Distribution dynamics and urbanization-related factors of Hantaan and Seoul virus infections in China between 2001 and 2020: A machine learning modelling analysis

Abstract

Objectives: The epidemical and clinical features of distinct hantavirus infections exhibit heterogeneity. However, the evolving epidemics and distinct determines of the two hantavirus infections remain uncertain.

Methods: Data on hemorrhagic fever with renal syndrome (HFRS) cases and genotyping were collected from multiple sources to explore the distribution dynamics of different endemic categories. Four modelling algorithms were used to examine the relationship between infected hantavirus genotypes in HFRS patients, as well as assess the impacts of urbanization-related factors on HFRS incidence.

Results: The number of cities dominated by Hantaan (HTNV) and Seoul (SEOV) viruses was projected to decrease between two phases, while the mixed endemic cities increased. Patients with SEOV infection predominantly presented gastrointestinal symptoms. The modeling analysis revealed that built-up land and real GDP demonstrated the highest contribution to HTNV and SEOV infections, respectively. The impact of nightlight index and park green land was more pronounced in HTNV-dominant cities, while cropland, impervious surface, and floor space of commercialized buildings sold contributed more to HFRS incidence in SEOV-dominant cities.

Conclusions: Our findings fill a gap for the three endemic categories of HFRS, which may guide the development of targeted prevention and control measures under the conditions of urbanization development.

Keywords: HFRS; Risk factor; SHAP; Urbanization; XGBoost.

Fig. 1

Geographic distribution dynamics and conversion of three endemic categories for HFRS in the mainland of China, 2001–2020. The observed and projected geographic distribution dynamics of the three endemic categories are presented by two phases: 2001–2010 (A, C) and 2011–2020 (B, D). The grey background indicates the areas with no genotyping data or no risk, and the other three colored backgrounds indicate the observed or projected areas of three endemic categories. The observed types of “HTNV-dominant” and “SEOV-dominant” indicate the areas where only HTNV and SEOV are reported, respectively, while type of “Mixed” indicates the areas where the two hantavirus genotypes are both reported. The predicted types of “HTNV-dominant”, “SEOV-dominant” and “Mixed” indicate the areas where the predictive proportions of HTNV are above 80 %, below 20 % or lying between, respectively. The regional conversion among the three endemic categories between the two phases is shown in (E).

Fig. 2

Modelling analysis of infections with different hantavirus genotypes based on clinical characteristics. (A) ROC curves of different models. (B) PR curves of different models. (C) Comparison of five evaluation indicators for different models. (D) Importance and effects of major variables indicated by SHAP values based on the optimal model (XGBoost). The variables are ranked in the importance according to their global SHAP values from top to bottom, with the y-axis indicating different variables and the x-axis indicating the SHAP values. Colors from yellow to purple indicate feature values from low to high. ROC: receiver operating characteristic. PR: precision-recall. SHAP: shapley additive explanations. XGBoost: extreme gradient boosting. AUC: area under curve. RF: random forest. GBM: gradient boosting machine. GLM: generalized linear model. WBC: white blood cell. PLT: platelet.

Fig. 3

Pooled exposure-response curves between HFRS incidence and urbanization-related factors in HTNV-dominant and SEOV-dominant cities during 2001‒2020 based on XGBoost model. (A) Built-up land (lag = 3). (B) Built-up land (lag = 2). (C) Population density (lag = 0). (D) Population density (lag = 3). (E) Green land (lag = 3). (F) Green land (lag = 3). (G) Real GDP (lag = 3). (H) Real GDP (lag = 2). (I) Impervious surface (lag = 0). (J) Impervious surface (lag = 0). (K) Nightlight index (lag = 1). (L) Cropland (lag = 3). (M) Built−up land (lag = 0). (N) Impervious surface (lag = 2). (O) Park green land (lag = 3). (P) Floor space of commercialized buildings sold (lag = 0). Pooled exposure-response curves of HTNV-dominant and SEOV-dominant cities were plotted in red and blue lines, respectively, with their 95 % CI indicated by the shaded areas. The x-axis (plotted in log scale) indicates the observed values of the urbanization-related factors, while the y-axis indicates the SHAP value, with the scattered points indicating the distribution of fitting results for each city every year. GDP: gross domestic product. CI: confidence interval.