NMR-based deep learning classification of raw cow's milk samples in different stages of mastitis.

Abstract

Background: Bovine mastitis is an inflammation of the mammary glands, usually caused by various pathogenic bacteria combined with glandular tissue injury of the cow's udder. Mastitis significantly affects the milk composition, since it changes the protein quality, the fatty acid content, the number and types of small-molecule metabolites and the concentrations of lactose and minerals. The main goal of this study was to evaluate nuclear magnetic resonance (NMR) data of raw milk samples without pre-treatment as a fingerprint for the deep reinforcement learning (DRL) protocol and to develop a reliable predictive model.

Results: The use of ¹H-NMR and diffusion-ordered spectroscopy (DOSY) NMR enabled identification of the most important metabolites in cow's milk and the composition of raw milk from healthy cows (two groups) and from cows suffering from subclinical (four groups) and clinical (one group) mastitis. Based on the obtained data, each sample was classified according to the stage and causative agent of mastitis. The 2D DOSY NMR spectra were further coupled with advanced multidimensional tensor decomposition methods and subjected to DRL. A combination of NMR spectroscopy and advanced tensor decomposition algorithms with state-of-the-art (DRL) methods was used to classify raw cow's milk according to different stages of mastitis.

Conclusion: An accurate and efficient classification model that comprises the chemical composition of the studied samples was built using the deep neural network. Classification accuracy based on the half-size training set was an excellent 82%. This model can further be exploited for diagnosis of diseases that affect the quality and content of milk in animal husbandry. © 2025 Society of Chemical Industry.