Deep Learning Prediction of Tick-Borne Disease Risk in Wild Ungulates

Abstract

Tick-borne diseases represent a growing menace to wild ungulate species where pathogenic agents, including Anaplasma phagocytophilum, Babesia spp., and Theileria spp., are spread by Ixodes, Amblyomma, and Dermacentor ticks. Increased climatic warming, changes in land use, and altered host distributions have led to expansion of tick habitat in areas not suitable previously, thus increasing the risk of disease for deer, elk, and moose. It is challenging to account for non-linear and multi-scale relations between climate, vegetation, and host movement with conventional statistics and correlative niche modeling approach. Materials and Methods: This study offers a deep learning-based approach, which utilizes a combination of a convolutional neural network (CNN) that extracts spatial habitat features from remote sensing covariates (land cover, NDVI, elevation, proximity to water) and a long short-term memory (LSTM) network for climate time series (temperature, precipitation, humidity). The ticks and hosts' occurrence data were combined with these covariates in a 1 km grid format. Results: The CNN-LSTM model performed better than the RF and ANN models by attaining maximum classification accuracy and least errors in identifying high-risk areas. This was mainly attributed to the ability of the model to predict high-risk areas in forest-edge and riparian habitats where there is a relatively high abundance of ungulates. Vegetation greenness and minimum winter temperatures proved to be the main predictive factors of risk. Conclusion: The deep learning techniques which integrate both habitat and climatic dynamics are accurate and can be scaled to identify tick-borne risks in wild ungulates