Modeling the relative role of human mobility, land-use and climate factors on dengue outbreak emergence in Sri Lanka

Dengue is a mosquito-borne viral disease that infects approximately 390 million people globally every year, particularly in tropical and subtropical countries [1, 2]. In Southeast Asia, dengue has become the leading cause of febrile illness for travelers [3]. The high number of infections combined with the lack, as yet, of a routinely used effective vaccine has made dengue a notorious public health problem [2, 4].

Dengue spreads through the bite of infected Ades mosquitoes, especially Aedes aegypti– the primary vector, with an estimated 15 to 17-day delay between the primary and secondary human infections [5]. Dengue outbreak control is a challenge for policy makers because Aedes aegypti mosquitoes are well adapted to high density urban environments and actively feed during the day [6‐8], thus presenting an elevated risk to humans. Urban settings provide an ideal habitat for Aedes aegypti breeding due to an abundance of discarded trash bags, plastic bottles, tires, and other containers that enable the formation of stagnant shallow water surfaces after precipitation [9]. Urban regions in developing countries are particularly vulnerable due to a lack of indoor plumbing infrastructure that, in conjunction with a lack of air-conditioning, results in higher human-mosquito exposure rates during the day. Additionally, because of the daytime feeding behaviors of Aedes aegypti, common vector control measures that work for night-biting mosquitoes, such as bed nets, fail to effectively control dengue transmission. Given these challenges, there is a need to better understand and predict dengue outbreaks and transmission risk within urban regions in developing countries so that vector control and surveillance resources can be optimally allocated.

The majority of dengue risk maps that identify dengue hotspots are static. Although they help reveal overall potential high-risk areas, challenges remain in their limited ability to provide temporally evolved risk maps and to support dengue early warning systems [10]. Therefore, explorations of both spatially and temporally varying factors in dengue risk dynamics are needed. Previous studies highlighted human mobility as a critical factor for dengue transmission [11‐17], which contrasts the more minor role travel plays in the spread of vector-borne diseases transmitted by night-biting mosquitoes [17]. While Aedes aegypti mosquitoes have a hard time dispersing geographically across large areas because they rarely travel more than 400 m from where they emerge as adults [18‐21], humans regularly travel much longer distances on a daily basis. As new dengue cases and clusters are regularly reported kilometers apart, it is likely that human mobility play a critical role in the spread of dengue outbreaks, i.e., infected humans introduce dengue into new mosquito populations at their trip ends. As an example, Vazquez-Prokopec, Montgomery [14] studied the pattern of dengue transmission using location-based contact tracing on infected dengue patients during a dengue outbreak centered at Cairns, Australia. They collected locations that the patients frequently traveled to during the daytime and 2–4 weeks prior to the onset of symptoms through phone interviews. The contact locations with a proximity of 100 m and a separation of 20 days were spatial-temporally linked into pairs and then chains to identify the plausible sites of dengue virus transmission. They showed that the complex pattern of dengue transmission was primarily driven by human mobility, and that targeted residual spaying could potentially reduce the probability of dengue transmission up to 96%. Their study highlights the importance of understanding dengue transmission patterns to optimize the allocation of dengue prevention and vector-control measures.

In addition to human mobility, recent studies have pointed to a strong association between climate conditions and dengue outbreaks at various locations and across different temporal resolutions [9, 22‐26]. Precipitation, mean temperature and temperature fluctuation were revealed to affect the population dynamics of Aedes aegypti mosquitoes and the dengue virus extrinsic incubation period [27‐31]. Specifically, a suitable average temperature and moderate temperature fluctuations are often favorable for dengue transmission [27], while an increase in precipitation is strongly associated with the onset of a dengue outbreak [24]. Humidity, a combined effect of precipitation and temperature, is also a common climate index to evaluate the environmental capacity for dengue emergence [22, 23, 32, 33]. Wesolowski, Qureshi [15] accounted for both climate and mobility in a study of dengue virus transmission over a large dengue outbreak period in Pakistan. They developed an epidemiological model that included temperature and relative humidity as input parameters for mosquito dynamics, as well as biting rate to capture the interactions between human and mosquito hosts. Human mobility was captured using mobile phone data of ~ 40 million subscribers to estimate the spatially explicit travel volume, albeit not differentiating infected and non-infected people. They showed that the emergence of dengue epidemics in a new region could be predicted using aggregated travel patterns from endemic areas in combination with the developed epidemiological model. While climatic factors were found to be significant for prediction, this was in part due to the large study region, i.e., country level, which has variable climatic suitability for the mosquito vector. The study region considered in our work is much smaller and has minimal climactic variability, thus alternative methods are required to distinguish site-specific risk.

Land-use patterns — indicators of human activities and potential breeding habitats — have also been linked to dengue outbreaks [34‐39]. Previous studies investigated the effect of land-use patterns on the spread of dengue and found that human settlements, water bodies, and mixed horticulture are the top three associated land-use patterns for dengue emergence in Malaysia [37]. In another study [38], areas surrounded by rice paddies and marshes/swamps were associated with a significantly higher population of dengue vectors during the rainy season in Thailand. Orchards (which often contain artificial water containers) and irrigation fields have also been shown to play an important role in dengue infections; however, their role varies given different local conditions. Sometimes, land-use type can be a proxy for other features, such as socio-economic factors, which may have a contradictory effect on dengue infections [39].

In this study, we present a statistical modeling framework to evaluate the relative role of human travel patterns, climate conditions, and land-use patterns on dengue outbreak dynamics in Negombo, Sri Lanka (Fig. 1), where dengue outbreaks are increasing in scale and prevalence. Dengue has been endemic in Sri Lanka since 1960s, and all four serotypes circulate in the region [41]. Dengue hemorrhagic fever (DHF) cases were rare until 1989, when the first major dengue outbreak occurred with approximately 200 cases and 20 deaths; DHF has since become endemic in the country [41‐43]. The 2017 outbreak involved many severe DHF cases, with evidence showing a dengue virus type 2 (DENV-2) being the causative serotype [44, 45]. DENV-2 had only be detected infrequently over the recent decades of dengue epidemics and is therefore associated with low immunity in the region [41]. As a result, more than 80,000 dengue cases including 215 deaths were reported nationally in less than 7 months between 2016 and 2017, a fourfold increase in the number of reported cases compared to the average number over 2010–2016 [46]. The region of Negombo, located in the Western province, experienced the greatest number of dengue cases in the country; approximately 45% of the cases nationwide by July 2017 (Fig. 1), and is the focus area of our study.

We applied a mixed-effects model, where the mobility data bridges the time-varying, spatially-invariant climate variables and the rasterized spatially explicit, time-invariant population and land-use variables, to capture the spatial-temporal dynamics of dengue transmission. Our model framework differs from previous studies that simulated the transmission process [14, 15, 47], and instead focuses on estimating the timing and location of new case introductions though a non-process based statistical model. Specifically, we focus on modeling the home locations of (newly infected) dengue patients, and assume dengue risk is increased in areas by infected individuals who travel to the area, providing the opportunity for vector transmission to result in new human cases. This assumption is consistent with previous studies that have shown visits to a household by infected people determines the infection risk in that household [13]. In addition, the study was conducted at a fine-grained spatial and temporal resolution — 1 km × 1 km spatially and 1 week temporally — providing an improved understanding of the role of mobility in the spread of dengue. While previous work studied the impact of mobility [14, 15], climate [9, 22, 24, 27], and land-use [37‐39] separately on dengue, the authors are unaware of any existing study that considers these factors within a single integrated framework. Thus, previous studies have been unable to quantify the relative contribution of each factor on the spatial-temporal patterns of dengue transmission as we do. The results from our study indicate that mobility is a more significant indicator of new dengue case clusters compared with land-use and climate factors. Furthermore, the case study in Sri Lanka provides critical insights into effective application of dengue prevention and vector control measures in developing regions.

A statistical model is applied to investigate the spatial-temporal dynamics of dengue outbreak with respect to a range of potential explanatory variables. The complete set of potential explanatory variables is listed in Table 1. Detailed descriptions of the data followed by a description of the methodology are provided below.

The results from this study illustrate the critical contribution of human mobility on the location and timing of new dengue cases, relative to land-use and climate variables. The results are sensitive to the patterns of travel during the week immediately preceding the appearance of new case reports. This was the variable with the greatest predictive power. Although, travel patterns 2 weeks prior were found to have an insignificant effect on dengue outbreak dynamics. Our results are consistent with Stoddard, Forshey [13], who concluded that visits to households by dengue infected individuals determines the infection risk, further validating our use of patient home locations in the model. Furthermore, the significance of mobility in outbreak prediction was found to be robust under both spatial resolutions of 1 km × 1 km and 5 km × 5 km.

In contrast to the role of mobility, which we found to be a consistently significant indicator of new dengue cases, the effect of land-use patterns on the number of new cases is sensitive to the spatial resolution of the models. Land-use variables played a larger explanatory role at the coarser spatial resolution of 5 km × 5 km (compared to the finer 1 km × 1 km resolution), particularly for smaller spatially-dominant land-use patterns such as Sea and StWtr. BuiltUp showed the strongest positive effect overall, indicating urbanization is associated with an increased risk of dengue outbreak (which is consistent with multiple previous findings [54, 55]). StWtr also showed significant positive effect, which is to be expected because standing water provides suitable mosquito breeding habitat [21]. The positive effect of Sea only appeared significant when human mobility was excluded from the model. Given the significant positive correlation between Sea and the number of trips (Table S4), it is likely that the large travel volume towards the area near the coastline makes the study region prone to dengue outbreaks. Previous studies have also found evidence that dengue mosquitoes can breed in brackish water [56], further supporting our finding of the distance to the sea as significant in some of the models. If this pattern holds in other regions, as seems likely, that fact can be used for the spatial prioritization of resource allocation for disease case and vector surveillance and control.

Among the climate factors, temperature-related variables including Tavg, Tmin, and DTR, were more strongly associated with the outbreak emergence than precipitation-related variables including Pre and RD, or RH, which is related to both. This finding is in accordance with [24], which concluded that “rainfall strongly modulates the timing of dengue (e.g., epidemics occurred earlier during rainy years) while temperature modulates the annual number of dengue fever cases.” Based on regression analysis, we found Tmin with a 10-day lead time to be the best climate-based predictor of new weekly dengue cases. The conclusions on climate influences may be different for other regions, for example, with longer lead time for the best climate predictor or more significant effects of precipitation on dengue risks [57]. However, given that our model is readily applicable, the model results on climate factors can be used for the temporal prioritization of resources in broader regions.

In addition to human mobility, climate, and land-use variables, which were included as fixed effects, population density was incorporated using random effects in the model because population is likely to have spatially heterogeneous effects on dengue outbreaks, as noted in the Methods. Based on the model results, the random-effects coefficients for population are mostly positive, as expected, indicating that higher population density is associated with a higher number of dengue cases (Figure S6). This is consistent with previous studies that have shown the negative impact of urbanization and population growth on dengue risk [58]. The most significant positive effect is seen north of the lagoon along the coastline, highlighting potentially high-risk areas, where higher populations are likely to facilitate the emergence of dengue outbreaks. A few cells resulted in negative random-effects coefficients, which may be due to confounding interactions between different variables included in the model, or alternative factors not captured in the model; these cells were few and only occurred in the model when the dominant mobility variable was included. It is possible the dominant role of mobility could over compensate for the impact of population, e.g., because people are likely to travel to crowded downtown areas, along the lagoon, or near the ocean where the large number of trips made to those regions could offset the impact of population. That the random-effects coefficients for population density are positive and negative lends support to the modeling decision to treat it as having enough stochasticity to qualify as a random-effects variable.

The results of this analysis have implications that are relevant to the design of measures to control dengue cases, such as allocation of resources for mosquito vector control. Previous global modeling of ecological suitability for dengue vector mosquito species (both Aedes aegypti and Aedes albopictus) have shown that the entire study area is a prime habitat for these species [59, 60]. This conclusion drawn from the global models finds validation in our analysis, which shows that climate and land-use variables are not the most strongly associated with dengue case outbreaks. Consequently, epidemiological risk based on vector ecology may be insufficient for the purposes of optimizing vector control resource allocation, as it is unable to distinguish between potential sites to target within the study area. Because travel into the sites is the most important predictor of new case clusters, it may well be time to optimize vector control resources based on mobility data, with the aim to prevent exposure to the day-biting mosquitoes, i.e., Aedes aegypti, at the highest risk locations. To the best of our knowledge, such a design for dengue control measures has not yet been tried in the field.

Finally, various limitations of this study should be noted. First, only dengue patients admitted to the CCMDDHF at the Negombo Hospital were included and surveyed in this study. Thus mild or asymptomatic cases, which account for the majority of dengue cases [1], were not accounted for in the study. Second, some patients infected in the study region may have gone to hospitals in other districts and would therefore not be included. Third, the mobility data was based on patients’ recollections over a 10-day period prior to hospital admission, and therefore may have inaccuracies due to human error in recalling the information. However, detailed analysis of the travel data revealed most trips recorded represent daily commuting routines. Thus, while some trips may be excluded due to human error, we believe the relative connectivity between cells is accurately captured by the survey responses. In addition, the sensitivity analysis revealed our results are robust to variation in the time of outbreak and trips variable. Fourth, the distance traveled and the time spent in a certain location were not considered due to the unavailability of relevant data, which is a limitation in most studies due to the lack of tools to precisely monitor the time-dependent location of individuals [61]. However, these factors have been shown to have little influence on dengue transmission [13, 16], and thus their exclusion does not invalidate the methodology used in this analysis. Fifth, by utilizing all the mobility data collected, we made an implicit assumption that the patients were infectious during the entire 10-day period prior to hospital admittance. This period does fall within the combined intrinsic incubation period (4–10 days) [62] and the early symptomatic period before admitted to the hospital. A sixth assumption was that the patients were infected at or around their home locations. This assumption is consistent with a wide variety of previous studies that revealed homes as the primary point of contact for dengue transmission [13, 63, 64]. Vazquez-Prokopec, Montgomery [14] tried to identify the most plausible transmission locations based on reported contact locations from a dengue outbreak in Cairns, Australia and found that only 10.2% of the identified transmission sites were at out-of-home locations, and a notable portion of them were within 1 km of the home locations. Given that our objective was not to model the transmission chains of dengue as in [14], assuming home locations as the site of infection provides reasonable support for predicting where infected individuals reside, and therefore the risk posed around homes of infected individuals. Lastly, the climate data were obtained from a single station, thus a homogenous climatic region was assumed for our study region. Therefore, the role of climate factors on the dengue outbreak may be underestimated.

While the modeling framework used here is readily applicable to other contexts, future work should investigate how widely transferable the model results are. More specifically using general mobility data (tracking movements for all residents); such as using mobile phone data as in [15], or transport planning data, which may be more readily available and cost effective; should be compared to the use of patient mobility surveys as in this study. Information on the attributes of the surveyed subjects such as ages and occupations may be worth exploration when available, as people in different age groups or occupations may report different movement patterns, such as the frequency of contacts per trip, which could therefore affect their associated risk of dengue transmission. Control measures that target specific sites within a cell or a land use type should be considered based on a combination of our study and other site-specific studies such as [65, 66]. For example, previous studies show that schools were associated with the highest ratio of positive larvae breeding sites over potential breeding sites, in contrast to residential premises at the lower end, based on an entomological survey in a subdistrict of the city of Colombo within our study region [66]. Such sites at higher risk should be specifically targeted after spatial and temporal prioritization of resources in vector control programs.

Our study highlights the potential value of using travel data to target vector control within a region. In addition to illustrating the relative relationship between various potential risk factors for dengue outbreaks, the results of our study can be used to inform where and when new cases of dengue are likely to occur within a region, and thus help more effectively and innovatively, plan for disease surveillance and vector control.