Machine Learning Prediction of Seasonal Ectoparasite Distribution in Wild Mammals

Abstract

Ectoparasites like ticks, fleas and mites are often zoonotic and are a public health and veterinary hazard, as wild mammal populations are a frequent natural reservoir for these parasites. Conventional ecological monitoring techniques are still very time-consuming and expensive, and are generally insufficient for monitoring fine-scale spatial and seasonal distribution changes. In order to address this, this study aims to develop an innovative machine learning approach that can predict and map the seasonal ectoparasite burdens of wild mammal hosts. The framework combines high-resolution climatic data, host demographic data, and spatial-temporal features to model non-linear ecological relationships, which are often not captured by traditional statistical modeling. Several algorithms such as Random Forest, Extreme Gradient Boosting (XGBoost) and Support Vector Machines (SVM) were tested using a robust spatial-temporal cross validation approach. The XGBoost model was the most successful algorithm, demonstrating an outstanding predictive performance for the AUC-ROC of 0.89 and root mean square error (RMSE) of 0.32 for the prediction of the parasite density. The main variables that determine the seasonality distribution patterns according to feature importance analysis through SHAP values were land surface temperature, relative humidity, and host population density. The results show that machine learning-based predictive modeling can be a valuable proactive early warning system tool for wildlife conservation, targeted parasite control, and public health monitoring of vector-borne diseases