Spatiotemporal Heterogeneity Analysis of Haemorrhagic Fever with Renal Syndrome in Anqiu City, China, 2000-2014

Background: The purpose of this study was to explore the dynamics of the occurrence of haemorrhagic fever with renal syndrome (HFRS) and find the potential spatiotemporal factors leading to the incidence of HFRS in Anqiu City. Methods: Monthly reported cases of HFRS and climatic data for 2000–2014 in Anqiu City were obtained. An autoregressive integrated moving average (ARIMA) model was used to fit the HFRS incidence prediction model and predict the epidemic trend in Anqiu City. Multiple linear regression method was used to analyze the temporal relationship between HFRS incidence and meteorological factors during the study period. Results: Spatial analysis results indicated that the annualized average incidence at the town level ranged from 2.18 to 6.09 per 100, 000 among 14 towns, and the western towns in Anqiu City exhibited high endemic levels during the study periods. With high validity, the optimal ARIMA (0, 1, 1) × (0, 1, 1)12 model could be used to predict the HFRS incidence in Anqiu City in 2014. The monthly trend in HFRS incidence was negatively associated with temperature and precipitation and positively associated with average wind speed. Multiple linear regression models showed that precipitation and relative wind speed were key climatic factors contributing to the transmission of HFRS. Conclusions: This study provides evidence that the ARIMA model can be used to fit the fluctuations in HFRS frequency in Anqiu City. Our findings add to the knowledge of the role played by climate factors in HFRS transmission in Anqiu City and can assist local health authorities in the development/refinement of a better strategy to prevent HFRS transmission. HFRS: Haemorrhagic fever with renal syndrome; CDC: Center for Disease Control and Prevention; GLM: Generalized linear model; ARIMA: Autoregressive integrated moving average; GAM: Generalized additive model; NARNN: Nonlinear autoregressive neural network; ADF: Augmented Dickey-Fuller; ACF: Autocorrelation functions; PACF: Partial autocorrelation functions; AIC: Akaike Information Criterion; BIC: Bayesian Information Criterion; MAE: Mean absolute error; MAPE: Mean absolute percentage error; MSE: Mean square error; Root mean error.

Background: The purpose of this study was to explore the dynamics of the occurrence of haemorrhagic fever with renal syndrome (HFRS) and find the potential spatiotemporal factors leading to the incidence of HFRS in Anqiu City. Methods: Monthly reported cases of HFRS and climatic data for 2000-2014 in Anqiu City were obtained. An autoregressive integrated moving average (ARIMA) model was used to fit the HFRS incidence prediction model and predict the epidemic trend in Anqiu City. Multiple linear regression method was used to analyze the temporal relationship between HFRS incidence and meteorological factors during the study period. Results: Spatial analysis results indicated that the annualized average incidence at the town level ranged from 2.18 to 6.09 per 100, 000 among 14 towns, and the western towns in Anqiu City exhibited high endemic levels during the study periods. With high validity, the optimal ARIMA (0, 1, 1) × (0, 1, 1)12 model could be used to predict the HFRS incidence in Anqiu City in 2014. The monthly trend in HFRS incidence was negatively associated with temperature and precipitation and positively associated with average wind speed. Multiple linear regression models showed that precipitation and relative wind speed were key climatic factors contributing to the  Some studies predicted HFRS epidemics using ARIMA models and obtained a basis for targeted prevention and control measures [ 15 , 16 ]. These studies showed that ARIMA model had better predictive performance, but there was still some inconsistency between models and/or regions, which make researchers difficult to choose the appropriate one to predict HFRS epidemic. This inconsistency may be due to the fact that there are many influencing factors, such as immunization, temperature, humidity, elevation, and the local rat species [ 5 22 ]. Therefore, the HFRS prediction model constructed from a particular region is not universal. To predict the epidemics of HFRS in a region accurately, a specific prediction model based on the actual data of the region need to be constructed.
Previous studies indicated that some areas in Shandong Province were moderate endemic areas with HFRS incidences between 5.0/100,000 and 30.0/100,000 from 1994 to 1998.
However, few studies have been conducted to explore the dynamics of HFRS occurrence and determine the potential spatiotemporal factors leading to this disease in Anqiu City.
In the present study, the spatiotemporal distribution patterns of HFRS cases were explored, the key climatic drivers of HFRS transmission were identified, and the optimal ARIMA model for predicting HFRS incidences was developed for Anqiu City. The results of this study can help predict the future trends of HFRS, which can be used to more accurately prevent and control HFRS in Anqiu City.

Study area
The study area covers Anqiu City in Shandong Province, which is located in the middle of the Shandong Peninsula. Anqiu City is located between latitudes 27°51' and 28°40' north and longitudes 111°53' and 14°5' east. Anqiu City is 217 km wide and 202 km long, with a total land area of 1760 km 2 . Anqiu City has a warm temperate continental climate influenced by the monsoon. The annual mean temperature in the study area is 12.

Spatiotemporal analysis of HFRS case characteristics
The spatiotemporal distribution characteristics, including the temporal and spatial distribution of HFRS cases from 2000 to 2014 in Anqiu City, were analysed according to the surveillance data from the infectious disease monitoring system. All HFRS cases were coded by administrative code using ArcGIS10.2 (ESRI Inc., Redlands, CA, USA),which were matched to the town-level polygon and point layers.

Temporal trend analysis
The ARIMA model is one of the most commonly applied time series prediction model.
ARIMA was designed to address highly seasonal data. There was a strong seasonality trend in this study, and we constructed ARIMA models for monthly HFRS incidences in Anqiu City from 2000 to 2014. The ARIMA model was defined as the number of autoregressive lags p, moving-average lags q and number of different passes d. The multiplicative seasonal ARIMA (p, d, q) × (P, D, Q)s model has apparent seasonal variation characteristics ,which is an extended ARIMA model. Similar to ARIMA models, the seasonal parameters include seasonal autoregressive lags P, seasonal moving-average lags Q, seasonal differences D, and the length of the seasonal periods. The Augmented Dickey-Fuller Unit Root (ADF) test was applied to estimate the stationarity of the time series. If the time series is not stationary, an appropriate difference can be used to make the series stationary. The Box and Jenkins strategy was used to construct the seasonal ARIMA model in this study [ 23 ]. The ARIMA model analysis includes three main iterative steps: model identification, parameter estimation and the model evaluation. The autocorrelation functions (ACFs) and partial autocorrelation functions (PACFs) of the transformed data were utilized to determine the seasonal and non-seasonal orders and identify an appropriate ARIMA model. The conditional least squares method was applied to estimate the model parameters. In model diagnosis, white-noise-test methods were employed to check whether the residuals were independent and normally distributed. Several models may be constructed, and the selection of an optimal ARIMA model is necessary, which is usually based on the Akaike Information Criterion (AIC), normalized Bayesian Information Criterion (BIC) and Ljung-Box Q test. In addition, the mean absolute error (MAE), mean absolute percentage error (MAPE) and mean square error (MSE) were selected as the measures to evaluate the ARIMA model. The root mean square error (RMSE) was also used to access the accuracy of the models. All of these analyses were conducted using SPSS (version 17.0, SPSS, Chicago, IL, USA).

Correlation analysis between meteorological factors and HFRS incidence
Bivariate linear analysis and multiple linear analysis were used to model the relationship between meteorological factors and monthly HFRS incidence. The meteorological factors utilized in this study mainly included average temperature, monthly precipitation, average wind speed, and relative humidity. The monthly HFRS cases is shown in Figure 1, which indicates that the occurrence of HFRS presented apparent seasonal character. There were two high peaks every year, the smaller epidemic occurred in spring between March and April and the larger one occurred in winter between October and November.

Spatial distribution of HFRS incidence
To account for the variations in incidence rates in small populations and areas, the annualized average incidence of HFRS per 100,000 in each town over the 15-year period was calculated. Furthermore, the annualized average incidences for each town were mapped in gradient colours. Figure 2(a-o) shows the annualized incidence for each town.

Parameter estimation
According to the parameter estimation and goodness of fit test results (Table 1 and Table   2), we selected ARIMA (0, 1, 1) (0, 1, 1)12 model as the best one . The goodness of fit analysis indicated that there was no significant autocorrelation among residuals with different lags.

Model testing and forecast analysis
The monthly data in 2000-2013 was used to construct the ARIMA (0, 1, 1) (0, 1, 1) 12 model (Table 3 and Figure 3) , and the monthly data in 2014 were used to test the model. The predicted data, actual data and the 95 % confidence limit for the predicted data for 2014 are shown in Table 3 and Figure 3. The predicted data didn't exactly match the observed data, but the observed data fell within the 95% confidence interval of the predicted data.

Multiple linear analysis of meteorological factors and monthly HFRS incidence
Bivariate linear analysis results showed that monthly mean temperature, monthly precipitation and wind speed were related to HFRS incidence. A multiple linear analysis method was further used to model the relationship between monthly HFRS incidence and meteorological factors. The results showed that monthly precipitation, average wind speed and relative humidity were important factors in HFRS incidence in Anqiu City (Table 4).

Discussion
Epidemiological surveillance is one of the important interventions in the prevention and control of infectious diseases. Time series analysis methods are very useful for dynamic prediction and effective control of the diseases. Our study showed that the occurrence of HFRS presented apparent seasonality and that there were two annual peaks: the smaller peak occurred in spring between March and April, and the larger peak occurred in winter between October and November. From a public health perspective, our results support the need to carry out deratization campaigns in spring and winter around Anqiu City as well as enhance population immunity by vaccination throughout the year. In addition, we applied a multiplicative seasonal ARIMA (0, 1, 1) (0, 1, 1) 12 model to analyse the HFRS surveillance data in Anqiu City, China. Based on the results above, the ARIMA (0, 1, 1)×(0, 1, 1) 12 model is reliable with high prediction accuracy and can be used to predict the HFRS incidence in the subsequent year in Anqiu City. The prediction results suggested that the HFRS ARIMA model has strong ability to forecast and predict the incidence of HFRS in Anqiu City. Therefore, it is very important and necessary to learn about the knowledge of HFRS forecasts, which can help health agencies to allocate health resources reasonably.
Together with time series analyses, the application of GIS in our study provides a way to explicitly quantify HFRS and further determine the epidemiological characteristics accounting for the increasing disease risk. Our results showed that the potential risk areas were mainly concentrated around towns west of Anqiu City in recent years. The reason for this result may be that most towns in the west area are mountainous regions where the living environment is poor and villagers have weak health consciousness. Our study also indicated that an increase in HFRS incidence in Anqiu City was observed over the past 5 years, especially in 2012 when there was a small increase ( Figure 2). This result may be due to the reconstruction and renovation of the old areas of Anqiu City in 2012, which included the renovation of a large number of houses, destroyed the rodent habitat, and made rats move frequently. These renovations also decreased the quality of the living environments of villagers and increased the chances of coming in contact with hantavirus, which further caused an increase in HFRS incidence. Based on our results, the government can allocate more health resources to high-risk areas and reduce the number of these resources used in low-risk areas to improve the effectiveness of interventions and the allocation of medical resources. Overall, both the temporal and spatial distribution patterns of HFRS in Anqiu City were studied. Disease prevention and control measures need to be strengthened in high-risk areas and during high-risk periods, which vary by month of the year.
In our study, correlation analysis and multiple linear analysis were used to further explore the relationship between environmental factors and HFRS incidence in Anqiu City. Our results showed that there was a negative association between temperature and HFRS incidence in Anqiu City, which was supported by other previous studies[ 12 25 ]. High temperatures may limit the time available to farmers for outdoor activity and work, thereby reducing the opportunity for contact between people and field mice, which is one of the most common agricultural pests and a natural vector of hantavirus. In addition, some studies have suggested that the breeding rate of rodents is highest at temperatures of 10-25°C, which are favourable conditions for outdoor activity and work [ 25 ]. Anqiu City is located in a warm temperate continental monsoon climate zone, and the annual average temperature is 12.2°C. However, inconsistent findings have been reported in other studies that indicated a positive association between temperature and the incidence of HFRS[ 24 ] [ 26 ]. This discrepancy might be due to different local conditions, such as different rodent compositions, different hantavirus serotypes, different environments and different climates in the study regions. Our data also indicated that precipitation was negatively associated with the incidence of HFRS; this finding is consistent with the findings of previous studies [ 27 , 28 ]. Abundant precipitation could have a negative impact on rodents by destroying their habitats. In addition, frequent rainfall may decrease the likelihood of rodent-tohuman contact, rodent-to-rodent contact, and virus transmission due to decreased rodent activity and reduced human exposure. However, several previous studies showed inconsistent findings of a positive association between precipitation and HFRS incidence [29][30][31]. There is no clear explanation for such differences, which may reflect the heterogeneity in local climate conditions. Further studies should be conducted in different regions to gain a better understanding of the impact of precipitation on HFRS. In addition, a positive association was obtained between average wind speed and HFRS incidence in our study, which was consistent with the results of a previous study