Seasonality of agricultural exposure as an important predictor of seasonal yellow fever spillover in Brazil

Yellow fever (YF), a zoonotic arbovirus, affects both humans and non-human primates (NHPs) in Africa and South America. In South America, YF transmission occurs in two cycles: sylvatic and urban. The sylvatic cycle involves transmission between NHPs via mosquitoes, with humans as incidental hosts. The urban cycle involves *Aedes aegypti* mosquitoes and can lead to devastating epidemics. In Brazil, since 1942, all YF cases have been linked to the sylvatic cycle, primarily in the North and Northwest. However, since 1998, there's been a significant expansion of risk areas, resulting in major outbreaks. Previous studies on YF seasonality have focused on climate, but a substantial proportion of human cases involve individuals in agriculture, a highly seasonal activity. This study aims to investigate the drivers of seasonal YF transmission in Brazil by examining the relative importance of climatic and agricultural seasonality in predicting YF occurrence in both humans and NHPs. The hypothesis is that the seasonality of agricultural activities, representing a proxy for exposure risk to the sylvatic cycle, significantly influences seasonal YF transmission patterns.

While the seasonality of YF has been previously noted, the underlying mechanisms remain unclear. Climate variations, influencing vector populations and transmission suitability, have been used to explain this seasonality and even for forecasting. However, given the significant involvement of agricultural workers in YF cases in Brazil, the seasonality of human exposure needs further consideration. Although the connection between agriculture and disease transmission is well-established, the role of agricultural seasonality in altering zoonotic disease transmission is understudied. Existing literature explores landscape changes affecting human exposure, vector composition, and reservoir host behavior, but lacks focused investigation on how agricultural seasonality directly impacts disease transmission. This study fills this gap by directly examining the impact of the seasonality of agricultural activities on YF spillover risk.

This study used random forest models to predict the occurrence of YF in humans and NHPs in Brazil. Data on YF reports (human and NHP) at the municipality level from 2003-2018 were obtained from the Brazilian Ministry of Health. These data were aggregated monthly. Four covariate groups were used: (1) agricultural output (number of farms producing eight crop types), (2) host demographics (number of NHP species, proportion of human population in agriculture, rural human population), (3) agricultural seasonality (planting and harvesting of eight crops), and (4) climate/vegetation seasonality (rainfall, temperature, and Enhanced Vegetation Index (EVI)). All possible combinations of these covariates were used to create 15 different models. Model performance was evaluated using the out-of-sample area under the receiver operating characteristic curve (AUC) and the Brier score. Spatial-block bootstrapping was used for out-of-sample validation. Variable importance was assessed using permutation importance and partial dependency plots. The analysis investigated the ability of the models to capture the monthly seasonality of YF reports and identified the most influential covariates.

YF reports showed strong seasonality in both humans and NHPs, although with different patterns. Human reports peaked in January, while NHP reports showed a less pronounced peak. Models including agricultural seasonality substantially outperformed climate-only models in predicting human YF reports (aggregate monthly correlation: 0.99 vs 0.80). The best-performing model (OHAC) included all covariate groups. Out-of-sample predictive performance, assessed using spatial-block bootstrapping, was comparable to in-sample performance for human reports but slightly lower for NHP and both combined reports, suggesting potential overfitting in the latter. Variable importance analysis revealed that climate covariates (temperature and rainfall) were highly important in all models, with agricultural seasonality particularly influential in explaining human YF reports. Specific crops, such as rice harvesting and peanut planting, were significantly associated with increased YF reporting. Geographical distribution analysis revealed YF reports across Brazil except for the Northeast, with hotspots in the Southeast, West and Amazonian states. The model OHAC accurately captured these spatial patterns.

The study's results demonstrate the crucial role of agricultural seasonality in predicting human YF cases in Brazil, complementing the established influence of climate. The superior predictive power of models incorporating agricultural seasonality suggests that increased human-environment interaction during agricultural activities drives exposure to the sylvatic cycle, leading to spillover events. The identified association between specific crops and increased YF risk provides actionable insights for targeted interventions. The findings challenge the existing reliance on climate-based models for YF prediction and highlight the need for integrated approaches considering both climatic and anthropogenic factors. The strong seasonality in YF reports, largely influenced by agriculture, raises concerns about potential underestimation of YF transmission outside the traditional surveillance period, especially in endemic zones. This has implications for disease surveillance, prevention, and response strategies.

This study provides strong evidence for the importance of agricultural seasonality in driving seasonal yellow fever spillover in Brazil. The superior performance of models integrating agricultural data demonstrates the need to consider human interaction with sylvatic habitats. Targeting interventions towards high-risk agricultural activities and time periods can optimize YF prevention and control efforts. Future research should focus on detailed investigation of specific agricultural activities and their interaction with sylvatic cycles at a higher spatial resolution to refine risk prediction and improve public health interventions.

The study's limitations include potential biases related to the surveillance of NHP YF cases, which varies across regions. The availability of data on agricultural activities was limited to eight crop types and the first administrative level, hindering a more granular analysis. The focus on presence/absence of reports instead of case numbers might not fully capture the true magnitude of transmission. Further geo-localization of reports and more detailed information on agricultural activities would enhance the study's accuracy and applicability.