Use of milk mid-infrared spectra to predict serum non-esterified fatty acid concentrations in Chinese Holstein cows

Abstract

Negative energy balance (NEB) in high-yielding cows during the peripartum period raises the risk of postpartum diseases. High-level concentration of non-esterified fatty acid (NEFA) is a good indicator of excessive NEB. The current low-cost and high-throughput mid-infrared (MIR) spectroscopy method is gradually applied to predict NEFA concentrations for NEB identification. The objective of this study was to compare different pre-processing methods and analysis models for optimal predictions of serum NEFA using milk MIR spectra. Four spectral pre-processing methods: standard normal variate, first-order derivative (FD), second-order derivative, and Savitzky-Golsy convolution smoothing, and four prediction models: partial least squares regression, ridge regression, lasso regression (LassoR), and random forest regression were investigated. In total, 366 collected serum and milk samples within the 1–7 weeks postpartum were randomly divided into the training (70%) and test (30%) sets for cross-validations. The results showed that the combined strategy of FD-LassoR model when parity and days in lactation information were considered resulted in the highest R² = 0.643, RMSE = 0.153 mmol/L, and highest residual predictive deviation = 1.665 of predictions on the test set. In addition, R² and RMSE values of FD-LassoR combined with other information were still higher than the other four prediction scenarios. Therefore, our study enables the optimal prediction of serum NEFA concentrations using milk MIR spectra in the further research and practical applications.