Rapid and Non-Destructive Prediction of Major Milk Fatty Acid Classes Using Handheld Near-Infrared Reflectance Spectroscopy Coupled with Chemometric Analysis

This study developed a non-destructive analytical approach based on handheld near-infrared (NIR) reflectance spectroscopy coupled with chemometric modeling for predicting major fatty acid classes in bovine milk. A total of 560 individual-cow milk samples from Holstein cows were collected from a single commercial dairy farm during one sampling period and analyzed using gas chromatography–mass spectrometry (GC–MS) as reference measurements. The concentrations of saturated fatty acids (SFA), monounsaturated fatty acids (MUFA), and polyunsaturated fatty acids (PUFA) determined by GC–MS were used as reference values for NIR prediction model development. Reflectance NIR spectra were collected over 900–1700 nm using a portable NIR spectrometer from previously frozen milk samples maintained at approximately 4 °C during measurement, and three replicate scans per sample were averaged. Seven preprocessing methods, multiple wavelength ranges, and different numbers of selected feature wavelengths were compared. Partial least squares regression models were developed using sample-level five-fold cross-validation. The optimal models for SFA, MUFA, and PUFA used raw, first-derivative, and Savitzky–Golay smoothed spectra, respectively, with CV-R2 values of 0.931, 0.887, and 0.910, RMSECV values of 0.2076, 0.1095, and 0.0120 g/100 g milk, respectively, and RPD values of 3.818, 2.988, and 3.342. These results indicate that handheld NIR reflectance spectroscopy combined with optimized chemometric strategies has potential for predicting major fatty acid class concentrations in bovine milk. However, the present models were evaluated only by internal cross-validation using samples from a single farm and breed during a limited sampling period; independent external validation across farms, seasons, breeds, and management conditions is required to further assess model transferability and generalizability.