Mining data from milk mid-infrared spectroscopy and animal characteristics to improve the prediction of dairy cow's liveweight using feature selection algorithms based on partial least squares and Elastic Net regressions

Abstract Body weight (BW) of dairy cows is relevant for breeding programs and farm management to assess the maintenance requirements, reproduction performance, or health status of cow. Currently, it is still difficult to follow BW changes of individual cows routinely in large herds. Combined with animal characteristics, milk mid-infrared (MIR) spectrum was proposed as an additional source of information to predict BW under the framework of dairy herd improvement (DHI) programs. However, the presence of less informative variables in the prediction equation could impact negatively its robustness. This research aims to improve the robustness of BW regression models by applying a feature selection before modeling. A total of 5,920 BW records composed of animal characteristics and milk MIR spectrum were collected from Holstein cows. Three feature selection algorithms were applied to select the most informative variables: partial least squares regression (PLS) combined with sum of ranking difference (PLS-SRD), PLS combined with uninformative variables elimination (PLS-UVE), and the output of Elastic Net regression (EN). Four herd independent validation sets and the corresponding remained calibration datasets having on average 163 and 1,708 records, respectively, were used to develop models using PLS or EN approaches. Ten-fold cross-validation was conducted to parametrize each model. Parity, days in milk (DIM), milk yield (MY), and two MIR spectral points were selected as relevant variables to predict BW. PLS (root mean square error of validation, RMSEp = 60 kg) and EN (RMSEp = 60 kg) regressions employing these 5 predictors were more robust than the models developed without MIR or using MIR without feature selection. The EN models had a cross-validation root mean square error of around 53 kg. The 2 MIR points explained up to 4.20% variation in predicting BW. The RMSE of validation sets using another brand of spectrometer were around 64 kg. This study confirms the possibility to predict an indicator of BW from animal characteristics and MIR variables. The variable selection procedures improved the model's robustness and transferability. The accuracy of BW prediction seems to be sufficient to provide useful information for breeding program and farm management decisions under a DHI framework.