A team of researchers from the Universidad Nacional de San Martín (UNSAM), led by Ezequiel Álvarez, is developing a Bayesian artificial intelligence model to predict dengue outbreaks in urban centers. The initiative is part of a collaboration with the provincial government of Buenos Aires on technological innovation.
At the beginning of October, the Buenos Aires Minister of Health, Nicolás Kreplak, announced a series of measures to combat dengue fever and called for vaccination. These include the use of telemedicine and the implementation of machine learning tools developed by the Unsam School of Science and Technology (ECyT) to prevent outbreaks of the disease.
“We expect the dengue season to be very big this year,” Kreplak told LA NACIÓN. In response, the health minister has opted for telemedicine to avoid the collapse of emergency rooms and hospitals, as "mild cases only require rest and hydration".
In addition to telemedicine, the province is also introducing an early warning system. Citizens can call 148 to report symptoms and receive medical assistance remotely. At the same time, this information will provide the health authorities with important data. Together with the information from epidemiological surveillance and the on-call service, they can create an artificial intelligence model to find out which blocks are most at risk of contracting dengue. “In this way, we will be able to take strategic action to reduce or contain the number of infections,” adds the minister.
The project
The development mentioned by Kreplak is an interdisciplinary project involving biologists, zoologists, epidemiologists and provincial government staff.
“In the dynamics of the dengue epidemic, as in any system, there are many interrelated processes that depend on certain orders of magnitude. For example, on the weather, on the number of cases reported in the last 20 days, on the population density in a block, on the number of calls to the 148 number, on the number of people visiting the emergency room, etc. All of this is part of the observed orders of magnitude. The number of infected mosquitoes in a block, the number of sick people in a block and the degree of water evacuation in a block, on the other hand, are among the unobserved variables,” describes Ezequiel Álvarez.
What is Bayesian artificial intelligence? It is an approach based on learning from experience combined with the application of Bayes' theorem. “In Bayesian machine learning, you take the part you know about, the relationships between processes, and from the quantities observed over time you can infer the true relationships. In this way, the model learns how the processes are related and can then infer the unobserved quantities using the information that comes from the observed quantities,” explains Alvarez.
According to the scientist, Bayes' theorem is a technique used when some data is ignored and other available and observed data is used. With this information, the probability of the unknown data is obtained. "We want to determine the probable distribution of the number of infected mosquitoes in each block of Greater Buenos Aires. This is our 'latent variable' that we want to determine. To do this, we use data on the number of calls to 148, population density, socioeconomic level, number of people in hospitals, vaccinations and the weather in recent days, among others. By linking all the variables and thanks to Bayesian AI, we will be able to determine the probability of mosquitoes in each block,” he says. Alvarez explains that although they do not know the exact number of mosquitoes, they can use the AI program to derive a probability distribution.
A Bayesian inference model is very useful to analyze a system as complex as the dengue epidemic, because infected people move, the mosquitoes that carry them move and so does the disease, as well as the time it takes for symptoms to appear and the days it takes for the mosquitoes to spread. There are many variables that are very complex and there are many processes that need to be considered.
Testing phase
Alvarez says that a first version of the model has already been tested. "It works very well and we expect it to get even better as it learns over time. At the moment, we are carrying out a preliminary analysis of last year's data using Bayesian AI to understand the correlation between the data. The final model still needs to be defined,” he says.
The expected date for receiving the data with the completed model is the beginning of December. According to Álvarez, it makes no sense to close it today if this program can continue to learn and improve until then. “To improve the model, for example, we create synthetic data and check the performance of the model. We hold group discussions and develop ideas on how we can incorporate more data. A few days ago, for example, we had the idea of adding geographical information about where people are campaigning,” he says.
The AI learns as it is used. “When the first data comes in for this year - we expect this to be between November and December - the model will learn and become more and more accurate when it comes to estimating variables that cannot be measured today, such as the probability of the number of sick mosquitoes in each block,” explains the scientist.
“When we run the model together with the data from the epidemic, we hope to find out that this probability distribution is higher in some places than in others. This information will be fundamental to take agile and efficient measures with the help of public policies to prevent dengue outbreaks",” concludes Álvarez.
The researcher hopes that this development will allow the relevant authorities to implement various measures to reduce dengue outbreaks, optimize public health resources and strengthen ongoing prevention; and as a result of all these measures, drastically reduce the number of people infected and the cost to the health system.
The most important thing about this study is the scalability of the project, which can also be used and exported for epidemics such as malaria or Crimean-Congo fever, among many others.
