A Landscape Epidemiological Approach for Predicting Chronic Wasting Disease: A Case Study in Virginia, US

Abstract

Many infectious diseases in wildlife occur under quantifiable landscape ecological patterns useful in facilitating epidemiological surveillance and management, though little is known about prion diseases. Chronic wasting disease (CWD), a fatal prion disease of the deer family Cervidae, currently affects white-tailed deer (Odocoileus virginianus) populations in the Mid-Atlantic United States (US) and challenges wildlife veterinarians and disease ecologists from its unclear mechanisms and associations within landscapes, particularly in early phases of an outbreak when CWD detections are sparse. We aimed to provide guidance for wildlife disease management by identifying the extent to which CWD-positive cases can be reliably predicted from landscape conditions. Using the CWD outbreak in Virginia, US from 2009 to early 2020 as a case study system, we used diverse algorithms (e.g., principal components analysis, support vector machines, kernel density estimation) and data partitioning methods to quantify remotely sensed landscape conditions associated with CWD cases. We used various model evaluation tools (e.g., AUC ratios, cumulative binomial testing, Jaccard similarity) to assess predictions of disease transmission risk using independent CWD data. We further examined model variation in the context of uncertainty. We provided significant support that vegetation phenology data representing landscape conditions can predict and map CWD transmission risk. Model predictions improved when incorporating inferred home ranges instead of raw hunter-reported coordinates. Different data availability scenarios identified variation among models. By showing that CWD could be predicted and mapped, our project adds to the available tools for understanding the landscape ecology of CWD transmission risk in free-ranging populations and natural conditions. Our modeling framework and use of widely available landscape data foster replicability for other wildlife diseases and study areas.

Keywords: CWD; chronic wasting disease; hypervolume; landscape epidemiology; prion; wildlife disease.

Figure 1

Natural history of chronic wasting disease in Virginia, US from 2007 to early 2020. (A) Virginia county colors represent average annual number of deer samples ranging from no samples (white) to highest sampling intensity in Frederick County (dark red; ~400 deer/year); mean cumulative number of samples is 143 white-tailed deer per county. In response to CWD detections, DWR increased sampling intensity and delineated disease management areas (dotted county lines). Our case study area (dark gray rectangle) focused on the northern tip of Virginia and portions of Maryland, Pennsylvania, and West Virginia. (B) Stacked bar chart shows sex ratios of CWD-positive cases from 2009 to 2019 hunting seasons. The apparent drop in the number of cases in 2019 is attributed to reallocation of DWR resources to prioritize sampling in non-CWD endemic counties. (C) Bar chart shows prevalence in CWD endemic Frederick County from DMA1 increasing over time. Details of higher prevalence values in some regions are lost due to administrative boundaries. (D) Horizontal bar chart shows hunter harvest as the predominant sampling method, followed by testing roadkill and clinical suspect cases. Please note that hunting seasons begin in late calendar months (typically November) and extend into early months of the following year.

Figure 2

Case study area delineation and current CWD distribution. Map shows the case study area outline (gray rectangle) that was determined using dissolved buffers (red line) of maximum dispersal distance (45 km) (50) around positive cases (circles). Colored circles show quadrant organization of CWD positives (n = 88) used in modeling in Virginia Department of Wildlife Resources Disease Management Areas (DMA) 1 and 2 (gray polygons). This case study area was used for acquiring landscape information (see modeling workflow in Figure 3).

Figure 3

Workflow of black-box landscape epidemiology analysis. Workflow displays our black-box framework and evaluation procedure. (A) We collected remotely sensed enhanced vegetation index (EVI) and cropped rasters to the extent of the maximum dispersal potential for our focal species as radii about disease records (for details, see Figure 2). (B) We performed a principal component analysis on the EVI data to reduce multicolinearity and generate dimensions in analysis. (C) Next, we selected only disease occurrence locations and partitioned data into geographic quadrants for calibration and evaluation datasets. (D) We extracted principal component data from the PCA-generated raster stack at both Harvest Location (the single 250 m × 250 m raster pixel value corresponding to the precise sampling coordinate) and Home Range scales (a summary of values from multiple 250 m × 250 m pixels surrounding precise sampling coordinate) inferred from the home range size of focal species. (E) We developed 24 hypervolumes using Gaussian kernel density estimation and a one-class support vector machine for all six quadrant combinations and both scales. (F) Each hypervolume was projected onto geography in the form of binary risk maps, but KDE hypervolumes were additionally projected into continuous risk maps. (G) We used models to generate risk maps and evaluated models using methods appropriate for the projection: cumulative binomial probability testing (for binary maps) (69) and partial ROC (for continuous maps) (68). To more rigorously test models, we penalized suitability inherent to calibration data and restricted each map to evaluation dataset quadrants (represented by “×”).

Figure 4

AUC ratio evaluation from partial ROC of continuous suitability maps in each geographic partition. Model evaluation according to different quadrants configuration used for calibration. Half-violin and raw data distribution plots denote bootstrapped AUC ratios obtained from the evaluation quadrants (not used in model calibration) for models based on Harvest Locations (gray) and Home Ranges (blue). Note that most configurations have AUC ratios >1, which is above the threshold for random expectation (red line; p < 0.001), except for one Harvest Location model calibrated with the northeast and northwest quadrants with non-significant predictions (p > 0.05). Ribbon abbreviations follow cardinal directions (i.e., NE, northeast; NW, northwest; SE, southeast; SW, southwest).

Figure 5

Hypervolume variation and characteristics by scale. (A) Plots show Jaccard's similarity index between hypervolume sets of the full CWD-positive dataset (n = 88) and those created from iteratively removing one occurrence record (leave-one-out). Note that the models never reach a Jaccard similarity index values at 1 denoted with dashed horizontal line, which indicates a failure to reach complete overlap and identical position and size in environmental space. (B) Half-violin plots and raw data distribution represent volumes extracted from hypervolumes created from leave-one-out iterations. Colors represent the scale for whether models were delineated from data at Harvest Locations (gray) or Home Ranges (blue). Note that hypervolumes from home ranges generally occupied smaller volumes in environmental space despite equal sample sizes.

Figure 6

Maps of projected CWD transmission risk from uncertainty analysis. Risk maps identify areas determined with more consistent probable risk (red) and less consistent risk (blue) for CWD transmission from jackknife analysis. We found more homogenous and widespread transmission risk being consistent among models using (A) Harvest Locations compared with (B) Home Ranges. Note that counties with considerable transmission risk include Rappahannock County (white outlined polygon). Lines indicate boundaries of states (thick black) and counties (thin black), while points (yellow circles) represent known CWD cases (n = 88). Overall, the amount of area predicted as consistently risky was higher in models generated from Harvest Locations.