Spatial epidemiological determinants of severe fever with thrombocytopenia syndrome in Miyazaki, Japan: a GWLR modeling study

Severe fever with thrombocytopenia syndrome (SFTS) is a fatal tick-borne viral disease that involves thrombocytopenia, caused by the SFTS virus (SFTSV), a phlebovirus in the order Bunyavirales [1, 2]. The disease is characterized by non-specific symptoms, most frequently involving fever, gastrointestinal symptoms, leukopenia, thrombocytopenia, and liver dysfunction, and sometimes involving hemophagocytic syndrome and hemorrhagic tendencies, leading to multiple organ failure [1‐4]. The disease was first detected in northeast China in 2007, followed by identification of its common clinical symptoms in 2009 and virus isolation in 2010 [1, 2]. The virus lifecycle and natural mechanisms of sustained transmission (e.g., reservoir animal species) remain to be fully clarified, but transmission via ticks is considered the most plausible route [5‐12]. SFTSV has been detected in cats, mice, and wild boars [5, 13‐15], and humans are believed to be an accidental host. Although SFTS was initially identified in China, the disease has also been observed in South Korea, Japan, and across all of East Asia [4, 16, 17]. However, the magnitude of cases in Japan has been limited; a total of 391 confirmed cases had been notified in Japan as of 12 December 2018, whereas SFTS cases in China had already numbered in the thousands by the same date.

In recent years, the frequency and geographic areas at risk of tick-borne infectious diseases are known to have expanded in North America, possibly caused by ecological changes of tick abundance that are possibly induced by climate change [18]. Geographic heterogeneity is evident in East Asia as well; for example, the eastern to central mountainous regions of China and Western Japan are known to be focal areas with a higher frequency of SFTS notifications than other regions [4]. Published studies have identified that at altitudes of 80–400 m [15, 19, 20], a number of climatological variables, including average temperature, humidity, and precipitation [10, 15, 19, 21, 22], as well as the density of livestock [7, 8, 10, 13, 17], are associated with an elevated risk of tick-borne infections. Other than published epidemiological studies, the epidemiological data in Japan imply that confirmed cases are frequently seen among farmers who live in rural regions or near forested areas [3, 4, 23]. The case fatality risk (CFR) among confirmed cases was about 30% in 2013 [4]; however, since then, both the CFR and absolute number of deaths have steadily decreased in Japan (e.g., of a total 72 cases in 2018, 3 deaths occurred, with a CFR about 5%) [24].

In Japan, SFTS is designated as a category IV infectious disease according to the Infectious Diseases Law, mandating physicians to notify all laboratory-confirmed cases to the government via the nearest health center [25]. For this reason, all 47 prefectural institutes of public health in Japan are equipped with one-step RT-PCR testing kits, for laboratory confirmation in every prefecture. The collected notification data are sent electronically to the headquarters of the National Institute of Infectious Diseases, and weekly notifications by prefecture are publicly reported. The expected micro-geographic locations of infection are not publicly reported for all of Japan, but suspected locations of infection are reported only among prefectures with high incidence, as described in the Methods below. Moreover, there have been no published epidemiological studies in Japan identifying micro-geographic areas at high risk of SFTSV infection; thus, the underlying mechanisms of infection in geographically high-risk areas have yet to be elucidated. The generally acknowledged method for preventing SFTS is to avoid tick bites by (i) ensuring that there is no exposure, such as by avoiding high-risk areas; (ii) keeping ticks away from exposed skin by wearing long-sleeved shirts, long pants, and high boots; (iii) using insect repellent, such as products containing DEET, on the body; and (iv) checking the entire body for ticks after activities associated with potential exposure. Despite this knowledge, many individuals have considerable difficulty in completely avoiding tick bites owing to their occupation or living circumstances [26, 27].

To develop effective future countermeasures, it is critical to identify and communicate the risk factors of infection, with high geospatial resolution at the community level. As a first step, in the present study we aimed to identify areas with a high risk of SFTSV infection and their characteristics using a geospatial dataset of SFTS cases in Miyazaki, Japan.

Figure 1a shows a map of all suspected geographic locations of SFTSV infection (n = 42, including 8 locations in 2017). Miyazaki Prefecture faces the Pacific Ocean on its eastern margin, and it borders Oita and Kagoshima prefectures on its northern and southern margins, respectively. The western margin of Miyazaki is a mountainous area bordering Kumamoto Prefecture. The eastern area aligned with the coast is referred to as Miyazaki Plain, where the population density tends to be higher than other areas. Although arbitrary, Fig. 1a indicates that cases of SFTS occurred near rivers, in transitional zones between farms and forest, and at altitudes that were not very elevated but also not close to sea level. Figure 1b shows a comparison of 34 locations at the end of 2016 and an additional 8 locations in 2017. An SFTS case was observed for the first time in 2017 only in mesh units overlapping with Takaharu town, which had no previously reported cases; otherwise, all mesh areas had at least one reported case prior to 2017.

Figure 1c and d shows the altitude and proportion of farmland at the second-order mesh level. Using univariate LR analysis, altitude (p < 0.01) was identified as a significant negative predictor of SFTS cases whereas farmland alone (p = 0.70), the proportion of elderly residents (p = 0.81), proportion of abandoned farmland (p = 0.46), and forest ratio (p = 0.38) did not have any significant associations with the presence of SFTS. Annual average precipitation (p = 0.94) was not significantly associated with the presence of SFTS whereas annual average temperature showed a positive association with SFTS (p = 0.05). Nevertheless, the significant association between SFTS and average temperature disappeared when altitude was modeled together with the average temperature and an interaction term of these variables was taken into account (p = 0.21 for temperature and p = 0.03 for altitude). Among non-significant variables, farmland showed significant correlations with the proportion of abandoned farmland (correlation coefficient, r = 0.75), forest ratio (r = 0.82), and proportion of elderly residents (r = 0.38). Of these, farmland was the variable that excellently reflects the changing land use from farmland to forest, and moreover, the observations depicted in Fig. 1a indicate frequent infections at mid-level altitudes and also in areas between farms and mountains; therefore, owing to the univariate results as well as the correlations among explanatory variables, we used only altitude and farmland as explanatory variables in the subsequent multivariate analyses.

Using altitude and farmland as input, we optimized the multivariate GWLR model. Table 1 shows the global terms of GWLR, revealing that altitude was a significant negative factor of infection. When the data at the end of 2016 were used to predict 2017, both altitude and farmland appeared to act as significant negative predictors. When all existing cases were predicted, GWLR outperformed the LR by means of the likelihood ratio test (p = 0.04). Similarly, when 2017 was predicted, the GWLR model significantly reduced deviations from the LR with a p-value < 0.001 (likelihood ratio test).

Figure 2a and c shows the predicted risk maps along with the evaluation of predictive performance using the ROC curve. The AUC of GWLR (73.9%) was greater than that of LR (72.4%) (Table 2). Figure 2b and d shows the predicted risk maps with the ROC curve of prediction in 2017. Again, the AUC of GWLR (76.6%) outperformed that of LR (75.6%). Except for the one case in 2017 in Takaharu town, the remaining cases (5 of 6 meshes containing a total of 7 cases) occurred in mesh areas where the risk of SFTS cases was above the median, and the concordance of the estimates with the actual observed locations (empty points) was verified.

Using the Youden index, the sensitivity and specificity of LR were estimated at 86.4 and 61.3%, and those of GWLR were 90.9 and 58.7%, respectively (Table 2). The PPV and NPV of LR were 39.6 and 93.9%, and comparable with those of GWLR, estimated at 39.2 and 95.7%, respectively. The actual observed mesh areas in Miyazaki, Nobeoka, and Nichinan cities were in line with high-risk areas, as indicated by the risk maps. In the prediction using only the 2017 data, the sensitivity and specificity of LR were estimated at 83.3 and 75.8%, and those of GWLR were 83.3 and 73.6%, respectively (Table 2). The PPV and NPV of LR were 18.5 and 98.6%, and those of GWLR were comparable, estimated at 17.2 and 98.5%, respectively. Precision and F1-score, given as the harmonic average of precision and sensitivity (=recall), were also comparable between the LR and GWLR models (Table 2).

Table 3 shows the summary statistics of local coefficients for altitude and farmland using the GWLR model. Both altitude and farmland yielded negative coefficient values, indicating that the local coefficients of these two variables always highlighted them as significant negative predictors of SFTS cases. Whereas the coefficients of altitude varied widely, those of farmland only varied slightly among the included mesh areas in Miyazaki. Similar patterns were seen when predicting 2017 data only, but variations in the estimated coefficients for farmland were greater than those predicted for the entire dataset.

As a sensitivity analysis, we examined possible explanatory factors, using 1 × 1-km mesh data and a multivariable LR model. Altitude (p < 0.001) and farmland (p < 0.02) consistently appeared to be significant explanatory variables, explaining the observed pattern well (AUC = 72.4%).

In the present study, we analyzed the spatial epidemiological risk of SFTS in Miyazaki, Japan, using 10 × 10-km mesh data and LR and GWLR models. By exploiting the spatial variation in the risk of SFTS, we attempted to identify environmental risk factors, exploring a number of variables including the proportion of abandoned farmland and the forest ratio, which are possibly associated with the vector ecology and risk of tick bites. From the results depicted in Fig. 1b, we hypothesized that most cases of SFTS in Miyazaki, Nobeoka, and Nichinan cities occur along the lower reaches of rivers at mid-level altitudes in locations between farmland and forest. Accordingly, we restricted the explanatory variables to altitude and farmland and identified these two variables as useful negative predictors of SFTS cases in Miyazaki.

Altitude has been identified as an important predictor of SFTS in China [15, 19, 20], indicating that the disease is unlikely to occur at very elevated locations. Published studies [3‐5, 11, 23] have also suggested that rural areas are likely to have higher levels SFTSV infection risk. The present study findings, showing a greater tendency of infections in rural areas, were well captured using farmland as an input variable. In Miyazaki, cases of SFTS in rural regions occurred in relatively lowland areas near rivers, which was well captured by jointly using altitude and farmland in the prediction model. Our modeling study findings lead us to conclude that the risk of SFTS is higher in geographic areas with a lower number of farms and at lower altitudes. The number of farms decrease as altitude increases, while altitude decreases as the location comes close to the ocean with greater number of farms than other locations, implying that areas with small number of farms and at mid-level altitudes may be at high risk of infection.

Considering geographic locations with a low density of farms, flatlands near hilly or mountainous locations along rivers were found to be hot spots of SFTSV infection. The implications of the present study results will greatly assist in future surveillance of both animals and vectors. For instance, local public health organizations regularly conduct surveillance of the tick distribution, and conduct sample collection and laboratory testing of Ixodes species to identify SFTSV infections [30]. The present study findings emphasize the need to selectively conduct frequent intensive ecological surveys in high-risk areas to identify the virus. Moreover, it is important to implement animal surveillance near high-risk areas, and to include surveys of wildlife species. The range of animal host species, including those acting as the interface between wildlife species and humans, could be better identified by conducting intensive surveys of animals in regions with high risk of SFTSV infections.

The present study includes several limitations. First, the total number of mesh areas with SFTS was limited to 22 of 97 mesh units; thus, the resolution of prediction was limited (e.g., with a greater number of cases, one might even consider using a 1 × 1-km mesh unit). This may be regarded as imbalanced data, which is a threat to stable prediction in logistic regression, and the performance of model prediction is known to be significantly hindered by class imbalances. Although the observed imbalance ratio remained smaller than 4:1, the recognized problem would not be eased unless more data become available and the imbalances are resolved [45]. What must be remembered in this regard is that the observed patterns may not have adequately reflected the distribution of explanatory variables. Owing to a related reason (i.e., limited precision of mesh data), we were unable to incorporate an original (more concrete) variable, i.e., location close to rivers and in transitional zones between farms, into our prediction using a 10 × 10-km mesh unit. Second, the suspected locations of infection rested on interviews; considering the map scale of the original data, there could be measurement error with respect to the precision of the suggested geographic points at finer spatial resolutions. Third, whereas we focused on a binary outcome (i.e., presence or absence of SFTS cases), some mesh areas had already experienced multiple cases of SFTS. Owing to a small total number of cases, prediction of the number of cases is a future task to be performed. Fourth, we ignored climatological variables including temperature, humidity, and precipitation owing to the absence of dynamic (time-dependent) case data over geographic space [46]. Fifth, infection risk depends on both (i) human behaviors that lead to tick bites and (ii) the risk of infection among ticks (and animals) [47‐49]. In the present study, we could not distinguish (i) and (ii) nor elucidate the detailed ecological mechanisms behind the observations [50]; this would require cross-disciplinary collaboration among epidemiologists and statisticians as well as entomologists, ecologists, veterinarians, and virologists.

Despite the number of important tasks that remain to be completed in future, in the present study, we successfully identified altitude and farmland as negative predictors of SFTS cases in Miyazaki, concluding that the risk of SFTS is high in geographic areas with decreased farmland and at middle altitudes. Our study findings not only indicate high-risk areas that are suitable for conducting ecological and animal surveys in the future, we also believe that it is critical to raise awareness among local residents in such areas about the existing risk of SFTS and to emphasize the importance of preventing tick bites at the local level.