Optimization of fecal near infrared spectroscopy for predicting organic matter digestibility in cows using local...

Accurate estimation of diet digestibility is essential for the efficient and sustainable production of milk and meat by ruminants. However, standard in vivo methods for assessing the nutritive value of ruminant feed are both time-consuming and costly. As a result, indirect methods based on the infrared absorbance of animal feces have been developed as practical alternatives. Most existing prediction models rely on global calibrations, which establish relationships between organic matter digestibility (OMD) and near-infrared (NIR) spectra across an entire sample population. However, using only a subset of samples, those most similar to the target sample, may improve prediction accuracy. This study compares a global prediction approach, partial least squares regression (PLSR), with several Local modeling techniques: Locally weighted PLS regression (LWPLSR), k-nearest neighbors Locally weighted PLS regression (KNN-LWPLSR), and an aggregated version of the latter (KNN-LWPLSR-AGG), for predicting OMD in cattle. A dataset of 466 fecal samples with corresponding in vivo OMD measurements was used. Of these, 299 samples were used for model calibration, while the remaining 167 were split into two groups: 76 for external validation and 91 for testing under routine conditions. Results showed no significant difference in prediction accuracy among the Local methods (P > 0.05). However, LWPLSR outperformed the global PLSR model (P < 0.05). The standard error of the in vivo standard reference method was estimated at 0.013 g/g, while the best NIR-based prediction error was 0.016 g/g. Given its balance between predictive accuracy and computational efficiency, LWPLSR is recommended for practical applications.