Bagging as a predictive method for landscape epidemiology of Lyme disease

Risk assessment of exposure to Lyme borreliosis, a tick-borne illness common in North America and endemic in several regions of Europe, is crucial for prevention, the prompt diagnosis of disease and the treatment of cardiac complications (Lyme carditis). In regions where the epidemic is prevalent, the presence of Borrelia burgdoferi (the agent of Lyme disease) could, in fact, be suspected in otherwise healthy patients with unexplained cardiac symptoms such as supraventricular tachiarrhythmia, atrioventricular block of unknown origin, and in dilated cardiomiopathy. The construction of predictive geospatial risk models based on critical ecological factors and rates of infection in ticks is thus proposed for a landscape epidemiology of Lyme disease. A computational method based on bootstrap aggregation (bagging) of tree-based classifiers integrated with a Geographical Information System (GIS) has been developed to predict the risk of Lyme disease in province of Trento. The risk of bite exposure was estimated from 438 field samples for a total of 3422 specimens of Ixodes ricinus. One hundred risk classification models were built by bootstrap resampling of a data base associating tick sampling to the variables extracted from GIS thematic digital maps of elevation, the type of vegetation and soil, exposure, and the density of roe deer (a key host for adult ticks). The final model was then obtained via the aggregation of the 100 sub-models and tested for predictivity also in comparison to previous results obtained with a single predictive classification tree (b632+ model) and with a classical logistic regression model developed on the same data set. Additional data sets and measures of prevalence of infection were then used to cross-validate and further extend the model. A tree-based model is now available that predicts local risk of tick presence over cells of size 50 x 50 meters. The overall accuracy of the bagging based model is 75% with respect to the positive samples (presence of ticks) and 76% with respect to the negative ones. The model stabilizes the previous single tree model prediction. A classical linear model based on stepwise variable selection exhibited poorer performance (av. accuracy: 65 +/- 5 %). On an additional data base, high infestation was correctly predicted by bagging in 81% of cases. Bagging has therefore improved the computational model for risk assessment of tick bites in Trentino. The resulting GIS based model is currently being expanded in order to directly predict the risk of B. burgdoferi infection for the purpose of reducing human exposure to Lyme disease and, thus, facilitate the recognition and the correct treatment of idiopathic heart failure