Space-time clusters of dengue, chikungunya, and Zika cases in the city of Rio de Janeiro

Brazil is a dengue-endemic country where all four dengue virus serotypes circulate and cause seasonal epidemics. Recently, chikungunya and Zika viruses were also introduced. In Rio de Janeiro city, the three diseases co-circulated for the first time in 2015-2016, resulting in what is known as the ‘triple epidemic’. In this study, we identify space-time clusters of dengue, chikungunya, and Zika, to understand the dynamics and interaction between these simultaneously circulating arboviruses in a densely populated and heterogeneous city. We conducted a spatio-temporal analysis of weekly notified cases of the three diseases in Rio de Janeiro city (July 2015 – January 2017), georeferenced by 160 neighbourhoods, using Kulldorff’s scan statistic with discrete Poisson probability models. There were 26549, 13662, and 35905 notified cases of dengue, chikungunya, and Zika, respectively. The 17 dengue clusters and 15 Zika clusters were spread all over the city, while the 14 chikungunya clusters were more concentrated in the North and Downtown areas. Zika clusters persisted over a longer period of time. The multivariate scan statistic – used to analyse the three diseases simultaneously – detected 17 clusters, nine of which included all three diseases. This is the first study exploring space-time clustering of dengue, chikungunya, and Zika in an intraurban area. In general, the clusters did not coincide in time and space. This is probably the result of the competition between viruses for host resources, and of vector-control attitudes promoted by previous arbovirus outbreaks. The main affected area – the North region – is characterised by a combination of high population density and low human development index, highlighting the importance of targeting interventions in this area. Spatio-temporal scan statistics have the potential to direct interventions to high-risk locations in a timely manner and should be considered as part of the municipal surveillance routine as a tool to optimize prevention strategies. Author summary Dengue, an arboviral disease transmitted by Aedes mosquitoes, has been endemic in Brazil for decades, but vector-control strategies have not led to a significant reduction in the disease burden and were not sufficient to prevent chikungunya and Zika entry and establishment in the country. In Rio de Janeiro city, the first Zika and chikungunya epidemics were detected between 2015-2016, coinciding with a dengue epidemic. Understanding the behaviour of these diseases in a triple epidemic scenario is a necessary step for devising better interventions for prevention and outbreak response. We applied scan statistics analysis to detect spatio-temporal clustering for each disease separately and for all three simultaneously. In general, clusters were not detected in the same locations and time periods, possibly due to competition between viruses for host resources, and change in behaviour of the human population (e.g. intensified vector-control activities in response to increasing cases of a particular arbovirus). Neighbourhoods with high population density and social vulnerability should be considered as important targets for interventions. Particularly in the North region, where clusters of the three diseases exist and the first chikungunya cluster occurred. The use of space-time cluster detection can direct intensive interventions to high-risk locations in a timely manner.


Introduction
Dengue has been endemic in Brazil for more than 30 years. Since 2010, all four dengue virus (DENV) serotypes circulate in the country [1]. The first chikungunya and Zika outbreaks in Brazil were detected in 2014 and 2015, respectively, both in the Northeast region. In 2016, 1.5 million dengue cases, 270 thousand chikungunya cases, and more than 200 thousand Zika cases were notified in the country [2]. Initially described as a benign disease, Zika quickly became a serious public health problem after the association of the disease during pregnancy with congenital malformations, such as microcephaly, was discovered [3][4][5] The co-circulation of DENV, chikungunya virus (CHIKV) and Zika virus (ZIKV), poses a serious public health and economic burden [6,7]. The Brazilian government has implemented dengue prevention and control measures in the form of vector-control interventions, but there is no evidence that vector-control has had a significant effect in reducing transmission in Brazil or other parts of the world [8]. The widespread presence of the vector (mainly Aedes aegypti but also Aedes albopictus), a highly mobile population, and low or lack of herd immunity resulted in simultaneous and overlapping outbreaks of all three diseases, a phenomenon that has been referred to as the 'triple epidemic'. Understanding the behaviour of dengue, Zika, and chikungunya, when they compete in time and space, is a step forward in improving the design of interventions for prevention and outbreak response [9]. Zika was monitored through sentinel surveillance [10,11]. Most notifications are made by physicians working in public health facilities, based on diagnostic protocols by the Ministry of Health. SINAN receives a large number of notifications and it thought to accurately represent the overall trend of the dengue situation in Brazil [12].
Considering DENV, CHIKV, and ZIKV share the same vectors and human hosts, we conducted a spatio-temporal analysis of notified cases to identify clusters and understand the dynamics of these diseases in a scenario of triple epidemics. Rio de Janeiro was the chosen city for this analysis for the following reasons: a history of large dengue epidemics with sustained transmission; the recent occurrence of CHIKV and ZIKV epidemics in 2015-2016; co-circulation of DENV, CHIKV and ZIKV; a high number of reported cases; the possibility to work with georeferenced cases in an intra-urban context; multiple environmental settings within the city; high

Study site
Rio de Janeiro is the second largest city in Brazil, with approximately 6,3 million inhabitants (2010 census), 1204 km² and 160 neighbourhoods (Fig 1). The city has the 45 th highest Human Development Index (HDI) of the country, of 0.799 (varying from 0.604 to 0.959 inside the city) [13,14]. The population density is 5249 inhabitants per km². Rio de Janeiro has a tropical climate, with temperature and rainfall varying depending on altitude, vegetation and ocean proximity. The average annual temperature is 23.7°C, and the annual accumulated precipitation is 1069 mm.
During the summer months (December to March), high temperatures (around 40ºC) and thunderstorms are common [15].
The 160 neighbourhoods are grouped into four large regions (North, South, Downtown and West), reflecting the geographical position and history of occupation. Almost all neighbourhoods are a mixture of very poor slums ("favelas") and more affluent areas of residence. The North region is very urbanized, with high population density, few green areas and very large favelas. Nearly 27% of the population of this region, almost 2.4 million people, lived in favelas in the 2010 demographic census [16]. The South region is the most popular tourist destination in Rio de Janeiro, with famous beaches, green areas, and neighbourhoods with the highest HDI of the city [13]. The Downtown region is the historical, commercial and financial center of the city, with many green areas and cultural centers. Finally, the West region has been urbanized and populated more recently, and is less densely populated [15].

Space-time analysis
For spatio-temporal detection of clusters, Kulldorff's scan statistic with a discrete Poisson probability model was applied for each disease individually and for the three diseases simultaneously (multivariate scan statistic with multiple data sets). The scan statistic uses moving cylinders across space (i.e. the base of the cylinder) and time (i.e. the height of the cylinder) to identify clusters, by comparing the observed number of cases inside the cylinder to the expected number of cases [17,18]. The detected clusters are ordered in the results section according to the likelihood ratio, such that the cluster with the maximum likelihood ratio is the most likely cluster, that is, the cluster least likely to be due to chance. The relative risk for each cluster is calculated as the observed number of cases within the cluster divided by the expected number of cases within the cluster, divided by the observed number of cases outside the cluster divided by the expected number of cases outside the cluster [19].
The multivariate scan statistic for multiple data sets was applied to simultaneously search for clusters of dengue, Zika and chikungunya that coincided in time and space. This calculates for each window the log likelihood ratio for each disease. Then, the likelihood for a particular window is calculated as the sum of the log likelihood ratios for the diseases with more than the expected number of cases. In the same way as for a single disease, the maximum of all the summed log likelihood ratios constitutes the most likely cluster [19,20].
For each model, Monte Carlo simulations (n=999) were performed to assess statistical significance. We considered statistically significant clusters (p-value < 0.05) that did not coincide in space (with no geographical overlap) and that included a maximum of 50% of the population of the city (nearly 3,1 million people). With only these parameters, two large clusters covering most of the city were detected (S1 Fig A), which is not useful if we are interest in identifying risk areas to direct interventions. After testing several combinations of temporal and spatial parameters (such as the size of the temporal window and maximum population at risk inside the cluster), we chose the combination that resulted in a reasonable number of clusters that aggregated close together and in similar locations that could also be targeted for local interventions (S1 Fig). The temporal window was set to be at least 1 week and a maximum of 4 weeks. Clusters were restricted to have at least 5 cases. In the output parameters, clusters were restricted to include a maximum of 5% of the population of the city (nearly 315 thousand people). SaTScan™ (version 9.5, https://www.satscan.org/) software was applied within R (version 3.4.4, https://www.r-project.org/), using the package rsatscan (version 0.3.9200) [21][22][23]. Maps were produced using the ggplot2 (version 3.1.0) package in R [24].

Dengue cases clusters
Scan statistics detected 17 dengue cases clusters (Table 2). Clusters were detected in different parts of the city (Fig 3A). The most likely cluster was located in the North zone of Rio de Janeiro city. Cluster 2 contained only one neighbourhood in the Downtown area with a relative risk of 172.67 (S2 Fig A). Clusters were detected within a short time period, from March to May 2016, except for cluster 15 that started in December 2015 (Fig 3B). The first dengue cluster in time was detected in the West zone (S3 Fig A).

Chikungunya cases clusters
For chikungunya, 14 clusters were detected (Table 3) of this region (Fig 4A, clusters 6, 9 and 13). The most likely cluster was located in the Downtown of Rio de Janeiro city and had the highest relative risk (S2 Fig B). Clusters were also detected within a restricted time period, between 27 March and 11 June (Fig 4B). The first chikungunya cluster in time occurred in the northern border of the city (S3 Fig B).  There were 15 Zika clusters, distributed all over the city, similar to the observed pattern for dengue ( Fig 5A, Table 4). The most likely cluster was located in the West of Rio de Janeiro city, a region where chikungunya clusters were rarely observed. This cluster also had the highest relative risk (S2 Fig C). In contrast to dengue and chikungunya, Zika clusters occurred over a longer period of time, between December 2015 and May 2016 (Fig 5B). The third most likely cluster occurred 8 weeks after the first one. The first Zika clusters in time emerged in the North of the city (S3 Fig C).

Dengue, chikungunya, and Zika multivariate clusters
The multivariate scan statistic for multiple data sets detected 17 clusters, of which nine showed dengue, chikungunya, and Zika occurring simultaneously; five showed overlapping dengue and Zika outbreaks; and three showed only outbreaks of Zika (Table 5, Fig 6). The most likely cluster was found in the Downtown region of the city.
Of the 160 neighbourhoods assessed, 57 (35,6%) had clusters for the three diseases coinciding in time and space. Of the nine simultaneous clusters, five were located in the North of the city, three in the West, and one in the Downtown.  260 displacement of Zika caused by chikungunya [27]. For Rio de Janeiro city, this might not be the case, as CHIKV caused only a few cases at beginning of 2016, and only started to rise when Zika cases decreased (the depletion of susceptible hosts). Therefore, we hypothesise that ZIKV circulation inhibited CHIKV, rather than CHIKV introduction displacing ZIKV.
Scan analysis successfully identified clusters of dengue, chikungunya, and Zika. The most likely cluster for each disease occurred in a different part of the city (North, Downtown, and West, respectively). Unlike for dengue and Zika, chikungunya clusters were rarely detected in the West of Rio de Janeiro, probably because the rainy and warm season ended before the disease could reach this region with a sufficient transmission rate to form clusters.
Zika clusters were detected over a longer period of time compared to dengue and chikungunya clusters. We hypothesise that this is a result of the ZIKV advantage in competing for Ae. aegypti mosquitoes: the Ae. aegypti has been described as a more efficient vector for ZIKV transmission than for DENV or CHIKV, even when co-infected [28,29]. Not only does Ae. aegypti transmit ZIKV at a higher rate, but it is also more easily infected by ZIKV compared to DENV and CHIKV. CHIKV, on the other hand, replicates better than ZIKV in Ae. albopictus cells [28]. While Ae. aegypti is highly adapted in urban settings, living preferably in domestic and peridomestic areas, Ae. albopictus prefers to live in areas with more vegetation. However, Ae. albopictus was recently identified distant from green areas in a densely urbanized complex of favelas in Rio de Janeiro, suggesting this species is adapting to anthropic environments [30]. Further studies are needed to understand the importance of Ae. albopictus in CHIKV transmission.
A previous study suggested that a Zika epidemic would prevent a subsequent dengue epidemic, as a consequence of cross-immunity [31]. Like DENV, ZIKV is a flavivirus, and the structural similarity between them results in cross-immunity. [32] Whether this cross-immunity leads to antibody-dependent enhancement (ADE, that results in more severe forms of the disease), protection, or neither, is still uncertain [33][34][35]. In our study, the number of dengue cases increased Dengue, chikungunya, and Zika clusters detected in Rio de Janeiro do not usually coincided in time and space, contrasting with a study in Mexico that found strong spatio-temporal coherence in the distribution of the three diseases [9]. In addition to virus interactions and competition for the resources for replication inside the vector, behaviour changes may also impact disease dynamics. A rise in the number of cases may promote vector-control activities, which in turn may decrease the number of cases and hinder the establishment of another arbovirus [36]. Also, wealthier areas may have better vector-control interventions, resulting in different spatial distributions.
Neighbourhoods in the North of the city were more likely to have simultaneous clusters of dengue, Zika and chikungunya, highlighted these areas as priority targets for interventions. This is especially important considering co-infections are possible and clinical outcomes are not clear for such cases [37]. As dengue has been endemic in Rio de Janeiro for the last three decades and Janeiro has already been identified as a hot spot for dengue and as a key region for dengue diffusion. Previous studies also identified Catumbi, a neighbourhood in the Downtown area, as a high-risk location for dengue [38,39]. In our findings, Catumbi comprised the most likely chikungunya cluster, the second most likely cluster for dengue and the third most likely for Zika.
Additionally, the clusters in Catumbi coincided in time (most likely cluster in the multivariate scan analysis). Further investigations should be conducted to understand why this neighbourhood in particular is a high-risk location for arboviruses.
The North of the city is marked by a combination of high population density and a lower HDI than the city average [13]. The high population density facilitates the mosquito-human contact and hence the chance of becoming infected. The link between poverty and arbovirus is controversial [40]. Nonetheless, locations with social and economic vulnerability more likely have poorer sanitary conditions and less efficient vector-control interventions, which would facilitate mosquito proliferation. In Rio de Janeiro city, areas in or near favelas were detected as hot spots for dengue [39]. Consistent with our findings, a study conducted in French Guiana indicated that, early in the epidemic, the poorest neighbourhoods would have a greater risk for CHIKV infection [41]. In the first dengue epidemic in a city of São Paulo state, Brazil, authors found a direct relationship between low socio-economic conditions and dengue [42]. We did not observe this relationship for dengue possibly because dengue has already had sustained transmission in the city for decades.
Some limitations affect this study. As our study population included only notified cases (i.e. only patients who sought medical care), asymptomatic cases were not captured. Mild cases usually are poorly captured by SINAN, but considering the disease awareness around Zika, people (especially women) were expected to be more concerned about seeking medical care in case of suspected Zika. As Zika, dengue and chikungunya share some symptoms, the disease awareness manifestations of dengue, Zika, and chikungunya also represent a limitation. This limitation is inherent of every study using notified cases, as only a small proportion of cases are laboratory confirmed. However, if misdiagnosis was common, we would not expect to detect differences in time and space of occurrences. In addition, the extensive experience of health care professionals working in Rio de Janeiro, in detecting and diagnosing dengue symptoms, is thought to reduce the probability of misdiagnosis.
A small percentage of cases (8%) that were not georeferenced (and hence, not included in this study) could potentially result in a selection bias. It is possible that cases occurring in favelas, where addresses are sometimes not standardized, have a higher chance of not being georeferenced.
Clustering was based on the neighbourhood of residence only, yet infection can happen at other places, such as the workplace. Scan analysis was not designed to understand diseases trajectory but are still helpful to help plan interventions. Also, the method detects circular clusters only, rather than clusters of irregular shapes.