Modelling tick bite risk by combining random forests and count data regression models

Abstract

The socio-economic and demographic changes that occurred over the past 50 years have dramatically expanded urban areas around the globe, thus bringing urban settlers in closer contact with nature. Ticks have trespassed the limits of forests and grasslands to start inhabiting green spaces within metropolitan areas. Hence, the transmission of pathogens causing tick-borne diseases is an important threat to public health. Using volunteered tick bite reports collected by two Dutch initiatives, here we present a method to model tick bite risk using human exposure and tick hazard predictors. Our method represents a step forward in risk modelling, since we combine a well-known ensemble learning method, Random Forest, with four count data models of the (zero-inflated) Poisson family. This combination allows us to better model the disproportions inherent in the volunteered tick bite reports. Unlike canonical machine learning models, our method can capture the overdispersion or zero-inflation inherent in data, thus yielding tick bite risk predictions that resemble the original signal captured by volunteers. Mapping model predictions enables a visual inspection of the spatial patterns of tick bite risk in the Netherlands. The Veluwe national park and the Utrechtse Heuvelrug forest, which are large forest-urban interfaces with several cities, are areas with high tick bite risk. This is expected, since these are popular places for recreation and tick activity is high in forests. However, our model can also predict high risk in less-intensively visited recreational areas, such as the patchy forests in the northeast of the country, the natural areas along the coastline, or some of the Frisian Islands. Our model could help public health specialists to design mitigation strategies for tick-borne diseases, and to target risky areas with awareness and prevention campaigns.

Assessment of vector-host-pathogen relationships using data mining and machine learning

“… Davi C. et al 2019 Prediction, Classification, DL SVM, ANN Q3 Genomics, Phenomics accuracy>86%, and sensitivity and specificity over 98% and 51%, respectively. using only genome markers 29,191,515 [46] This study present simulated global distribution of Aedes aegypti and Aedes albopictus at a 5 × 5 km spatial resolution with high-dimensional multidisciplinary datasets and machine learning methods Ding F. et al 2018 Prediction SVM, GBM, RF Q2 Population, Other SB level RF (AUC) of 0.973 and 0.974, respectively, GBM (AUC of 0.971 and 0.972, respectively) and SVM (AUC of 0.963 and 0.964, respectively) statistically significant 31,821,325 [44] Model tick bite risk using human exposure and tick hazard predictors, represents a step forward in risk modelling by combining a well-known ensemble learning method, Random Forest, with four count data models of the (zero-inflated) Poisson family. Garcia-Marti I et al 2019 Prediction, Classification RF, Ensemble Q2 Population stdev = 3.15) Pearson/Kendall coefficient Species/organism 29,114,054 [38] In this study combine techniques in serial block-face scanning-electron microscopy and deep-learning–based image segmentation algorithms to visualize the distribution, abundance, and interactions of Ophiocordyceps unilateralis sensu lato fungus inside the body of its manipulated host.…”

“…The model accuracy range was 82–97%. The problems addressed included: the identification of high risk snail habitats as a function of Schistosoma japonicum infection [43] , modelling of tick bite risk based on ecological factors [44] , predicting the global distribution of Aedes mosquitoes and the effects of seasonal changes on their range [45] , [46] and the prediction of Dengue virus outbreak risk based on climate [47] , [48] .…”

Count Regression and Machine Learning Approach for Zero-Inflated Over-Dispersed Count Data. Application to Micro-Retail Distribution and Urban Form

“…decision trees) and strong estimators (i.e. models for count data) as proposed by [ 34 ] or with multi-output modelling approaches [ 35 ]. Advanced cross-validation techniques specifically conceived for highly spatially correlated data [ 36 ] might be also considered.…”

Effect of restoration measures in Atlantic rivers: A 25‐year overview of sea and riverine brown trout populations in the River Bidasoa

“…RF is a statistical ensemble method based on the combination of a multitude of decision trees, which is used to determine the mean prediction of the individual trees (Breiman, 2001). It can capture the overdispersion or zero‐inflation inherent in count data (Garcia‐Marti et al., 2019), allows freedom from normality and homoscedasticity assumptions and does not require previous data transformation or a separate test set for cross‐validation as it is performed internally during the run (Breiman, 2001).…”