Abstract
This study aimed to improve the prediction of AME and AMEn of maize for laying hens by evaluating the contribution of anti-nutritional factors (ANF) as predictor variables. Prediction models were developed using two sets of independent variables: a nutrient-only set (starch, crude protein, and ether extract; IVC1) and an extended set additionally including ANF (crude fiber, total arabinoxylan, ash and phytic acid; IVC2). Five algorithms (Linear Regression, Ridge, LASSO, Elastic Net, and Random Forest, RF) were trained with Bayesian hyperparameter optimization and evaluated using repeated nested cross-validation (3 × 10-fold) to obtain unbiased performance estimates. With nutrient-only inputs, RF achieved the highest accuracy, with R²_CV values of 0.645 for AME and 0.669 for AMEn, and RMSE_CV values of 22.9 and 21.3 kcal/kg DM, respectively. Including ANF further improved RF performance to R²_CV = 0.748 (RMSE_CV = 19.1 kcal/kg DM) for AME and R²_CV = 0.758 (RMSE_CV = 18.1 kcal/kg DM) for AMEn. Wilcoxon signed-rank tests confirmed that the improvements from IVC1 to IVC2 were consistent for all algorithms (P < 0.001). In conclusion, incorporating ANF as predictor variables substantially improves the prediction of maize AME and AMEn for laying hens beyond conventional nutrient composition. Among the tested algorithms, RF combined with Bayesian hyperparameter optimization and repeated nested cross-validation provided the most accurate and robust models.