Using artificial neural networks to predict pH, ammonia, and volatile fatty acid concentrations in the rumen

Abstract

The objectives of this study were (1) to predict ruminal pH and ruminal ammonia and volatile fatty acid (VFA) concentrations by developing artificial neural networks (ANN) using dietary nutrient compositions, dry matter intake, and body weight as input variables; and (2) to compare accuracy and precision of ANN model predictions with that of a multiple linear regression model (MLR). Data were collected from 229 published papers with 938 treatment means. The data set was randomly split into a training data set containing 70% of the observations and a test data set with the remaining observations. A series of ANN with a range of 1 to 9 artificial neurons in 1 hidden layer were examined, and the best one was selected to compare with the best-fitted MLR model. The performance of model predictions was evaluated by root mean square errors (RMSE) and concordance correlation coefficients (CCC) using cross-evaluations with 100 iterations. When using the ANN to predict ruminal pH and concentrations of ammonia, total VFA, acetate, propionate, and butyrate, the RMSE were 4.2, 41.4, 20.9, 22.3, 32.9, and 29.7% of observed means, respectively. The RMSE for the MLR were 4.2, 37.8, 18.3, 19.9, 29.8, and 26.6% of the observed means. The CCC for ruminal pH, ruminal concentrations of ammonia, total VFA, acetate, propionate, and butyrate were 0.57, 0.49, 0.45, 0.40, 0.52, and 0.40, using the ANN, and 0.37, 0.48, 0.40, 0.29, 0.43, and 0.35, using the MLR. Evaluations of the MLR and the ANN indicated that these 2 model forms exhibited similar prediction errors, with 4.2, 39.6, 19.6, 21.1, 31.3, and 28.1% of observed means for pH, ammonia, total VFA, acetate, propionate, and butyrate. Although the ANN increased the precision of predictions related to ruminal metabolism, it failed to improve the accuracy compared with the linear regression model.