Resumen
- Some mathematical models have been developed to assist in decision-making at different stages of livestock production system. Most mathematical models applied to bovine in vitro embryo production (IVP) use the average number of produced embryos by donor cow. However, these simulations cause biased results on embryo production yield. Thus, the use of stochastic models, which have origin in random events, appears as useful tool for definition of more accurate and economic strategies in commercial IVP. Furthermore, due to individual variability in IVP by donor cow, identification of variables having most influence in achieving pregnancy may be a useful tool for characterizing a donor embryo for its production. Artificial neural networks (ANN) and decision trees have proved to be successful in different fields of study such as medicine, genetics and animal production. This study aimed to compare different methods to predict pregnancy in commercial IVP. A real dataset was used, consisting of 9,697 in vitro produced embryo transfers. Dataset was analyzed by using logistic regression (LR); feed-forward multilayer perceptron neural network (MLP) with a hidden layer, without momentum constant and learning rate; two MLPs with momentum (0.8) and learning rate (0.2) with one and two hidden layers; and a decision tree by ID3 algorithm. For comparison, only explanatory variables that were significant (P < 0.05) on LR were used in ANNs: type of embryo (fresh or cryopreserved), type of semen used on IVF (sexed or not) and embryo developmental stage (expanded blastocyst and hatched blastocyst). All analyses were compared by predictive capacity (Table 1). Analysis by MLP without momentum and learning rate correctly classified more than 70% of positive pregnancies; but, incorrectly classified almost 60% of non-pregnancies. The ID3 algorithm was unable to design a suitable decision tree, but it was able to indicate the variables with higher impact on final response, in decreasing order: type of embryo (fresh or cryopreserved), type of semen used on IVF (sexed or not), embryo developmental stage (expanded blastocyst and hatched blastocyst). In general, all analyses were very similar, which may be explained by use of categorical explanatory variables. Use of quantitative variables as well momentum and learning rate should be considered for sequential studies.