Detection of cows with insemination problems using selected classification models

In the present study, the detection of cows with artificial insemination (AI) difficulties using selected statistical and machine learning methods is presented. Cows were divided into two classes: those that conceived after one or two services (“GOOD”) and those that required more than two services per conception (“POOR”). The best performance was exhibited by one of the artificial neural networks (ANN) and the multivariate adaptive regression spline (MARS) method (AIC, BIC, RMS and accuracy); whereas logistic regression (LR) and classification functions (CF) were of somewhat lower quality. The detection of cows with AI difficulties, performed on the basis of the test set comprising new instances, showed that the ANN and MARS were more precise in comparison with the statistical methods. Sensitivity and specificity were over 85% for the perceptron with two hidden layers (MLP2) and MARS and approximately 80% or lower for LR and CF. From among variables determining the AI category, the average calving interval and cow body condition index were the most important. Other significant variables were lactation number, pregnancy length, sex of calf from previous calving and cow age. The prognoses obtained using ANN and MARS can be used for the appropriate preparation of cows for AI.

Research highlights▶ The possibility of practical detection of cows with insemination problems using the multivariate adaptive regression spline (MARS) method and artificial neural networks (ANN). ▶ Finding that detections of cows with insemination problems performed by means of machine learning methods (MARS and ANN) are more conservative than detections performed using traditional methods – logistic regression (LR) and classification functions (CF). ▶ Indication that changes of body condition during artificial insemination compared to the average body condition during the production season can determine insemination difficulties.