Abstract #M150

# M150
Application of partial least squares regression to predict feed intake using feeding behavior traits in growing Holstein heifers.
J. R. Johnson*1, G. E. Carstens1, C. Heuer2, N. Deeb2, 1Texas A&M University, College Station, TX, 2STgenetics, Navasota, TX.

The objective of this study was to evaluate the use of partial least squares (PLS) regression models to predict feed intake of growing Holstein heifers (n = 609; Initial BW = 246 ± 43 kg), using feeding behavior traits. DMI and feeding behavior traits were measured using a GrowSafe System for 70 to 100 d (15 trials) while heifers were fed a corn-silage based ration. Nineteen feeding behavior traits were evaluated: frequency and duration of bunk visit (BV) and meal events, head-down duration (HDD), average meal length, maximum non-feeding interval, time-to bunk (TTB), corresponding day-to-day variation (SD) of these traits, and ratios of HDD per BV duration, HDD per meal duration, and BV events per meal event. Test-set validation techniques were used to calibrate and validate PLS prediction equations for DMI using performance (mid-test BW0.75and ADG) and feeding behavior traits as independent variables. Validation groups were built by randomly selecting 3 trials to be used for validation, with results presented as the average of 5 iterations, with each trial being used for validation only once. For calibration, independent variables were excluded if variable of importance in projection (VIP) scores were less than 0.80. Overall, 12 of 19 feeding behavior traits were included (VIP >0.80) in the final model. The base model (Mid-test BW0.75and ADG) accounted for 54% of the variation in individual-animal DMI. Inclusion of feeding behavior traits to the base model increased the R2 of validation from 0.54 to 0.74 and reduced the model SE from 0.78 to 0.72. The R2 and SE of model calibration and validation were similar (0.75 and 0.73 vs. 0.74 and 0.72, respectively), indicating that the PLS model was robust in predicting individual-animal DMI across trials. No differences (P = 0.93) were found between observed (8.59 ± 1.67 kg/d) and model predicted (8.59 ± 1.50 kg/d) DMI. Overall, performance and feeding behavior traits accounted for 74% of the variation in individual-animal DMI using PLS regression models. These results indicate that future prediction models for individual-animal DMI may benefit from the inclusion of feeding behavior traits.

Key Words: dairy cattle, partial least squares (PLS), feeding behavior