vitor Rolla and Luiz Velho, both from Impa.
Both the program code and the data used are open to others who want to improve the model or use it to understand other diseases and conditions. According to Souza, the code can be useful for analyzes in which time series (the sequence of climatic data, for example) have some relevance, for example malaria and yellow fever.
The effective use of the new tool to predict dengue outbreaks depends on the interest of governments and funders so that the research can continue.
Dengue cases, when identified early and properly treated, have low lethality, less than 1%. Even so, the disease represents an important burden for countries in Latin America and Asia due to the large number of infected people, estimated at 390 million in the world, annually, with a quarter presenting clinical manifestations.
