Sistema neural para predição de brucelose em rebanhos bovinos

Abstract

Brucellosis is one of the most important diseases in Brazilian cattle rearing, and animal defense programs have difficulty, especially operational, in its control. The use of computer intelligence tools, known as Artificial Neural Networks (ANN) can be of great use in sanitary vigilance and epidemiologic services to aid in the diagnosis. This work had the objective of developing a mathematical model using NAA to classify herds positive and negative for Bovine Brucellosis. In addition, forms of previous data treatments were compared in order to improve the efficiency of the ANN model. The development of the neural system was comprised of the following stages: 1) obtaining the data (interview with cattle farmers); 2) pre-processing of the data (variable selection and quality analysis); 3) ANN training (definition of the ANN architecture); 4) result presentation (discriminate the herd as positive or negative – prediction model). The work was conducted with data from the investigation performed by the official animal sanitary defense service of the state of Minas Gerais, Brazil (MAPA/IMA), from September of 2010 to December of 2012. The ANN was trained with different models for comparison: 1) with all database variables, with no pre-selection (49 variables), and 2) with pre-selection of the most relevant variables. The pre-selection methods compared were: 1) association test by chisquare method, considering the statistical significance levels of 5% (p<0.05) and 10% (p<0.10): row models; 2) Fisher Linear Discriminant with 10, 20 or 30 most relevant variables: three models. The Tukey test with 95% of probability was used for comparison between the sensitivity, specificity and efficiency values of the models resultant of this processing. All models showed the best results for 83.33% of sensitivity. The best results for specificity ranged from 53.89 to 66.60%, and for efficiency, from 68.61 to 74.96%. There was no significant difference (p>0.05) for efficiency and sensitivity between the models. Regarding the specificity, the pre-selection model by chi-square (model 2 – p<0.10) reached the highest value, however, the mean was significantly similar to the other models (p>0.05). The results showed that the application of the ANN for the prediction of positive and negative herds is promising, and the preselection of variables to obtain simpler models must be used, given that the efficiency is maintained. This tool must be used for animal defense for reducing operational costs and guidance of disease control policies.