As concern for animal welfare grows, the challenge of carrying out individual assessments on a large scale remains unresolved. This study aimed to investigate the potential of random regression test-day models (RR-TDM) to predict individual welfare scores of dairy cows throughout the entire lactation, using routinely collected milk yield and composition records. A total of 2,381 cows from 30 herds in the Walloon Region of Belgium were assessed between 2019 and 2024. The protocol used in this study focused on four key principles: proper feeding, suitable housing, good health, and appropriate behavior. An initial welfare assessment was conducted on all cows to calculate the SGI-I (Score Global Individuel de bien-être), a composite welfare score based specifically on feeding, housing, and health criteria. Subsequently, a more comprehensive evaluation (SGI-II), which also incorporated behavioral assessments, was performed on a subset of 1,229 cows. The data consisted of test-day records of milk yield (MY), fat percentage (FP), protein percentage (PP), somatic cell score (SCS), SGI-I and SGI-II. To evaluate the potential of a multiple-trait random regression test-day model (RR-TDM) to predict SGIs, a five-fold cross-validation procedure was performed. For each trait (SGI-I and SGI-II), the dataset was partitioned into a training set (75%) and a validation set (25%) where SGI scores were masked. (Co)variance components were then estimated using the combined data from both the training and validation stes; then, SGI solutions were predicted for all animals in the population, using MY, FP, PP, and SCS as predictors. This entire process of partitioning, masking, estimation, and prediction was repeated five times for SGI-I and five times for SGI-II to ensure the stability and robustness of the predictions. Prediction accuracy was evaluated using Pearson's correlation, root mean squared error (RMSE), prediction error (PE), absolute prediction error (APE), and Cohen's Kappa. Mean (standard deviation; SD) SGI-I and SGI-II were 69.92 (10.28) and 65.90 (9.65), respectively. The mean Pearson correlation between predicted and observed scores was 0.76 (95% CI: 0.73 - 0.80) for SGI-I and 0.72 (95% CI: 0.66 - 0.77) for SGI-II. Cohen's Kappa values indicated substantial agreement, with κ = 0.75 (95% CI: 0.71 - 0.79) for SGI-I and κ = 0.72 (95% CI: 0.66 - 0.78) for SGI-II. The mean R2 was 0.58 (RMSE = 6.66) for SGI-I and 0.51 (RMSE = 7.23) for SGI-II. These findings demonstrate the potential of RR-TDM to predict SGI for the entire lactation using available test-day records of milk yield and composition in Holstein dairy cows.
Chen et al. (Sun,) studied this question.