A Geographic Information System Approach to Map Tick Exposure Risk at a Scale for Public Health Intervention

Abstract

Tick-borne disease control and prevention have been largely ineffective compared to the control of other vector-borne diseases. Although control strategies exist, they are costly or ineffective at large spatial scales. We need tools to target these strategies to places of highest tick exposure risk. Here we present a geographic information system (GIS) method for mapping predicted tick exposure risk at a 200 m by 200 m resolution, appropriate for public health intervention. We followed the approach used to map tick habitat suitability over large areas. We used drag-cloth sampling to measure the density of nymphal blacklegged ticks (Ixodes scapularis, Say (Acari: Ixodidae)) at 24 sites in Addison and Rutland Counties, VT, United States. We used a GIS to average habitat, climatological, land-use/land-cover, and abiotic characteristics over 100 m, 400 m, 1,000 m, and 2,000 m buffers around each site to evaluate which characteristic at which buffer size best predicted density of nymphal ticks (DON). The relationships between predictor variables and DON were determined with random forest models. The 100 m buffer model performed best and explained 37.7% of the variation in DON, although was highly accurate at classifying sites as having below or above average DON. This model was applied to Addison County, VT, to predict tick exposure risk at a 200 m resolution. This GIS approach to map predicted DON over a small area with fine resolution, could be used to target public health campaigns and land management practices to reduce human exposure to ticks.

Keywords: Ixodes scapularis; Lyme disease; bioclimatic modeling; remote sensing; risk-mapping.

Fig. 1.

Locations of 24 sampling sites, indicated by stars, in Rutland and Addison Counties, Vermont. Addison County is highlighted as this is the area over which tick density is predicted. Inset map shows the regional location of the study area.

Fig. 2.

Example of the arrangement of tick sampling transects (solid lines) and 100 m predictive buffer regions (dotted lines) in A) Addison County sites and B) Rutland County sites. The center of each site, represented by the dot, is the centroid of all sampling transects at that site. Both subfigures have the same scale. Only the 100 m predictive buffer is shown, the larger radii ones are excluded for clarity.

Fig. 3.

Summary of steps taken to analyze remotely sensed data before and after random forest analysis. A) Before random forest analysis we summarized large-scale data products within the 4 buffer sizes around our field sites (products and their summary statistics are listed in Table 2). B) After random forest analysis we summarized these same data products to a 200 m² fishnet of Addison County, input these variables into the random forest model and generated a table of predicted DON at the 200 m by 200 m scale in Addison County.

Fig. 4.

Observed versus predicted DON based on 100 m buffer random forest model. This was the best performing model and explained 37.7% of variation in DON. The solid line gives observed equals predicted, i.e., if the model predicted exact values. The dotted line is the observed versus predicted trendline. The thin horizontal and vertical lines give the mean DON. Points in the lower-left square were correctly predicted to have below-average density and those in the upper-right were correctly predicted to have above-average density.

Fig. 5.

Partial dependence plots for the six most-predictive variables explaining DON at the 100 m scale. These plots show the marginal effect of each variable on predicted DON. See Table 3 for the importance value for each variable.

Fig. 6.

Partial dependence plots for the six most-predictive variables explaining DON at the 2000 m scale. These plots show the marginal effect of each variable on predicted DON. See Table 3 for the importance value for each variable.

Fig. 7.

Predicted DON in Addison County, VT at 200 m resolution. Predictions are based on extrapolations from the 100 m buffer random forest model. The model was applied to the forested areas of Addison County, VT because all sampling sites were forested. The upper inset shows predicted DON on Snake Mountain, a population recreation area. The lower inset shows the regional context for the map.