Prediction of lamb survival using machine learning algorithms with neonatal lamb behaviors and maternal behavior score in Kivircik lambs

Abstract

The aims of this study were to examine the relationship between lamb and ewe behaviors in postnatal 3-hour and lamb survival using machine learning (ML) algorithms and to determine the best ML classifier to predict lamb survival. The research data consisted of postnatal 3-hour behavior records of 43 Kivircik ewes and their 65 lambs, along with preweaning survival information of lambs. The prediction of lamb survival was performed on three datasets containing different features using decision tree, support vector machine (SVM), multilayer perceptron, logistic regression, random forest (RF), K-nearest neighbors, and boosting (B) ML algorithms. The accuracy, precision, recall, and F1 score values of the RF algorithm were 0.931, and the area under curve value was 0.966 for dataset 1, which included parity of dam, birth type and sex of lamb, and birth weight predictors, as well as postnatal lamb and ewe behaviors as features. In dataset 2, which includes principal component scores instead of lamb and ewe behaviors, the RF approach made classification with an accuracy of 0.909. In dataset 3, which includes the maternal behavior score as a feature, the RF and SVM algorithms showed similar performance (0.909 accuracy). These results indicate that, by using lamb and ewe behaviors in the postnatal 3-hour with ML methods, it is possible to classify lambs as either surviving or dying before weaning with high accuracy. In addition, it was determined that the ML algorithm that best adapted to the current study data was the RF classifier.