Multiple regression and machine learning based methods for carcass traits and saleable meat cuts prediction using non-invasive in vivo measurements in commercial lambs

Abstract

The aim of this study is to investigate the performance of multiple linear regression and machine learning methods to predict carcass traits and commercial meat cuts in lambs using non-invasive in vivo measurements. Bayesian networks were also investigated as an alternative method for feature selection. Phenotypes from 74 nondescript breed lambs were measured. A leave-one-out cross-validation strategy was performed and predictive ability statistics (R2, RMSE) were assessed. Moderate to high prediction accuracies were observed, with R2 values across models ranging, from 0.36 to 0.88 for carcass traits and 0.65 to 0.84 for meat cuts predictions. Results suggest that support vector machine algorithm is a potential alternative to the traditional multiple linear regression model. Further, the results of this study suggest that both stepwise and Bayesian network procedures could be useful as pre-selection tools of the input variables in non-parametric approaches and the best method for feature selection may be trait and model dependent.