Ecology and geography of hemorrhagic fever with renal syndrome in Changsha, China

Changsha is the capital city of Hunan Province in central China, between 27°51'– 28°40'N and 111°53'– 114°5'E (Figure 1). Changsha is about 233 km long and 90 km wide, with a total land area of 118,000 km² and population of 6 million. Changsha is composed of the alluvial plain of the Xiangjiang River, which is a part of the Central China Plain. The Xiangjiang and its branches flow throughout the city, and cultivated land is widely distributed along the rivers. Subtropical double-harvest rice is the main crop, and most farmers reside less than 50 m from their farmlands. Traditional farming methods provide an opportunity for wild rodent propagation, offering suitable living conditions and sufficient food resources to increase transmission of HFRS between rodents and from rodents to humans. Furthermore, HFRS rodent hosts are widely distributed in Changsha, including A. agrarius, R. norvegicus, M. musculus, Rattus flavipectus and Rattus rattoides. The infection rate of hantaviruses in some rodent species is relatively high; e.g., the rates of R. norvegicus and A. agrarius in 2005 were 3.17% and 2.80%, respectively. Meanwhile, a dense human population resulting from rapid urbanization in the Changsha-Zhuzhou-Xiangtan metropolis has increased interaction between people infected with hantavirus and those susceptible to infection. These factors provide an opportunity for HFRS to persist and spread in Changsha. Since 2005, Ningxiang County has been established as a national surveillance site to monitor the relationship between humans and HFRS host animals.

The study protocol was reviewed by the institutional review board of the Hunan CDC and ethics approval was not required.

Data on reported HFRS cases in Changsha from 2005 to 2009 (Figure 2) were obtained from the Hunan Notifiable Disease Surveillance System (HNDSS) [11]. There were a total of 327 case reports, which included the sex, age and residential address of each patient, plus onset date of symptoms for each case. The HNDSS HFRS data do not differentiate HTNV from SEOV infections. All cases were geocoded by residential address at town or village level using Google Earth. First, the city of Changsha was entered, and then the name of the county or district. Finally, a more specific address, the name of town or village, was added. To ensure accuracy of the geocoding results, Google Maps was also used in the same way.

Surveillance on hantavirus infections among rodent hosts was carried out in Ningxiang County, according to the protocol established by Chinese Centers for Disease Control and Prevention. Representative villages were selected for rodent surveillance, according to the distribution of HFRS epidemics and landscape elements. Indoor and outdoor rodent traps with peanuts were set at night, and recovered in the morning. The traps were placed about 1 km away from villages in locations where rodents were most likely found, such as the edges of rivers and roads, on ridges and in yards. Traps were set in spring (March to April) and autumn (September to October). More than 100 traps per patch were placed indoors at approximately 12–15-meter intervals for three consecutive nights, and more than 200 traps per patch were placed outdoors (every 5 meters in each row, with 50 meters between rows). A total of 22,060 traps were set, and 728 rodents were captured from 2005 to 2009. Relative population density of rodents was used as an indicator of abundance. This is calculated as the number of rodents captured, divided by the number of traps.

Data on eco-geographical variables (possibly contributing to hantavirus transmission among rodent hosts) were collected and processed in ArcGIS 9.3 (ESRI Inc., Redlands, CA, USA) (Table 1). Data for each variable were converted to the same geographic projection (Gauss-Kruger-Xian 1980) and clipped to the study area. Data on elevation, precipitation, temperature, precipitation seasonality and temperature seasonality were obtained from WorldClim (http://​www.​worldclim.​org/​). Slope and aspect values were calculated from elevation in ArcGIS 9.3. Land use data were obtained from the Second National Land Survey (http://​www.​mlr.​gov.​cn/​zt/​dierciquanguotud​idiaocha/​index.​htm). The compound topographic index (CTI, also called topographic wetness index), a function of both slope and upstream contributing area per unit width orthogonal to the flow direction, was obtained from the United States Geological Survey (http://​eros.​usgs.​gov). The Human Footprint Index was obtained from the Center for International Earth Science Information (http://​www.​ciesin.​columbia.​edu/​). Population in 2005 was obtained from the Changsha Statistical Yearbook, and population density calculated at the district and county level. The NDVI value from 2005 to 2009 was obtained from the National Aeronautics and Space Administration (http://​ladsweb.​nascom.​nasa.​gov/​data/​), and then average monthly values were calculated. All environmental variables were resampled to raster data with a spatial resolution of 0.00833° (nearly 1 km).

ENMs were developed to reflect the internal relationship between environmental variation and the distribution of HFRS. ENMs simulate and predict disease risks by exploring the relationship between the disease pattern and landscape elements such as land use type, temperature, precipitation, and elevation [9]. We used the Genetic Algorithm for Rule-set Production (GARP) for ENMs development. Occurrence points were divided randomly into three parts: extrinsic testing data (50%), used for model evaluation; training data (25%), used for model development; intrinsic testing data (25%), used for intrinsic evaluation and tuning of model rules. Distributional data were converted to raster layers. Then, 1,250 points randomly sampled from the training and intrinsic testing data and 1,250 points randomly sampled from the study region were used to calculate predictive accuracy (Table 2). GARP works iteratively for rule selection: a method is chosen from a set of possibilities (logistic regression, bioclimatic rules, atomic rules and range rules), and it is then applied to the training dataset to develop or evolve a rule. Rules evolve in ways that mimic DNA evolution (e.g., point mutations, deletions and crossing over). At each iteration, the change of predictive accuracy (Table 2) from one iteration to the next is used to evaluate a particular rule [12, 13]. The ecological niche modeling here was done using a desktop version of GARP, which is publicly available (http://​www.​nhm.​ku.​edu/​desktopgarp).

To build different subset models for the entire occurrence area, an algorithm threshold of 0.01 was selected, with 1,000 iterations as an upper limit for each replicate. Because of the stochastic nature of GARP in producing different outputs at different replicates, best practice approaches were required [14]. Ten best subsets were selected, with a threshold level of 0% extrinsic hard omission and 50% commission. GARP outputs are rasterized coverages of the study area. Two different pixel values were used to show the absence or presence of related species; 0 means absence (including areas in which no rules can be applied) and 1 means presence. Then, we calculated each pixel value by summing the "best subset models", using the Raster Calculator function in the Spatial Analyst extension of ArcGIS 9.3, producing final predictions of potential distributions with 11 thresholds (integers from 0 to 10). A pixel was predicted to be present if the threshold ≥ 5 of the 10 best subset models.

We developed a series of tests of model predictive ability. In each case, the developed models and predictions tested were based on independent suites of HFRS data. The HFRS data were divided into quadrants, above and below the median longitude and median latitude: (i) west versus east of the median longitude (hereafter “WE” tests); (ii) north versus south of the median latitude (hereafter “NS” tests); (iii) on-diagonal (upper left-hand and lower right-hand quadrants) and off-diagonal (upper right-hand and lower left-hand quadrants, hereafter “DIAG” tests). In each case, we developed both reciprocal predictions, testing the ability of ENMs to predict the distribution of HFRS in regions where no sampling was available. Finally, data from 2005 through 2008 were used to develop the ENMs, and 2009 data were used to validate the prediction (hereafter “Time” tests).

To investigate relative contributions of the 25 environmental variables, a jackknife procedure was used, which was performed by fitting the model with n-1 variables and successively omitting one variable. Then, we compared the predictive accuracy of the subset models with the comprehensive model (based on n variables), to eliminate variables that lead to over-fitting and to select key risk factors. Occurrence data from 2005 through 2008 were used to perform the jackknife procedure, and the average area under the receiver operating characteristic (ROC)[15] curve (AUC) was used to evaluate predictive accuracy of the model. After the jackknife procedure, the final HFRS prediction was executed again by GARP, using available variables.

The relative population density of rodents has important effects on HFRS occurrence. Rodent species composition varies with land use type. Since a certain land cover type is often associated with the presence of hantavirus-infected rodents, hantaviruses are usually associated with distinct rodent species. The annual proportion of HFRS cases was based on five main land use types (cultivated land, orchard, forest, urban and water). We calculated the annual density of various types of rodents (A. agrarius, R. norvegicus, M. musculus, R. flavipectus and other species) and annual total rodent density from 2005 through 2009 (Figure 3). Then, we used a Markov chain Monte Carlo (MCMC) method [16] to investigate the relationship between rodent density and HFRS incidence. Matrix R represents HFRS cases in different land use types; rows correspond to years, and columns to HFRS risks. Matrix C represents the density of different types of rodents; rows correspond to years, and columns to rodent density. Then, we calculated the coefficient matrix (Formula 1). Next, the relationship between annual HFRS cases and total rodent density was analyzed using Formula 2. Related environmental variables and rodent monitoring data were used to analyze potential associations between rodent species composition and HFRS incidence in different land use types. The MCMC method was carried out using Matlab software (vR2009b) (MathWorks Inc., Natick, MA, USA)[16].

where R _ij is the proportion of HFRS incidence in risk area j in year i; β is the coefficient matrix; C _iu is the proportion of relative population density of rodents u in year i; γ _k is the density of rodents in year k; ψ _k is the number of predicted HFRS cases.

All tests in this study found that independent sets of test points coincided with ENMs predictions, which were significantly better than random expectations (Table 3). Results of model testing showed that test points were well predicted (at threshold levels of 5–10) in all subsets, except those of the WE testing. In these tests (especially "east predicts west"), fewer test points were correctly predicted, with fewer areas predicted present (Table 3). However, results of all spatial calibration tests (Figure 4) correctly predicted the independent test points, as evident from the significant level of binomial probability (P< 0.01). Prediction of the overall study area at one time scale also implied a good model fit (Figure 4 and Table 3).

Results of the jackknife procedure showed that all environmental variables, except CTI and human population density, contributed to model predictive ability. Exclusion of temperature seasonality, land use, elevation and NDVI (mainly in November, January and June) resulted in the greatest deviations (Table 4).

Data from 2005 through 2008 were then used to develop the final prediction model, and 65 HFRS cases in 2009 were used to validate the prediction. Population density and CTI were excluded, and 10 best subset models were calculated for a prediction map (Figure 5). The results show that ENMs independently predicted 58 of 65 cases in high risk areas (45 cases were predicted by all 10 subset models) in 2009. Evaluation data were merged with 10,000 randomly selected background points and entered into a ROC analysis, to derive the AUC. The average AUC was 0.855 (95% CI: 0.821–0.886, SD = 0.017, P< 0.01), indicating that model performance was satisfactory.

The exploratory visualizations describe ecological characteristics of the predicted results of HFRS occurrence in Changsha, with the 10,000 randomly selected points in areas predicted absent versus present (Figure 6). The results show that HFRS occurrence was restricted to areas of strong temperature seasonality (Figure 6b), with mean annual temperature around 17.5°C (Figure 6a), precipitation below 1600 mm (Figure 6a), and elevation below 200 m (Figure 6c). Risk was concentrated in cultivated and urban land (Figure 6d). The risk level of HFRS is correlated with an increase in area of cultivated and urban land, and a decrease in forested areas (Figure 7). The NDVI value in areas predicted present is lower than in areas predicted absent, but with substantial seasonal variation (Figure 6e and Figure 6f).

Correlation analysis showed that the number of HFRS cases was remarkably associated with rodent density (r = 0.84, P = 0.07). The prediction model provided reasonably good fits of the relationship between rodent density and number of HFRS cases (Figure 8), in which θ ₁ is 10.81 and θ ₂ is 25.39. The coefficient matrix used to show the potential influence of rodent species composition on different areas was calculated as

The coefficient matrix implies that HFRS cases in urban areas and cultivated land are mainly caused by R. norvegicus and other species (including Rattus niviventer, R. flavipectus, Microtus fortis and shrew species, with total proportion less than 5%), and that risk in forested areas is from A. agrarius. In terms of population quantity, R. norvegicus and A. agrarius are considered the dominant species, with varying impacts on HFRS occurrence in different risk areas of Changsha.