Estimation & Prediction

Integrating Dynamic Lactation Traits, Resilience Indicators, and Genomic Data to Predict Later-Lactation Milk Yield

---

Abstract

Accurate prediction of later-lactation performance from early-life phenotypes remains a key objective for genetic improvement and precision management in dairy production. In this study, the ability of lactation information to predict subsequent lactation cumulative milk yield was evaluated using dynamic curve descriptors, resilience indicators, and biological covariates derived from automatic milking system (AMS) records. Daily milk yield records from 710 Holstein cows were cleaned, smoothed to remove outliers, and retained only when lactations exceeded 280 days in milk. First (L1) and second (L2) lactation trajectories were modelled using the EMOTIONS package, where 47 lactation-curve models were fitted to each cow and combined in an ensemble model weighting the individual predictions based on the AIC of each model. In addition to this ensemble representation, two alternative curve-fitting scenarios were constructed from the same model fits: (i) a best-individual model, selecting the parametric function with the lowest AIC per cow; and (ii) a best-average model, selecting the function with the lowest mean AIC across cows. Model residual were used to derive resilience indicators (log-variance, lag-1 autocorrelation, skewness). Cumulative 305-d yield in second and third lactation were extracted as prediction targets. To predict subsequent lactation cumulative yield, three complementary representations of production were evaluated: rolling 30-day weighted means, biologically interpretable curve metrics (peak yield, day at peak, DMY at 305 DIM, decay slope), and principal component scores from the interpolated daily yield series. All datasets included resilience indicators and cow-level covariates such as birth year, age at first calving, and genomic breeding values for milk yield at 305. Two modelling approach were compared. Feature-based models (Ridge regression, Random Forest, XGBoost, and a feedforward neural network) were trained to directly predict next-lactation 305-d yield. In parallel, a 1D convolutional neural network (CNN) was trained in a sequence-to-sequence setting to reconstruct daily milk yield across DIM, and cumulative yield was obtained by summing predicted daily values. A Wood-function curve was fitted as a parametric baseline. Predictive performance was evaluated using cow-wise partitioning and reported as RMSE and R².Predicting L1-to-L2 cumulative yield was challenging for feature-based models (best R² = 0.11), whereas the Wood baseline achieved moderate accuracy (R² = 0.29). For L2-to-L3, predictability improved substantially with the CNN (best R² = 0.53), consistent with greater continuity across mature parities. For quartile classification, the Wood baseline provided the strongest performance (accuracy = 0.397 for L1-to-L2), with clearer discrimination of extreme quartiles than intermediate classes. Overall, the results highlight both the potential and the current limitations of predicting later-lactation outcomes from early-life milk-yield profiles. Despite the modest sample size, these results establish the phenotypic baseline necessary to evaluate whether AMS-derived resilience metrics can be effectively integrated into future large-scale breeding evaluations.

Keywords: 2026

How to Cite:

Ferrari, C., Bernini, F., Strillacci, M., Fonseca, P., Cánovas, Á. & Bagnato, A., (2026) “Integrating Dynamic Lactation Traits, Resilience Indicators, and Genomic Data to Predict Later-Lactation Milk Yield”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2286432. doi: https://doi.org/10.31274/wcgalp.24003

Rights: 1