Predicting dry matter intake in Canadian Holstein dairy cattle using milk mid-infrared reflectance spectroscopy and other commonly available predictors via artificial neural networks

Authors: SHADPOUR, SAEED - University Of Guelph; CHUD, TATIANE - University Of Guelph; HAILEMARIAM, DAGNACHEW - University Of Alberta; OLIVEIRA, HINAYAH - University Of Guelph; PLASTOW, GRAHAM - University Of Alberta; STOTHARD, PAUL - University Of Guelph; LASSEN, JAN - Aarhus University; Baldwin, Ransom - Randy; MIGLIOR, FILIPPO - University Of Guelph; BAES, CHRISTINE - University Of Guelph; TULPAN, DAN - University Of Guelph; SCHENKEL, FLAVIO - University Of Guelph

Submitted to: Journal of Dairy Science
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 5/31/2022
Publication Date: 10/1/2022
Citation: Shadpour, S., Chud, T.C., Hailemariam, D., Oliveira, H.R., Plastow, G., Stothard, P., Lassen, J., Baldwin, R.L., Miglior, F., Baes, C., Tulpan, D., Schenkel, F.S. 2022. Predicting dry matter intake in Canadian Holstein dairy cattle using milk mid-infrared reflectance spectroscopy and other commonly available predictors via artificial neural networks. Journal of Dairy Science. 105(10):8257–8271. https://doi.org/10.3168/jds.2021-21297.
DOI: https://doi.org/10.3168/jds.2021-21297

Interpretive Summary: Dry matter intake (DMI) is central to understanding feed efficiency but measuring DMI in a commercial system for individual cows is expensive. Mid-infrared reflectance spectroscopy (MIRS) on milk samples could be an inexpensive alternative. The objectives of this study were to develop algorithms to predict DMI from MIRS and validate the robustness of developed prediction models. A total of 7,398 milk samples from 509 dairy cows distributed over Canada, USA and Denmark were analyzed. Including MIRS data into prediction models improved weekly average DMI prediction in dairy cattle, but it seems MIRS predicts DMI mostly through its association with production traits, which may limit its use. The predictive ability of nonlinear Artificial Neural Network (ANN) programs over linear ANN indicates possible nonlinear relationships between weekly average DMI and the explanatory variables. In general, ANN using Bayesian regularization yielded slightly better weekly average DMI predictions compared to ANN using the Levenberg-Marquardt and scaled conjugate gradient training algorithms.

Technical Abstract: Dry matter intake (DMI) is a key factor influencing feed efficiency of animals, but measuring DMI in a commercial system for individual cows is expensive. Mid-infrared reflectance spectroscopy (MIRS) on milk samples could be an inexpensive alternative to predict DMI. The objectives of this study were to (1) assess whether milk MIRS data could improve DMI predictions using artificial neural networks (ANN); (2) investigate the ability of different ANN architectures to predict unobserved DMI; and (3) validate the robustness of developed prediction models. A total of 7,398 milk samples from 509 dairy cows distributed over Canada, USA and Denmark were analyzed. For each milk spectra record, the corresponding weekly average DMI (kg/d), test-day milk yield (MY, kg/d), fat yield (FY, g/d), and protein yield (PY, g/d), standardized metabolic body weight (SMBW), age at calving (AC), year of calving (YC), season of calving (SC), days in milk (DIM, d), lactation number (LN), country (CN), and herd (HD) were available. The weekly average DMI was predicted with various ANN architectures using covariate sets as follows: set 1 (MY); set 2 (MY, FY, and PY); set 3 (MY, FY, PY, and SMBW); set 4 (505 MIRS wavelengths), set 5 (MY, SMBW, and 36 principal components of MIRS wavelengths); set 6 (MY, FY, PY, SMBW, and 36 principal components of MIRS wavelengths); set 7 (MY, FY, PY, SMBW, and 505 MIRS wavelengths). All covariate sets also included AC and DIM. In addition, the classification effects of SC, CN, HD and LN were always included in the model. The explored ANN architectures consisted of three training algorithms (Bayesian regularization, Levenberg-Marquardt, and scaled conjugate gradient), two types of activation functions (hyperbolic tangent and linear), and from 1 to 10 neurons in hidden layers. Models were compared using cross-validation based on leaving out 10% of records (validation A) and leaving out 10% of cows (validation B). Using MY alone to predict weekly average DMI resulted in inferior fitting statistics compared to the other covariate sets. The predictive correlation (r) ranged from 0.478 to 0.649 and root mean squared error (RMSE) ranged from 2.862 to 3.328 (set 1). Adding FY and PY to MY improved weekly average DMI prediction (set 2; r = 0.49 to 0.674; RMSE = 2.779 to 3.308). The weekly average DMI prediction improved by adding SMBW as an additional predictor trait (set 3; r = 0.521 to 0.696; RMSE = 2.7 to 3.249). Greater improvement in weekly average DMI prediction was observed when the principal components of MIRS wavelengths were added to the model, along with MY and SMBW (set 5; r = 0.666 to 0.767; RMSE = 2.417 to 2.826). Combining FY and PY to the covariates included in the set 4 did not have a major impact on weekly average DMI prediction (set 6; r = 0.672 to 0.77; RMSE = 2.401 to 2.802). Using MIRS together with MY, FY, PY, and SMBW as predictors resulted in a slightly superior fitting statistic (set 7; r = 0.685 to 0.778; RMSE = 2.373 to 2.833). Including MIRS data improved weekly average DMI prediction in dairy cattle, but it seems MIRS predicts DMI mostly through its association with production traits, which may limit its use. The predictive ability of nonlinear ANN over linear ANN indicates possible nonlinear relationships between weekly average DMI and the explanatory variables. In general, ANN using Bayesian regularization yielded slightly better weekly average DMI predictions compared to ANN using the Levenberg-Marquardt and scaled conjugate gradient training algorithms.