Estimates of rabies risk from vampire bats to livestock in Colombia

Abstract

Objective.

The goal of this study is to share with the community a spatial model of vampire-bat rabies (VBR) spillover risk developed with the government of Colombia to better inform current anti-rabies vaccination efforts. This model was designed to determine likely areas where vaccination should be prioritized across Colombia.

Materials and methods.

This spatial analysis designed by the Instituto Colombiano Agropecuario (ICA) sought to develop a predictive data-driven model, with information obtained from passive surveillance. Predictive identification of VBR hotspots of transmission risk within Colombia was made by incorporating VBR epidemiological data with environmental and landscape variables.

Results.

This study identified clustered patterns of the current geographic risk of VBR in Colombia, which were used to inform a national effort for rabies vaccination. Areas of VBR spillover risk were associated with distance to forested area, climate, and high cattle density in the Caribbean and Andean regions. At least 87% of Colombia could be identified as “at risk” to VBR spillover based on environmental conditions. That is, the majority of the country has conditions similar to sites were VBR spillover to livestock has occurred in the past. The model identified 88 municipalities as being at high risk for VBR spillover (spillover risk ≥0.70). More years of vaccination data are needed for an accurate evaluation of the rabies vaccination program in cattle in Colombia.

Conclusions.

This study provides a refined VBR transmission-risk map, which reveals key geographic to be prioritized for vaccination and govertment effort for VBR control.

INTRODUCTION

Vampire-bat rabies (VBR) is a well-documented and impactful bat-borne pathogen in Latin America which causes a significant economic burden (1,2,3). Thousands of livestock are lost annually in the region due to VBR, with at least 133,801 infections being reported in cattle since 1970 (3,4,5). Despite its endemicity, outbreaks persist across Latin America, especially in countries like Mexico, Brazil, Perú, Ecuador, and Colombia (5,6,7,8,9). In fact, nearly 3000 individual outbreaks of VBR in cattle have been reported in Colombia alone (5). While incidence remains stable in some areas, outbreaks in previously unaffected regions suggest a potential spatial expansion of VBR (5,10,11). For instance, Mexico has experienced ongoing outbreaks in historical VBR-free municipalities, with the main bat reservoir, Desmodus rotundus, potentially expanding its geographic range northward towards the continental US (5,12,13). Similarly, Peru has reported wave-like spatial expansion of VBR in inter-Andean valleys where the disease has previously been absent (2). In Colombia, passive surveillance revealed that VBR outbreaks doubled between 2010 and 2019, particularly in the Caribbean region, suggesting an actual rise in spillover transmission (14). A rabies vaccination program was initiated in Colombia by the government in 2015, and since then the number of rabies annually seems to have an historical decrease (14). Nevertheless, specific locations where VBR occurs is unclear and identification of sites to vaccinate is largely based on institutional conjecture and expert opinion, which limits the precision and efficacy of livestock-vaccination programs. Surveillance for rabies virus within D. rotundus has been limited, as surveillance and monitoring has been historically targeted toward human and livestock cases (12,15). Furthermore, VBR has been found to be rare in D. rotundus where infections tend to be nonlethal and immunize bats (16,17).

In tropical countries of Latin America, such as Colombia, VBR occurs in endemic and epidemic foci year round (15). Traditional livestock vaccination strategies typically rely on responsive rather than proactive vaccination efforts (13,16). As such, it is critical to develop public health and veterinary health policies that promote proactive vaccination informed by risk analysis derived from epidemiological and environmental analysis (18). Data-driven risk analysis to guide vaccination schemas may, however, have cultural, political, and economic challenges, including shortages of vaccine, logistical limitation to access at-risk populations, and limited funding (19).

Livestock vaccination efforts against VBR in Colombia are led by the Instituto Colombiano Agropecuario (ICA) (17). Since 2003, ICA has implemented vaccination in areas considered to be “VBR hotspots of risk” based on retrospective VBR outbreak data at the administrative level (17). Aiming to develop a quantitative proxy of VBR risk of transmission to livestock in Colombia, ICA led a spatial analysis in an inter-institutional agreement with the Unidad de Planificación Rural Agropecuaria (UPRA), an entity that developed the technical execution of the model (variable management and model calibration), while the ICA supervised, guided, and evaluated the results for its application in vaccination programs. This study describes ICA’s recent effort to develop and implement a VBR-risk map, incorporating VBR epidemiological data with environmental variables. This map has been in operational use since 2021 to inform a rabies vaccination program in Colombia. The article illustrates a stakeholders’ effort for data-driven rabies control in livestock in Latin America as an example of precision epidemiology.

MATERIALS AND METHODS

Outbreak data collection.

VBR outbreaks data at the locality level from ICA official records were used to map VBR outbreaks across Colombia from 2017–2019. Locations of cattle farms with VBR spillover in geographic coordinates format were used for the modeling of VBR spillover risk. To address possible sampling or surveillance bias and spatial autocorrelation within the outbreak location data, the spillover locations were resampled to one per pixel following the pixel size of environmental variables (19). Pixel size was set using the highest resolution of the available data (500 meters). The study area was defined using World Geodetic System 1984 (WGS84) reference system and clipped to a rectangular study area that encompassed the entirety of continental Colombia (11.72°N, 3.80°S, 77.37°W, 67.67°W). Data within the study area that fell within the Pacific Ocean or Caribbean Sea were classified as “NA”, and therefore removed from consideration during the model calibration process.

A background variable selection was conducted based on known ecological determinants of VBR spillover propensity in Colombia, and based on on-the-ground field expertise from local authorities (17,20,21). As such, no formal correlation assessment was conducted to eliminate variables since one of the goals of this exploration was to evaluate the relative contribution of these variables regardless of collinearity. Landscape and climate variables were collected and resampled at 500 meter spatial resolution from the Instituto de Hidrología, Meteorología y Estudios Ambientales (IDEAM)) and the Instituto Geográfico Agustín Codazzi (IGAC) via the Unidad de Planificacion Rural Agropecuaria (UPRA) (Table 1). Landscape and climate variables included distance to waterbody (m), elevation (m), potential evapotranspiration (mm), relative humidity (%), annual precipitation (mm), solar radiation (kWh/m²), and annual temperature (°C). Furthermore, variables relative to forest composition were collected from the IDEAM scaled LANDSAT forest map, including distance to forested area and distance to location of deforestation. Cattle density (head of cattle per square kilometer) was derived from the ICA cattle census database, which accounts for surveyed number of cattle per administrative area at the municipality level (22). Cattle census data were used from the year 2019. All background variables were resampled to 500 meter resolution using the terra and raster packages in R statistical software version 4.1.0 and R studio version 2023.09.01 (23,24) to align rasters and standardize resolution prior to model calibration.

Table 1.

Background environmental variables. List of background environmental variables used for each modeling effort. All variables were collected or resampled to 500-meter resolution.

Variable Name	Variable Unit	Source
Temperature	°C	Institute of Hydrology, Meteorology and Environmental Studies (Ideam). (2014). Distribution map of multi-year average annual mean temperature 1981–2010, scale 1: 100,000. Accessed December 2022.
Solar Radiation	kWh/m2	Institute of Hydrology, Meteorology and Environmental Studies (Ideam). (2014), Distribution map of the annual average daily solar brightness (hours) multiannual average 1981–2010, scale 1:100,000. Accessed December 2022.
Precipitation	mm	Institute of Hydrology, Meteorology and Environmental Studies (Ideam). (2014). Distribution map of average annual precipitation 1981–2010, scale 1:100,000. Accessed December 2022.
Relative Humidity	%	Institute of Hydrology, Meteorology and Environmental Studies (Ideam). (2014). Distribution map of annual average relative humidity (%) scale 1:100,000. Accessed December 2022.
Potential Evapotranspiration	mm	Institute of Hydrology, Meteorology and Environmental Studies (Ideam). (2014). Climatological Atlas of Colombia - Layer of potential Evapotranspiration zones. Scale 1:100,000. Accessed December 2022.
Elevation	m	Agustín Codazzi Geographic Institute (IGAC). (2014). Digital terrain model, spatial resolution: 90 meters. Accessed December 2022.
Distance to Waterbody	m	Agustín Codazzi Geographic Institute (IGAC). (2019). Basic cartography, scale 1:100,000. Accessed December 2022.
Distance to Deforestation	m	Institute of Hydrology, Meteorology and Environmental Studies (Ideam). (2016). Non-Forest Forest Map Colombia - Continental Area (LANDSAT Fine Scale) year 2010, version 5. Accessed December 2022.
Distance to Forest	m	Institute of Hydrology, Meteorology and Environmental Studies (Ideam). (2016). Non-Forest Forest Map Colombia - Continental Area (LANDSAT Fine Scale) year 2010, version 5. Accessed December 2022.

Model calibration, evaluation, and projection were completed using MaxEnt version 3.4.4. (25). MaxEnt is a commonly-used correlative ecological niche modeling algorithm which utilizes presence-background comparison, and as such does not require true absence locations and thus more accurately abides by the nature of the available epidemiological data (26,27,28). Occurrence points were portioned into 70% training points and 30% testing points with a random seed. Parameterizations within MaxEnt were set using auto-feature classes and a regularization multiplier of three (25). MaxEnt regularization was set to three to mitigate output overfit, as can happen with default settings (29). MaxEnt output features were also utilized to create response curves, and a Jacknife resampling of all landscape and climate variables measured variable contribution to the final model. Model evaluation was conducted using 100 replicates of calibration and cross-validation. A binomial test was also used with 30% of the data retained for testing and calculated the omission rate of the final model using the kuenm package in R version 4.1.0 (30).

Models were summarized with a median of the cumulative log outputs raster of all 100 replicates, and were then combined in ArcGIS Pro software (31). As such the final output consisted of a continuous span of values which can be interpreted as the similarity of the projected pixel to conditions in sites of known spillover (32). This value can also be interpreted as “VBR spillover risk” based on these assumptions, and will be referend to as such from here onward. In ArcGIS pro software, the final raster was visualized using seven bins of projected spillover risk values from low (0.05) to high (0.98). The average minimum training presence of all 100 replicates was used as a threshold to classify if pixels had sufficient projected spillover risk to be considered “at risk” for VBR. This was based on the assumption that the lowest projected spillover risk value at which a spillover event has previously occurred can be used as a minimum benchmark for which VBR spillover is plausible. Any pixel which had a spillover risk value greater than 0.70 was classified as “high risk” based on this same assumption of similarity to known spillover locations. These thresholds were based on previous assessments of bovine rabies distribution in Colombia, and the institutional standards established therein pertaining to VBR spillover risk projections (9,14,21).

RESULTS

Projected VBR spillover risk distribution.

The final model had an omission rate of 0.02 and was statistically significant per the binomial test (p<0.01), suggesting that independent occurrence data were predicted better than by chance. The area under the receiver operating characteristic curve (AUC) was 0.87. Relative variable contribution to the MaxEnt models revealed that distance to forest and cattle density were the two variables which contributed the most to regularization gain. Average percent variable contributions across all 100 replicates included distance to forest (42.8%), temperature (17.5%), annual precipitation (15.2%), cattle density (14.3%), solar radiation (4.24%), potential evapotranspiration (2.8%), and elevation (2%) (Figure 1). The remaining variables (i.e., distance to waterbody, relative humidity, and distance from deforestation) contributed less than 1% to the overall distribution of projected VBR spillover risk in Colombia. In summary, VBR spillover risk was higher in areas close to forests, in warm climates, and with high cattle density.

Figure 1.

Response Curves of Risk for VBR in Response to Variation in Environmental Variables. Response curves denoting how each background variable is related to suitability prediction. Predicted probability of presence (suitability) changes as each background variable varies, and these curves indicate the marginal effect of each variable individually. Each curve shows the mean response of the associated variable for the 100 replicate MaxEnt runs (red) and the mean +/− one standard deviation (blue).

In terms of model output, 87.1% of Colombia could be identified as being at some risk for VBR spillover based on the minimum training presence threshold used to define spillover risk (spillover risk ≥0.03). That is, the majority of the country has conditions for VBR spillover risk similar to other sites were VBR spillover to livestock has occurred in the past. In a more applied context, geographic patterns of projected VBR spillover risk indicate that more risk exists in northern portions of Colombia (spillover risk ≥0.70), particularly in the Caribbean region (Figure 1). We identified 88 municipalities as having high risk for VBR spillover (Figure 2, Table 2). Moderate to low risk extended along the Andes Mountain range throughout Colombia into more central areas of the country (spillover risk 0.5–0.7).

Figure 2.

Projected rabies virus spillover risk map of Colombia. A. Categorical visualization of projected VBR spillover risk as represented by the median raster cumulative log output from all 100 replicates of the model. Green indicates low projected VBR spillover risk. Red indicates high projected VBR spillover risk. Numbers associates with each colored risk category are representative of MaxEnt value outputs and can also be interpreted as environmental similarity of areas with previous VBR spillover relative to the background predictor variables. B. Continuous projected VBR spillover risk as represented by the median raster cumulative log output from all 100 replicates of the model. Yellow indicates low projected VBR spillover risk. Red indicates high projected VBR spillover risk, and denotes areas that should be prioritized for VBR prevention initiatives. Numbers associates with each colored risk category are representative of MaxEnt value outputs, and can also be interpreted as similarity between site of known VBR spillover and background predictor variables.

Table 2.

High VBR spillover risk municipalities.

Department	Municipality	AP VBR
Sucre	Morroa	0.93
Magdalena	Chivolo	0.92
Sucre	Sincelejo	0.92
Sucre	Toluviejo	0.92
Bolívar	Turbaná	0.91
Magdalena	Ariguaní	0.91
Cesar	San Diego	0.9
Sucre	Corozal	0.9
Sucre	Sampués	0.9
Sucre	San Juan de Betulia	0.9
Bolívar	Turbaco	0.89
Cesar	Gamarra	0.89
Magdalena	Pivijay	0.89
Magdalena	Plato	0.89
Sucre	Sincé	0.89
Cesar	Astrea	0.88
Cesar	Bosconia	0.88
Magdalena	San Sebastian de Buenavista	0.88
Sucre	San Antonio de Palmito	0.88
Sucre	San Onofre	0.88
Bolívar	El Guamo	0.87
Cundinamarca	Girardot	0.87
Cundinamarca	Guataquí	0.87
Magdalena	Santa Ana	0.87
Sucre	Colosó	0.87
Sucre	Galeras	0.87
Sucre	Los Palmitos	0.87
Sucre	Tolú	0.87
Magdalena	San Zenón	0.86
Cesar	Aguachica	0.85
Magdalena	Pedraza	0.85
Tolima	Ambalema	0.85
Bolívar	Arjona	0.84
Bolívar	Mahates	0.84
Bolívar	María la Baja	0.84
Boyacá	Chinavita	0.84
Cesar	El Paso	0.84
Magdalena	El Piñón	0.84
Magdalena	Tenerife	0.84
Bolívar	San Estanislao de Kostka	0.83
Bolívar	Santa Catalina	0.83
Tolima	Flandes	0.83
Bolívar	Cartagena de Indias	0.81
Cesar	Río de Oro	0.81
Cundinamarca	Agua de Dios	0.8
Bolívar	Río Viejo	0.79
Bolívar	Zambrano	0.79
Cesar	El Copey	0.79
Cundinamarca	El Colegio	0.79
Sucre	Ovejas	0.79
Bolívar	El Carmen de Bolívar	0.78
Córdoba	San Andrés de Sotavento	0.78
Cesar	Tamalameque	0.78
Cundinamarca	Beltrán	0.78
Magdalena	Cerro de San Antonio	0.78
Sucre	Chalán	0.78
Cundinamarca	Nilo	0.76
Tolima	Armero	0.76
Tolima	Espinal	0.76
Bolívar	San Martín de Loba	0.75
Bolívar	Soplaviento	0.75
Cesar	La Gloria	0.75
Cundinamarca	Guayabal de Síquima	0.75
Córdoba	Chinú	0.74
Córdoba	Ciénaga de Oro	0.74
Córdoba	Moñitos	0.74
Córdoba	San Bernardino de Sahagún	0.74
Cundinamarca	San Juan de Río Seco	0.74
Magdalena	Remolino	0.74
Tolima	Melgar	0.74
Bolívar	San Juan Nepomuceno	0.73
Bolívar	Simití	0.73
Cesar	Chimichagua	0.73
Cundinamarca	Bituima	0.73
Cundinamarca	Tocaima	0.73
Tolima	Coyaima	0.73
Tolima	Natagaima	0.73
Tolima	Piedras	0.73
Boyacá	Guateque	0.72
Córdoba	Puerto Escondido	0.72
Cesar	Becerril	0.72
Cundinamarca	Jerusalén	0.72
Tolima	Venadillo	0.72
Cesar	Pelaya	0.71
Bolívar	San Jacinto	0.7
Cundinamarca	Chaguaní	0.7
Cundinamarca	Vianí	0.7
Cundinamarca	Zipacón	0.7
Tolima	Guamo	0.7

AP VBR= Average Projected VBR Spillover Risk;

The Average projected VBR spillover risk for all municipalities which exceeded 0.7 in Colombia. Numeric values are representative of average MaxEnt value outputs across all 100 replicates and can also be interpreted as similarity to areas of known VBR spillover relative to the background predictor variables.

DISCUSSION

This article describes a MaxEnt ecological niche model of rabies transmission from vampire bats to cattle developed and in use by ICA in Colombia. In Colombia, VBR is a burden on the livestock industry despite ongoing vaccination campaigns implemented by the Animal Health Direction of ICA (8,13,17). Following the need for estimation risk models, this spatial modeling constitutes an inter-institutional effort led by the ICA, in agreement with the UPRA, whose technical participation allowed the construction and calibration of the predictive model to map and quantify risk of VBR spillover in Colombia after the increased number of outbreaks in 2014. The MaxEnt model, which yielded a geographic and numeric representation of projected VBR spillover risk, was mainly influenced by distance to forest, climate, and cattle density. The resulting map of geographic VBR spillover risk hotspots showed that northern portions of Colombia in the Caribbean region have a higher predicted risk (Figure 2). The Caribbean region was identified as a priority region for the national vaccination program. This map was then used by ICA officials to inform the rabies vaccination plan in tandem with other logistics and security factors.

Distance to forest, climate (i.e., temperature and precipitation), and cattle density were the primary drivers of VBR spillover risk based on the MaxEnt model. The spillover and environmental data indicate that predicted spillover risk for VBR is highest in areas that are close to forests, in warm climates, with high cattle density. As such, landscape change for agricultural reasons or development in warm tropical areas such as the Caribbean region could contribute to the persistent VBR burden in Colombia (Figures 1 and 2). Forest conversion for livestock expansion, likely increasing livestock densities at the same time, facilitate contact with vampire bats. While temperature and precipitation had a lesser influence on the model, temperature may be associated with vampire bats distribution in warmer areas like the coasts, eastern plains, and inter-Andean valleys (33,34). Precipitation might be linked to dispersal patterns, potentially increasing during dry seasons as previously suggested (13). These factors highlight the urgency for improved vaccination and monitoring logistics, especially in the Caribbean region where VBR has extended throughout endemic and non-endemic areas reaching rabies-free municipalities. Continual monitoring will also be necessary, as many of the factors associated with VBR spillover risk identified here are not static, and may change over time. As such more longitudinal assessments may be necessary. The Caribbean region emerged as an area with the highest projected VBR spillover risk, followed by the Andean region (Figure 2). These findings align with historical land-use changes in Colombia, which includes deforestation concentrated in the Andean and Caribbean regions (35). The higher predicted risk in the Caribbean region also coincides with high livestock densities and frequent VBR outbreaks previously identified in this system (11,13,14).

The risk map developed and used by ICA to inform the vaccination program suggests prioritizing areas with high likelihood of VBR spillover risk to cattle, e.g., predicted risk >0.7. The risk estimation, however, is not static and landscape conversion and changes in cattle density are expected to modulate the risk estimate. Regular assessment is essential for rabies prevention to capture temporal variation in VBR outbreaks and the effects of intervention on the size, direction, frequency, and location of spillover events. While annual level assessments would continue to be beneficial for VBR management, biennial or even quarterly assessment of outbreaks may allow for a more seasonal assessment of outbreak patterns. Nevertheless, resource and time limitation may prevent this type of temporal refinement in data collection. In conclusion, this article describes a refined map of VBR potential transmission used in Colombia as an example of precision epidemiology to guide prevention efforts. With time, these modeling efforts may contribute to transition from reactive to proactive vaccination schemes in Colombia. Future inclusion of more contemporary vaccination and outbreak data (i.e., after the year 2019) may also allow robust assessments of the prevention of VBR spillover risk.