Seasonal dynamics of influenza in Brazil: the latitude effect

Influenza is caused by viruses that infect the upper respiratory tract. It is characterized by sudden onset of high fever, myalgia, headache, and severe malaise, non-productive cough, sore throat, and rhinitis [1]. The disease severity ranges from mild illness to severe cases that may evolve to death, particularly among high-risk groups. Though prevention is possible with vaccines updated annually [2], worldwide epidemics are estimated to result in 3 to 5 million severe cases per year, causing 250,000 to 500,000 deaths [3]. Influenza epidemics have been well studied in high-income countries within temperate regions [4‐6]. In these regions, the seasons are well marked, resulting in a well behaved disease pattern [7]. In the tropics, seasons are not that well marked resulting in a less predictable seasonal dynamics. There are reports of semi-annual or year-round influenza activity [8, 9]. Studies of influenza in the tropics started recently. We found reports for: Brazil [10, 11], Vietnam [12], Madagascar [13], Myanmar [14] and China [15].

There are also a few ecological studies linking the influenza dynamics to differences in latitude. Some of them associate these differences to different climatic variables [10, 15‐22]. These studies use a myriad of methods, sometimes more akin to a data visualization than to statistical analysis. Moreover, ad hoc definitions of seasonality are commonly used. However, to assess the complex interactions of influenza with the environment, a thorough understanding of the temporal dynamics and the periodicity of influenza is needed.

Brazil is an interesting case study due to its continental dimension. Its territory ranges from 5 degrees north to 35 degrees south of the Equator line, encompassing six climatic sub-types [23]. The mortality rate by influenza averages 1.01 deaths per 100,000 inhabitants per year [24]. A further advantage of a single country study is that it minimizes issues of differences in the data collection procedures.

Studies on the ecology of influenza and its temporal dynamics in Brazil are recent. It started when Alonso et al. [10] produced a state-level study, using monthly mortality rates for the 1979-2001 period. Within a Fourier analysis framework, they show a relationship between latitude and deaths, both in intensity as in the time location of the peak activity. Later, Mello et al. [11] analyzed influenza isolated activity from two state capitals: São Paulo in the Southeast region and Belém in the North of the country. Using this rather limited sample, they advocate the need to use latitude information into the national vaccination calendar; and Schuck-Paim et al. [25] studied the impact of the 2009 influenza pandemic relative to mortality rates for all Brazilian states and also showed significant latitude differences.

The main goal of this article is to study the dynamics of influenza in Brazil at state-level, to verify where and when it exhibits seasonal behavior and explore its characteristics. We hypothesized that the diversity of environments across states would yield a variety of behaviors and just with dynamic techniques it would be possible to unearth the different seasonal patterns.

We used weekly data on SARI cases (hospitalization and deaths) related to influenza, to assess whether a significant seasonal signature is detected at the state-level, across the diverse Brazilian geographic states/scenarios. We adopted a combination of two statistical methods: (i) the wavelet analysis, a modern, dynamic version of the classical Fourier methodology, used here to detect periodic behavior; and (ii) Circular Statistics, a set of tools designed to model variables known to be periodic.

This merger allowed us to uncover a rich pattern of seasonal behavior, with marked dissimilarities between the subtropical states in the South and the tropical states in the North. Comparing the seasonal aspects of different states from a continental tropical country, encompassing so many different latitudes, will support a better understanding of the interactions between disease and ecology, improving strategies for prevention and control.

Our database contains weekly SARI incidence time-series from all 26 Brazilian states plus the federal district. We refer to them as the 27 Brazilian states. 21 states are situated in the tropical climate zone, corresponding to 81% of the territory. Towards the South lies the subtropical climate zone, with six states representing 14% of the territory. A smaller region is the semi-arid zone, occupying the other 5% of the territory. No state capital lies in it [23].

For a first overview of the data, we conducted an exploratory data analysis before the use of the time series methods. Figure 1 shows the distribution of incidence from all Brazilian states, ordered by latitude, north on top.

The boxplots on Fig. 1 highlight the overall tendency of higher SARI incidence in the higher latitudes of the south region. The state with the largest median incidence is Parańa, followed by Rio Grande do Sul and Santa Catarina. These are the southernmost states of the country, the only ones: (i) lying entirely below the Tropic of Capricorn, and (ii) with subtropical climate [23].

But this latitude effect alone does not capture the true dynamics of the disease. Qualitatively, the heat map on Fig. 1 demonstrates a clear trend towards influenza infection in winter months of lower latitude states. It shows the yearly evolution of the peak week of activity for each state. There, blue and red colors are around summertime, whereas the greens are close to the winter. It is only in the bottom part of the graph (South region) that we see a relation between the winter season and the SARI peak activity over the years.

In this study, we analyzed a 7-year weekly dataset of Brazilian state-level SARI time-series incidence from 2010 until 2016. The analyses have two primary goals: (i) inferring the existence of annual seasonality pattern for each state, and (ii) describing how summary statistics related to the seasonal activity can vary in a country with an extensive range of latitudes.

It is the first study (known to us) to use Brazilian SARI incidence to infer about influenza seasonality and evaluate this across all Brazilian states.

Influenza within temperate climate zones is known for their well defined seasonal pattern. This pattern is similar to those found in the Brazilian Southern (subtropical) states, the ones that are most affected by the disease, as shown in Fig. 5 and also by [38, 39].

Some authors have shown the existence of a latitude relationship with the spread of influenza [4, 10, 15‐17, 19, 20, 22]. Most of them associate it to climatic variables like temperature, precipitation, humidity or hours of sunshine [5, 15, 17, 18, 20, 40].

In particular, Alonso et al. [10] conducted an important study on the pat- tern of mortality by influenza and pneumonia in Brazil. Their main finding was a wave from the north, a pattern of disease peak activity starting earlier in the North of the country in May, with delayed peak activity in the South, in July.

Alonso et al’s study describes a well-behaved pattern for the aggregate Brazilian data: in their case, monthly deaths from 1979 to 2001, whereas our dataset comprises weekly SARI hospitalization/deaths cases from 2010 to 2016. A similar pattern is observed in only the southern states in the data presented here. The behavior of the time series for other states do not exhibit this trend. The behavior of the time series for other states tend to be far more erratic in nature, and seasonal behavior is found only in some cases.

Their approach did not benefit from the modern wavelet technique. Hence, no statistical analysis for reassuring that a given state-level series is seasonal is conducted. Also, for some states the confidence intervals reported are quite large: for the Rondônia, Acre and Mato Grosso the reported 95% confidence interval have lengths of 10.5, 9.6 and 11.4 months. In our dataset, these states exhibit no seasonal behavior.

According to our results, we propose a different hypothesis for the traveling wave: it does not start in the Amazon but instead in the Northeast, moving south along the coast. Our proposition is based is several pieces of evidence.

First, the seasonal signature in the North region as a whole is weak. From the six states of the region only one, the state of Amazonas was found to be seasonal. Additionally, seasonality trends are not as robust for this region. An inspection of the Amazonas power spectrum shows a seasonal behavior only in the second half of the sample. Hence, should more stringent criteria for seasonality be adopted, Amazonas would no longer be considered seasonal. To confirm the lack of seasonality for Amazonas, an inspection of the rose diagrams on Fig. 4 shows very complex phenomena, with clear multimodality. Events in March and July are as important as the seasonal peak in May. These suggest that perhaps the seasonal behavior detected in the state of Amazonas is accidental, a type I statistical error.

Alonso et al were puzzled that the wave started from extremely sparsely populated areas in the Amazon basin where roads are limited. Therefore, the wave would have no conceivable connection with the movement of people, since travel by air would likely move the disease directly to the most populated areas in the Southeast region.

Amongst the states that exhibit seasonal behavior, Amazonas peaks first in week 18, closely followed by Ceará (week 19), in the Northeast. Ceará is a far more likely starting point for the influenza season. Its capital, Fortaleza, has around 4 million inhabitants in the metropolitan area, the largest in the Northeast region. A possible explanation for why the wave starts in Ceará is the rainy season spanning from March to May, peaking around week 15. It is true that influenza is associated with cold and dry climate in temperate regions, but little is known in places where it never gets cold. In Fortaleza, the monthly maximum temperature oscillates between 29.1C^° and 30.4C^° year round. In fact, locals call this rainy season as the winter season, the only time of the year with a somewhat cooler sensation. The rainy season, therefore, may be the trigger for the disease.

Another piece of evidence suggesting that the wave starts in Ceará is the better mechanism of how it may move southbound. Viboud et al. [41] found that movement of people to and from work was a key determinant of regional disease spread. According to their result, a wave starting in Ceará is far easier to explain. The reason is that some of the main roads of the country are located alongside the coastline from Ceará to Rio Grande do Sul, so the movement of people and cargo there is constant and intense.

Regarding the five administrative regions in Brazil, two seem disconnected from this wave from the Northeast: the North and the Midwest regions. They possess 3.07 and 11.15 km of paved roads per million square km of territory, respectively. The numbers for the Northeast, Southeast and South regions (comprising most of the states in the wave) are 24.28, 37.00 and 31.23, respectively. For a rough comparison, the USA corresponding number is 450, while its GDP is ten times that of Brazil. [42]. So, it is possible that their isolation impacts the spread of influenza in this area.

Finally, we must mention that the state of São Paulo also exhibits a deviation from the expected pattern. It is Brazil’s most populated state, contributing to 1/3 of its GDP, where a severe influenza epidemic occurred in 2016. Its peak was in week 14, whereas the gravity center from 2010 to 2015, that is, excluding 2016, was the 21.03 week. The exercise of excluding the 2016 cases was conducted and resulted in São Paulo being seasonal. We, however, maintained its non-seasonal status in the study.

Climate, location and road infrastructure seems to play a complex role in the influenza seasonality and the spread dynamics. We aim not at ex- plaining all these factors, but rather at obtaining a practical result: a good characterization of the wave and consequent identification of the epidemic peaks across different latitudes. This characterization is rather important to improve the vaccine logistics in a continental country with limited resources. Today, Brazil follows a common vaccination calendar [11].

We restricted ourselves to a single definition of dynamic seasonality, using time averages. The flexibility of the wavelet analysis allows many other possibilities. The power spectrum offers a p-value for each observation in each frequency. For instance, Paireau et al. [43], consider the time series to be seasonal using a criterium of three consecutive years of significant p-values in a given frequency.

Another way our analysis could be further enriched is through the use of multimodal models in the circular context. We tested for a uniform distribution against a van Mises which is unimodal. A nested test for an increasing number of von Mises in a mixture model is feasible, but have not been implemented.

Little is known on why the seasonality of influenza varies with latitude [44]. Our results can be used to generate hypotheses about factors potentially involved in this unknown relationship. No single environmental factor has been shown to be the driver: in India, Chadha et al. [22] mention the rainfall; in China, temperature, hours of sunshine, and maximum rainfall [15] appears to have important relationships to the disease spread.

We conducted a dynamic time-series analysis to advance the understanding of seasonal influenza dynamics in tropical lands. In Brazil, only 12 of the 27 states were found to have annual-seasonal influenza activity. Among them, we identified that the peak activity starts in Ceará, the largest city of the Northeast region in 19^th week, traveling southbound by areas covered by roads with constant movement of people. The peak in the Rio Grande do Sul, the southernmost state of Brazil occurs in the 27^th week. Our results can be used to help the country vaccine logistics, identifying states where an earlier schedule is desirable, and others where some postponement will not harm. The span of the peak activity is more than two months apart, so the benefits of a noncommon vaccination calendar are not trivial.