Predictors of bovine Schistosoma japonicum infection in rural Sichuan, China

Abstract

In China, bovines are believed to be the most common animal source of human schistosomiasis infections, though little is known about what factors promote bovine infections. The current body of literature features inconsistent, and sometimes contradictory results, and to date, few studies have looked beyond physical characteristics to identify the broader environmental conditions that predict bovine schistosomiasis. Because schistosomiasis is a sanitation-related, water-borne disease transmitted by many animals, we hypothesised that several environmental factors - such as the lack of improved sanitation systems, or participation in agricultural production that is water-intensive - could promote schistosomiasis infection in bovines. Using data collected as part of a repeat cross-sectional study conducted in rural villages in Sichuan, China from 2007 to 2016, we used a Random Forests, machine learning approach to identify the best physical and environmental predictors of bovine Schistosoma japonicum infection. Candidate predictors included: (i) physical/biological characteristics of bovines, (ii) human sources of environmental schistosomes, (iii) socio-economic indicators, (iv) animal reservoirs, and (v) agricultural practices. The density of bovines in a village and agricultural practices such as the area of rice and dry summer crops planted, and the use of night soil as an agricultural fertilizer, were among the top predictors of bovine S. japonicum infection in all collection years. Additionally, human infection prevalence, pig ownership and bovine age were found to be strong predictors of bovine infection in at least 1 year. Our findings highlight that presumptively treating bovines in villages with high bovine density or human infection prevalence may help to interrupt transmission. Furthermore, village-level predictors were stronger predictors of bovine infection than household-level predictors, suggesting future investigations may need to apply a broad ecological lens to identify potential underlying sources of persistent transmission.

Keywords: Buffalo; Cattle; China; Machine learning; Prevention and control; Schistosomiasis.

Fig. 1.

Village-level prevalence of schistosomiasis in bovines in 2007, 2010 and 2016. The prevalence of schistosomiasis infection in bovines is indicated for each village included in 2007, 2010 and 2016. The darker the shade, the higher the prevalence. While villages in the lower left corner of each map were found to have repeatedly high infection prevalence, other villages achieved 0% infection in bovines in later years (upper left and center right of maps), while still others emerged as high prevalence villages in 2010 (center of maps) and 2016 (upper left of maps) after years of low infection prevalence in bovines. Service Layer Credits: World Terrain Base Sources: Esri, USGS, NOAA. OpenStreetMap Data Extracts for China, Asia: Data/Maps Copyright 2018 Geofabrik GmbH and OpenStreetMap Contributors.

Fig. 2.

Variable importance rankings and direction of association for candidate predictors of bovine Schistosoma japonicum infection in 2007, 2010 and 2016. Variable importance rankings are based on a composite of mean decrease in accuracy scores for 10 random forest (RF) models for each model type (full, lean and sensitivity (Sens.)) and collection year. The direction of association was determined through logistic regression, using tertile categories for continuous variables to assess evidence for non-linearity. A P-value <0.2 was used to indicate evidence of a between-group difference and, when a between-group difference was found, the direction of association is indicated. See Supplementary Table S3 for detailed logistic regression results.

Fig. 2.

Variable importance rankings and direction of association for candidate predictors of bovine Schistosoma japonicum infection in 2007, 2010 and 2016. Variable importance rankings are based on a composite of mean decrease in accuracy scores for 10 random forest (RF) models for each model type (full, lean and sensitivity (Sens.)) and collection year. The direction of association was determined through logistic regression, using tertile categories for continuous variables to assess evidence for non-linearity. A P-value <0.2 was used to indicate evidence of a between-group difference and, when a between-group difference was found, the direction of association is indicated. See Supplementary Table S3 for detailed logistic regression results.

Fig. 2.

Variable importance rankings and direction of association for candidate predictors of bovine Schistosoma japonicum infection in 2007, 2010 and 2016. Variable importance rankings are based on a composite of mean decrease in accuracy scores for 10 random forest (RF) models for each model type (full, lean and sensitivity (Sens.)) and collection year. The direction of association was determined through logistic regression, using tertile categories for continuous variables to assess evidence for non-linearity. A P-value <0.2 was used to indicate evidence of a between-group difference and, when a between-group difference was found, the direction of association is indicated. See Supplementary Table S3 for detailed logistic regression results.

Fig. 3.

Changes in agricultural practices and the relationship between bovine Schistosoma japonicum infection and agricultural predictors over time. For each of the agricultural predictors included in this analysis, boxplots are used to represent the distribution of uninfected (white) and infected bovines, for household-level (left) and village-level variables (right) in 2007, 2010 and 2016.