Abstract
The aim of the present study was to: 1) check whether it would be possible to detect cows susceptible to mastitis at an early stage of their utilization based on selected genotypes and basic production traits in the first three lactations using ensemble data mining methods (boosted classification tress – BT and random forest – RF), 2) find out whether the inclusion of additional production variables for subsequent lactations will improve detection performance of the models, 3) identify the most significant predictors of susceptibility to mastitis, and 4) compare the results obtained by using BT and RF with those for the more traditional generalized linear model (GLZ). A total of 801 records for Polish Holstein-Friesian Black-and-White cows were analyzed. The maximum sensitivity, specificity and accuracy of the test set were 72.13%, 39.73%, 55.90% (BT), 86.89%, 17.81%, 59.49% (RF) and 90.16%, 8.22%, 58.97% (GLZ), respectively. Inclusion of additional variables did not have a significant effect on the model performance. The most significant predictors of susceptibility to mastitis were: milk yield, days in milk, sire’s rank, percentage of Holstein-Friesian genes, whereas calving season and genotypes (lactoferrin, tumor necrosis factor alpha, lysozyme and defensins) were ranked much lower. The applied models (both data mining ones and GLZ) showed low accuracy in detecting cows susceptible to mastitis and therefore some other more discriminating predictors should be used in future research.