Predicting reticuloruminal pH and subacute ruminal acidosis of individual cows using machine learning and Fourier-transform infrared spectroscopy milk analysis

Low reticuloruminal pH (rpH) for a prolonged period could lead to SARA. This disease negatively affects cow health and is associated with monetary losses for the dairy industry. The aim of this study was to predict rpH and SARA separately using different machine learning (ML) models applied to Fourier transform infrared spectroscopy (FTIR) spectra obtained from routine DHI milk analysis of individual cows. A total of 107 primiparous and multiparous Holstein cows were selected from 12 commercial farms in Québec, Canada, and their rpH was continuously monitored for 150 d using wireless boluses. In parallel, 2,634 individual milk samples were collected in the morning and afternoon and analyzed to obtain FTIR spectra. After the cleaning process, 1,744 samples remained, evenly divided into 872 morning (a.m.) and 872 afternoon (p.m.) samples. The FTIR and rpH data were combined to create 3 equally balanced datasets for ML model development: one for a.m. samples, one for p.m. samples, and one composed of both a.m. and p.m. samples, with 872 samples in each dataset. Various spectra preprocessing methods were evaluated, including using the first derivative of the spectra and filtering with 3 different sets of spectra. Additionally, different ML algorithms, including partial least squares, random forest, and gradient boosting, were used to predict rpH and SARA. A total of 36 different models were developed and evaluated for both rpH and SARA prediction. All ML models were assessed using 3 different cross-validation (C-V) methods: nested 10-fold, nested leave-one-farm-out (LOFO), and nested leave-cows-out (LCO) C-V. For rpH prediction, the best performance was achieved using nested 10-fold C-V with median R 2 values of 0.26, 0.26, and 0.22 for the a.m., p.m., and a.m./p.m. datasets, respectively. However, these performances were likely overoptimistic as none of the models evaluated using nested LOFO or nested LCO C-V obtained R 2 higher than 0.12. Unlike rpH, SARA prediction accuracies evaluated using nested LOFO (a.m.: 59%, p.m.: 69%, a.m./p.m.: 64%), and nested LCO (a.m.: 67%, p.m.: 66%, a.m./p.m.: 64%) were closer to the nested 10-fold C-V. These results indicated that rpH was likely not predictable from FTIR, but SARA can be predicted separately and directly from FTIR with 69% accuracy from routine DHI milk samples of individual cows.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.