The effects of meteorological factors on hemorrhagic fever with renal syndrome in Yichun, China: a major Apodemus-type endemic city

Abstract

Historically, Yichun City has borne the highest burden of Hemorrhagic Fever with Renal Syndrome (HFRS) in China, with incidence exhibiting marked fluctuations between 2005 and 2023. Despite this, HFRS-weather associations in Yichun remain poorly characterized. We analyzed weekly time-series data on HFRS cases and meteorological factors (mainly temperature, humidity, and precipitation) from 2005 to 2023 using random forest models to identify optimal lags, distributed lag non-linear models, and generalized additive models. We identified a symmetric unimodal relationship for temperature (peak risk at 13.66 °C; RR = 1.77, 95% CI: 0.92–3.39), a right-skewed pattern for relative humidity, and a distinct inverse J-shaped pattern for precipitation. Lag-response analysis indicated that the strongest associations occurred at 26 weeks for temperature, 12 weeks for relative humidity, and 16 weeks for precipitation. Temperature showed complex delayed responses with both protective and risk-increasing phases (~ 4–6 months), while precipitation and humidity primarily exhibited protective effects. Formal interaction testing revealed no statistically significant two-way or three-way interactions among meteorological variables (all P > 0.05), indicating that each factor operates independently rather than synergistically. These findings offer actionable evidence for designing targeted, climate-informed HFRS prevention strategies in Yichun and similar high-burden regions.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1038/s41598-026-59003-x.

Introduction

Hemorrhagic Fever with Renal Syndrome (HFRS) is a rodent-borne zoonosis caused by hantaviruses and characterized by fever, hemorrhage, headache, back pain, abdominal pain, and acute renal failure and, in severe cases, death¹. Humans primarily acquire hantaviruses through inhalation of aerosolized particles from excreta of infected rodent reservoir. In China, HFRS is predominantly caused by two hantavirus species: Hantaan virus (HTNV), carried by the striped field mouse (Apodemus agrarius), and Seoul virus (SEOV), maintained by the Norway rat (Rattus norvegicus)². HTNV typically causes more severe disease, with a pronounced autumn-winter seasonal peak, whereas SEOV-associated HFRS tends to be milder and peaks in spring³. Ecologically, A. agrarius thrives in rural agricultural landscapes, while R. norvegicus is an anthropophilic species commonly found in urban environments⁴. HFRS is endemic in all 31 provincial-level administrative divisions in China. Despite public health interventions, including rodent control, environmental sanitation, and vaccination, the country continues to account for 70–90% of global HFRS cases, with 207,850 recorded between 2005 and 2024⁵. Among Chinese provinces, Jiangxi exhibits one of the highest HFRS burden, reporting 8,981 cases and a case fatality rate of 1.16% from 2005 to 2021⁶.

Rodent-borne viral diseases are among the most climate-sensitive infectious diseases. Numerous epidemiological investigations have demonstrated that variations in temperature, precipitation, and humidity significantly influence the transmission dynamics and geographic distribution of HFRS^7–10. Nevertheless, the direction, magnitude and lag structure of these meteorological effects remain inconsistent across settings and warrant further investigation. Discrepancies in modeling strategies have compromised the reproducibility of findings, and pooled estimates that fail to account for regional heterogeneity in climate zones and dominant reservoir species may yield misleading, or even erroneous, conclusions¹¹. To address this ecological complexity, recent multi-city studies have employed stratified analytical frameworks. For example, Luo et al. analyzed 151 cities across three climate zones and estimated HFRS-weather associations separately within each zone¹². Similarly, Chang et al. classified 285 cities into three rodent-driven categories based on the seasonal incidence pattern of HFRS and examined meteorological associations within each group¹³. These efforts highlight a critical methodological prerequisite: accurate identification of local climate zone and dominant reservoir species for each study location. Unfortunately, China’s national disease surveillance system does not routinely differentiate between hantavirus genotypes in clinical HFRS cases, and high-quality, longitudinal rodent density data remain scarce. Moreover, published genotyping data are often derived from non-representative sampling, introducing potential selection bias that may compromise the validity of reservoir-based city classifications¹¹. Given these limitations, multi-city analyses risk substantial ecological bias when estimating meteorological effects on HFRS due to unaccounted heterogeneity in reservoir ecology, land use, and microclimate. In contrast, single-city studies can yield more robust, interpretable, and policy-relevant estimates of local HFRS-weather relationships^14–17. Consequently, prioritizing understudied, high-endemic cities is essential, not only to clarify localized HFRS-weather dynamics but also to strengthen the evidence base for early warning systems currently constrained by the limitations of aggregated, multi-city models.

Yichun City, a prefecture-city of Jiangxi Province, serves as the regional epicenter of HFRS, accounting for over 50% of all provincial cases¹⁸. Surveillance records spanning the past six decades document approximately 57,000 HFRS cases and over 1,879 deaths in Yichun alone. Recent trends show that incidence rose steadily, peaked in 2018, and subsequently declined^6,19, a pattern reflecting a persistent and substantial disease burden. Although R. norvegicus is a confirmed carried, A. agrarius is the primary reservoir responsible for human infections in this region, a role strongly supported by Yichun’s extensive agricultural terrain and humid subtropical climate, which favor the proliferation of this species^18,20. Despite this high burden, the HFRS-weather associations in Yichun remains poorly characterized. To address this gap, we conducted a weekly time-series analysis using random forest models to identify optimal lag structures. We then fitted a distributed lag non-linear model (DLNM) within a generalized additive model (GAM) framework to quantify the delayed effects and interactions of temperature, humidity, and precipitation on HFRS incidence. Specifically, we examined the non-linear relationships and lagged effects of meteorological factors, tested for potential two-way and three-way interactions among these variables, and assessed their independent contributions to HFRS risk. Our findings advance understanding of the local climatic drivers of HFRS in a high-risk setting. These insights may inform the development of region-specific early epidemic warning systems and support the evidence-based allocation of public health resources to mitigate future outbreaks.

Methods

Study area

Yichun (27°33’ − 29°06’ N, 113°54’ − 116°27’ E) is situated in the northwest region of Jiangxi Province in Central China (Fig. 1). The region is 174 km wide and 223 km long, with a total land area of 18,700 km². Yichun includes a municipal district (Yuanzhou) and nine counties (Jing’an, Fengxin, Shanggao, Yifeng, Tonggu, Wanzai, Gao’an, Zhangshu, Fengcheng) with a population of about 5.58 million in 2024. Yichun has a humid subtropical climate and an annual average temperature of 17.5 °C, relative humidity of 80%, and precipitation 1,680 mm.

Fig. 1.

Geographic location of Yichun in Jiangxi Province, China. The map was created in ArcGIS 10.7 using the political boundaries from the National Geomatics Center of China for illustrated sample location only.

Data collection

HFRS is a class B notifiable infectious disease in China, diagnosed according to the Diagnostic Criteria for Epidemic Hemorrhagic Fever issued by the Ministry of Health (https://www.nhc.gov.cn/wjw/s9491/200802/39043.shtml). Medical institutions are required to report HFRS cases within 24 h in a standardized format through the web-based surveillance system. Daily data on HFRS cases from January 1, 2005, to December 31, 2023, were obtained from the China Disease Control and Prevention Information System. Daily meteorological data for the same period, including average temperature (°C), relative humidity (%), and cumulative precipitation (mm), were retrieved from the China Meteorological Data Sharing Service System (http://data.cma.cn/). City-level daily average values of meteorological factors were calculated as the mean values of the data from the four climate monitoring stations in Zhangshu, Jing’an, Yifeng, and Yuanzhou. To address the overdispersion caused by excessive zero-values in daily HFRS records, raw daily data were aggregated to a weekly scale. We derived weekly averages for temperature (TEM) and relative humidity (RHU), while weekly cumulative totals for HFRS case counts and precipitation (PRE).

Data analysis

Descriptive analysis

Descriptive statistics, including mean, standard deviation, median, and interquartile range (P25, P75), were used to characterize HFRS incidence and meteorological factors. Temporal trends were visualized to inspect patterns over the study period.

Spearman’s rank correlation analysis was conducted to assess the relationships between weekly HFRS case counts and meteorological factors, as well as the inter-correlations among meteorological variables. To mitigate multicollinearity, variables with an absolute correlation coefficient greater than 0.7 (|r|> 0.7) were not included simultaneously in the model, thereby preserving the accuracy of variable importance estimates in subsequent analyses.

To identify optimal lag structures for the distributed lag models, a Random Forest (RF) regression framework with 1,000 trees was utilized. The temporal dimension was explicitly modeled by incorporating lag periods of up to 26 weeks based on prior knowledge of HFRS incubation and reporting patterns²¹. The significance of each meteorological predictor was evaluated using the %IncMSE metric, which measures the increase in prediction error (mean squared residuals) when values of a given variable is permuted in the out-of bag (OOB) samples, thus elucidating the complex, time-lagged associations between weather patterns and HFRS outbreaks.

These preliminary analyses informed the selection of optimal parameters for subsequent modeling.

Exposure-risk analysis

To quantify the exposure-risk relationships, a DLNM was applied. The time-series framework employed a “cross-basis” function to simultaneously describe the exposure-response curve and its distributed lag effects. Given the relatively low number of weekly HFRS cases, the DLNM was embedded within a GAM framework with a quasi-Poisson link function to control for overdispersion which was confirmed through preliminary diagnosed tests. The basic model structure was:

1

Where E(Y_t ) represents the expected case count for week t, and α is the intercept. The terms , , and represent the cross-basis functions for temperature, precipitation, and relative humidity, respectively, capturing the non-linear and delayed effects of meteorological variables on HFRS incidence. , , and denote the degrees of freedom (df) for the natural cubic splines used in the cross-basis functions. Factor (year) and factor (month) controlled for long-term trends and seasonal variation.

The quasi-Akaike Information Criterion (QAIC) was employed to select the optimal df, with lower values indicating better model fit. The optimal df for both the exposure-response and lag-response dimensions was determined though QAIC comparison to be 3 for all meteorological variables. Relative risks (RRs) and 95% confidence intervals (CIs) were calculated using the median of each variable as the reference (18.71 °C for TEM, 82.05% for RHU, and 23.16 mm for PRE). To assess the impact of weather extremes, the 1st and 99th percentiles were analyzed, while the 25th and 75th percentiles were used to represent moderate conditions.

Interaction effect analysis

To investigate potential interactions among meteorological variables, we extended the GAM framework to explore two-way smooth interactions:

2

Where β denotes the intercept, and s() represents penalized spline function. k refers to one of the meteorological factors (TEM, RHU, and PRE), while x and z represent the other two meteorological factors. The term s₁ (k, x) is a two-dimensional smooth function that captures the non-linear interaction between the pair of meteorological variables (k, x), whereas s₂(z) models the independent effect of an additional meteorological variable. We included factor(year) and factor(month) to control for long-term trends and seasonal variations.

Sensitivity analysis

A sensitivity analysis was conducted to test the robustness of our primary findings. First, we assessed the impact of key model specifications, such as variations in the df for the cross-basis functions. Second, given the relatively high correlation between precipitation and relative humidity, we performed additional sensitivity analyses to evaluate the potential impact of multicollinearity. Specifically, we compared the cumulative exposure-response associations of temperature derived from the full model (adjusted for both PRE and RHU) with those from reduced models excluding PRE or RHU individually. The stability of the exposure-response curves across these models was used to confirm the robustness of the main estimates.

All statistical analyses were performed using R software (version 4.3.0; R Foundation for Statistical Computing, Vienna, Austria) with the “randomForest”, “dlnm”, “splines”, “mgcv”, and “ggplot2” packages. A P-value of less than 0.05 was considered statistically significant. Spatial mapping of epidemiological data was conducted using ArcGIS software (version 10.7; Esri Inc, Redlands, CA, USA).

Ethics approval and consent to participate

This study was conducted in accordance with the Declaration of Helsinki. The study protocol and the use of anonymized surveillance data were reviewed and approved by the Ethics Committee of the Nanchang City Center for Disease Control and Prevention (Approval No: NCCDC-2023-007). The requirement for informed consent was waived by the same ethics committee because this study involved the analysis of pre-existing, fully anonymized public health surveillance data, and the research presented no more than minimal risk to the individuals whose data were used.

Results

Descriptive Analysis

A total of 4,599 HFRS cases were reported in Yichun City between 2005 and 2023. As summarized in Table 1, the weekly average case count was 4.66, with a peak of 31 cases in a single week. Figure 2 displays the time series of HFRS cases alongside meteorological factors, highlighting a clear seasonal cycle in all climate variables, with a colder spring, wet and warm summer followed by a colder, drier winter, and illustrating their correlation with the seasonality of HFRS incidence.

Fig. 2.

Time series of weekly HFRS cases and meteorological factors in Yichun City, 2005–2023. Abbreviation: HFRS, hemorrhagic fever with renal syndrome.

Table 1.

Summary statistics for weekly HFRS cases and meteorological factors in Yichun City, 2005–2023.

Variables	Mean	S.D.	Minimum	P (25)	Median	P (75)	Maximum
HFRS cases	4.66	4.03	0.00	2.00	4.00	6.00	31.00
TEM	17.55	8.40	-3.05	10.01	18.75	25.25	30.52
RHU (%)	80.29	7.14	51.27	76.74	81.97	85.45	92.85
PRE (mm)	33.85	35.88	0.00	6.04	23.95	51.17	252.14

Abbreviation: HFRS, hemorrhagic fever with renal syndrome; S.D., standard deviation; TEM, weekly average temperature; RHU, weekly average relative humidity; PRE, weekly cumulative precipitation.

Discussion

Yichun has long been a focus point of HFRS epidemics in China, with incidence exhibiting marked fluctuations during the study period⁶. Understanding the HFRS-weather associations is increasingly critical in the context of climate change, as it informs risk mitigation strategies and supports anticipatory public health planning^22–24. To our knowledge, this study represents the first quantitative assessment of the effects of temperature, relative humidity, and precipitation on HFRS incidence in Yichun, revealing complex non-linear and lagged associations with distinct temporal patterns.

The observed time lags between meteorological exposure and HFRS incidence are biologically plausible, reflecting the sequential nature of ecological processes that drive transmission: vegetation growth influences food availability which affects rodent reproduction and population dynamics, ultimately modulating hantavirus spillover to humans^25–27. In Yichun, the strongest associations with weekly HFRS cases occurred at lags of 26 weeks (approximately 6 months) for temperature, 12 weeks for relative humidity, and 16 weeks for precipitation, suggesting a medium- to long-term pathway linking climate conditions to human case reporting through reservoir ecology. Notably, this lag structure varies geographically: in Elunchun, climatic effects on HFRS emerged about one month earlier than in the neighboring Molidawahaner²⁸. While the drivers of this disparity remain unclear, potentially involving differences in surveillance sensitivity or, more likely, distinct transmission ecologies, the consistent identification of a 6-month lead time offers a practical window for epidemic forecasting. This interval enables public health authorities to implement timely interventions, including targeted rodent control, vaccination campaigns, environmental management, and community-based early warnings.

Our analysis revealed distinct exposure-response relationship for each meteorological variable. Temperature showed a unimodal temperature-HFRS association, peaking at moderate temperatures (13.66 °C), consistent with previous studies^11,14,29. This pattern reflects temperature’s multifaceted influence on rodent reproduction, survival, and viral persistence^30,31. Notably, our optimal temperature estimate (13.66 °C) differs from other regions (10 °C in Weifang¹⁴, and 22 °C in Guangzhou³², likely reflecting regional variations in climate, reservoirs ecology, and viral strains. Consequently, future multi-city analyses should stratify cities by epidemic type and climate region to accurately assess location-specific climatic drivers of HFRS.

Relative humidity demonstrated a right-skewed unimodal relationship, while precipitation showed an inverse J-shaped, suggesting that moderate moisture conditions favor transmission by maintaining suitable habitable and food resources for rodent reservoirs^33,34. However, extreme conditions (both very low humidity and high precipitation) showed protective effects, possibly due to environmental stress disrupting transmission cycles. Yichun exemplifies this high-risk profile with its humid subtropical monsoon climate and landscape composition (“six parts mountain, one part water, two parts fields, and one part roads and settlements”) fostering ideal rodent habitats. With high average relative humidity (80.27%) and abundant precipitation, Yichun’s climate aligns with typical profiles of HFRS-endemic areas in China, which are predominantly humid or semi-humid^35,36. Monitoring fluctuations in temperature, relative humidity, and precipitation could serve as a valuable component of integrated HFRS surveillance, enabling timely deployment of preventive measures.

The distinct lagged patterns reveal different temporal pathways: temperature showed complex delayed responses with both protective and risk-increasing phases (~ 4–6 months), while precipitation and humidity primarily exhibited protective effects. These delayed responses likely reflect the time required for meteorological changes to influence rodent population dynamics and virus persistence, supporting the use of seasonal temperature forecasts for early warning^26,37.

Despite visually suggestive patterns in the 3D response surfaces (e.g., elevated risk under low temperature–high humidity or low temperature–low precipitation), formal tests revealed no statistically significant two-way or three-way interactions among temperature, relative humidity, and precipitation (all P > 0.05). This indicates that the influence of each meteorological factor on HFRS risk operates largely independently—rather than synergistically or antagonistically—in this setting. The absence of interaction supports modeling HFRS risk using additive effects of individual variables, with temperature emerging as the dominant driver. While ecological plausibility exists for combined effects (e.g., cold-damp conditions favoring rodent survival), our data suggest such interactions are either weak or masked by other confounders (e.g., land use, host density). Thus, predictive models may prioritize univariate lagged temperature signals over complex interaction terms.

This study has several limitations. First, we were unable to account for the potential impact of HFRS vaccination, a key intervention in China. Vaccination coverage in Yichun has been sporadic and non-systematic, precluding reliable quantification in our models. Second, despite the central role of rodent reservoirs, we could not incorporate rodent density data due to the absence of continuous, standardized surveillance. Trapping-based indices are often fragmented and difficult to scale temporally or spatially. Third, although HFRS case data were drawn from China’s national notifiable disease system, underreporting is likely, particularly for mild or asymptomatic infections, or among populations with limited healthcare access, potentially leading to underestimation of true incidence. Finally, the lack of virological confirmation (i.e., SEOV vs. HTNV) for individual cases prevented virus-specific analysis of climatic analyses. Given the distinct ecologies of these hantaviruses, this limitation may obscure nuanced transmission dynamics.

In conclusion, these findings have important implications for HFRS surveillance and prevention. The consistent 6-month lag for temperature effects provides a reliable forecasting window for public health authorities to implement targeted interventions, including enhanced rodent control, vaccination campaigns, and community awareness programs. The lack of significant interactions suggests that temperature-based early warning systems may be sufficient without incorporating complex multivariate models, making implementation more feasible at local scales.

Over the 19-year study period, none of the meteorological variables exhibited substantial interannual variability. Weekly conditions consistently followed stable intra-annual patterns: TEM fluctuated between − 3.05 °C and 30.52 °C, RHU ranged from 51.27% to 92.85%, and PRE varied from 0.00 to 252.14 mm. Deviations from these patterns occurred only during occasional extreme weather events.

Correlation between meteorological factors and HFRS cases

Spearman correlation analysis indicated that the association between meteorological factors and HFRS incidence varied by lag weeks. At lag 0, TEM was negatively correlated with HFRS cases (r = -0.40, p < 0.05). In contrast, the linear correlations for PRE and RHU were week or non-significant at lag 0 (Table 2). These results suggest the presence of complex non-linear relationships and delayed effects, necessitating further analysis using the RF model.

Table 2.

Spearman correlation analysis of HFRS cases and meteorological factors in Yichun City, 2005–2023.

Variables	HFRS cases	TEM (℃)	RHU (%)	PRE (mm)
HFRS cases	1.00***
TEM	-0.40***	1.00***
RHU (%)	-0.09**	0.32***	1.00***
PRE (mm)	-0.02	0.19***	0.68***	1.00***

Abbreviation: HFRS, hemorrhagic fever with renal syndrome; TEM, weekly average temperature; RHU, weekly average relative humidity; PRE, weekly cumulative precipitation. Notes: * P < 0.05, ** P < 0.01, ** P < 0.001;

RF model performance and variable importance

The RF model, incorporating meteorological variables with lags of up to 26 weeks, explained approximately 50.0% of the variance in weekly HFRS cases (R² = 0.500). Applying a variable importance threshold of 5.0% IncMSE, 40 lagged predictors were retained, including 24 TEM-related variables (current week and lags 2 - 26), 9 PRE-related variables (lags 4–9, 13–14, and 16), and 7 RHU-related variables (lag 4, and lags 7–12). Thus, the maximum relevant lag periods were identified as 26 weeks for TEM, 16 weeks for PRE, and 12 weeks for RHU.

TEM consistently ranked as the most influential predictor. TEM_lag20 exhibited the highest importance score (%IncMSE = 16.46), followed by TEM_lag4 (16.00) and TEM_lag5 (15.81). Among the top 20 variables, 15 were TEM-related, and their importance persisted across almost the entire 26-week lag period, indicating a prolonged cumulative effect of temperature on HFRS transmission. In contrast, PRE and RHU exhibited more localized peaks: RHU_lag9 (12.33) and RHU_lag12 (11.36) were the dominant humidity predictors, while PRE_lag7 (10.10) was the top PRE predictor, suggesting shorter-term impacts on disease dynamics (Fig. 3).

Fig. 3.

Importance of lagged meteorological variables on HFRS incidence based on the Random Forest model. The Y-axis represents the percentage increase in Mean Squared Error (MSE) when a specific variable is excluded, indicating its contribution to the model’s predictive accuracy. Abbreviation: TEM, weekly average temperature; RHU, weekly average relative humidity; PRE, weekly cumulative precipitation.

Overall cumulative risk of meteorological exposure

In the DLNM, the median value of each meteorological factor was set as the reference to estimate the RR and 95% CI for each variable. Figure 4 presents the overall cumulative effects of meteorological factors on HFRS risk. Temperature (Fig. 4A) exhibited an approximately symmetric unimodal association with HFRS risk over the 26-week lag period. The risk peaked at around 13.66 °C (RR = 1.77, 95% CI: 0.92–3.39). Although the point estimate indicates a higher risk, the association was not statistically significant at the peak (as indicated by the black solid line). Elevated risks (RR > 1) were observed within the moderated temperatures range (9.02 –18.75 °C), while both extremely low and high temperatures were associated with reduced risks. Precipitation (Fig. 4B) displayed a distinct inverse J-shaped association. The highest risk was observed at very low precipitation levels (peak at 23.50 mm) and decreased rapidly as precipitation increased. While the curve shows an upward trend at extremely high precipitation levels (> 200 mm), the estimates are statistically unstable with very wide confidence intervals, likely due to the sparsity of data points (as shown in the rug plot) in this range. Therefore, the risk at extreme precipitation should be interpreted with caution. Relative humidity (Fig. 4C) showed a right-skewed unimodal pattern over the 16-week lag period. The risk increased with humidity up to a peak (around 76.72%), after which it declined. However, the peak risk was not statistically significant. Notably, low relative humidity (below 65.83%, RR = 0.66, 95% CI: 0.44–0.99) was associated with a statistically significant reduction in HFRS risk (RR < 1).

Fig. 4.

Cumulative exposure-response relationships between meteorological factors and HFRS. Shown are the relative risk (RR, solid lines) and 95% confidence intervals (shaded areas) for weekly average temperature (A), relative humidity (B), and cumulative precipitation (C). The reference levels (RR = 1, indicated by horizontal dashed lines) are set at 18.75 °C, 23.95 mm, and 81.97%, respectively. Vertical dashed lines indicate the values associated with the highest risk (peaks). Rug plots at the bottom of each panel show the distribution of observed data.

In summary, moderate temperatures and low precipitation levels appear to be the primary meteorological drivers for HFRS transmission in Yichun, whereas extreme weather conditions (very high heat, extreme rainfall) do not show a robust positive association with incidence.

Effects of meteorological factors across exposure levels and lag weeks

Fig. 5 presents the distributed lag patterns associated with different intensity levels (the 1st, 25th, 75th, and 99th percentiles) of each meteorological factor on HFRS risk at various lag weeks. For low temperature (1st and 25th percentiles), the risk decreased initially, bottomed out around lag week 13, and then rebounded. Both the 1st and 25th percentiles showed peak RR at lag 0 and lag 26. Notably, at lags 25–26, the risk associated with moderately low temperatures (10.01 °C) was significantly elevated (RR > 1; Supplementary Table S1), indicating a delayed adverse effect. In contrast, high temperatures (75th and 99th percentiles) were associated with a reduced risk (RR < 1) during the early lag weeks. However, this protective effect diminished over time; the risk gradually increased as the lag period extended, approaching or exceeding the reference level (RR ≈ 1) by the end of the lag period (lag 18).

Fig. 5.

Plots of relative risk (RR) by lag at specific percentiles of meteorological factor. Shown are the RR of HFRS across lag weeks for the 1st, 25th, 75th, and 99th percentiles of temperature, relative humidity, and precipitation; points represent estimates and error bars denote confidence intervals. The reference levels (RR = 1) are 18.75 °C, 23.95 mm, and 81.97%, respectively.

For precipitation, the effects varied by intensity and lag duration. For low rainfall (1st and 25th percentiles) and moderately high rainfall (75th percentile), the relative risk peaked within the current week (lag 0) but was not statistically significant. As the lag period extended, the risk gradually declined, shifting towards a protective effect (RR < 1, Supplementary Table S2), particularly evident during lag weeks 3–15. In contrast, extremely high precipitations (99th percentiles) exhibited a distinct delayed pattern: while the immediate risk was not elevated, a statistically significant and stable protective effect emerged at lag weeks 13 to 16, where the confidence intervals fell entirely below the null value.

For relative humidity, a statistically significant protective effect was observed at the 1st percentile (extremely low humidity). Specifically, the relative risks were significantly less than 1 between lag weeks 3 and 10, as the 95% confidence intervals fell entirely below the null value (Supplementary Table S3). For the 25th percentile (moderately low humidity), relative risks were marginally above 1.0 during the first 8 weeks but lacked statistical significance. In contrast, for low (25th percentile), high (75th percentile), and extremely high (99th percentile) humidity levels, the associations were generally not statistically significant, with confidence intervals mostly crossing the null line across the lag periods.

Interactive effects of meteorological factors on HFRS

GAMs were employed to investigate the interaction effect of meteorological factors on HFRS incidence (Fig. 6). The results revealed no statistically significant two-way interactions among the meteorological variables (all P > 0.05; Supplementary Tables S4-S6). Specifically, the interaction between temperature and relative humidity was not significant (P = 0.172; Supplementary Table S4), nor was the interaction between relative humidity and precipitation (P = 0.518; Supplementary Table S6). The interaction between temperature and precipitation also failed to reach statistical significance (P = 0.084; Supplementary Table S5). While the three-dimensional interaction surfaces in Fig. 6 show some visual patterns—such as potentially higher risk under conditions of low temperature combined with high humidity (Fig. 6A), elevated risk at high humidity particularly under low precipitation levels (Fig. 6C), and increased risk with concurrent low temperature and low precipitation (Fig. 6B)—these apparent patterns were not statistically significant. The lack of significant interactions suggests that the effect of one meteorological variable on HFRS incidence does not substantially depend on the levels of the other variables in the model. Instead, the individual effects of meteorological factors, particularly temperature, appear to drive the observed associations without strong synergistic or antagonistic interactions among them.

Fig. 6.

The effect interactions of the associations between meteorological factors and HFRS incidence. Abbreviation: HFRS, hemorrhagic fever with renal syndrome.

Sensitivity analysis

The results of the sensitivity analyses confirmed the robustness of our primary findings. As detailed in Supplementary Figures S1 and S2, the estimated associations remained stable across variations in the df for the cross-basis functions. When the df for the exposure dimension increased to 5, the effects of temperature, precipitation, and relative humidity became more unstable at the boundaries, as evidenced by wider confidence intervals. Furthermore, the strong consistency between models with and without additional meteorological covariates (Supplementary Figure S3 vs. Figure 3) provided additional support for the reliability of our results.

Regarding the potential collinearity between precipitation and relative humidity, we compared the cumulative exposure-response curves of temperature across the full model and the reduced models (excluding PRE or RHU individually). As shown in Supplementary Figure S4, the overall non-linear patterns remained highly consistent. The peak risk for temperature consistently occurred around 14 °C across all three models. Although excluding RHU slightly attenuated the peak relative risk and excluding PRE slightly elevated it, these variations did not substantially distort the main exposure-response relationship, indicating that the simultaneous inclusion of PRE and RHU did not introduce significant bias.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Shu Yang: Data curation, Formal analysis, Writing-original draft, Writing- review & editing. Xiaobo Liu: Data curation, Formal analysis. Xin Qi: Formal analysis, Methodology. Hui Li: Conceptualization. Guihua Zhuang: Resources, Conceptualization, Writing – review & editing.

Funding

This study was supported by the Science and Technology Bureau of Nanchang City, China (grant number 2020-133-18). The funder had no role in study design, data collection, and analysis, decision to publish, or preparation of the manuscript.

Data availability

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.