Standardization of milk infrared spectra for the retroactive application of calibration models

Abstract

The objective of this study was to standardize the infrared spectra obtained over time and across 2 milk laboratories of Canada to create a uniform historical database and allow (1) the retroactive application of calibration models for prediction of fine milk composition; and (2) the direct use of spectral information for the development of indicators of animal health and efficiency. Spectral variation across laboratories and over time was inspected by principal components analysis (PCA). Shifts in the PCA scores were detected over time, leading to the definition of different subsets of spectra having homogeneous infrared signal. To evaluate the possibility of using common equations on spectra collected by the 2 instruments and over time, we developed a standardization (STD) method. For each subset of data having homogeneous infrared signal, a total of 99 spectra corresponding to the percentiles of the distribution of the absorbance at each wavenumber were created and used to build the STD matrices. Equations predicting contents of saturated fatty acids, short-chain fatty acids, and C18:0 were created and applied on different subsets of spectra, before and after STD. After STD, bias and root mean squared error of prediction decreased by 66% and 32%, respectively. When calibration equations were applied to the historical nonstandardized database of spectra, shifts in the predictions could be observed over time for all investigated traits. Shifts in the distribution of the predictions over time corresponded to the shifts identified by the inspection of the PCA scores. After STD, shifts in the predicted fatty acid contents were greatly reduced. Standardization reduced spectral variability between instruments and over time, allowing the merging of milk spectra data from different instruments into a common database, the retroactive use of calibrations equations, or the direct use of the spectral data without restrictions.

Strategies for noise reduction and standardization of milk mid-infrared spectra from dairy cattle

“…The effectiveness of standardization for reducing prediction errors when transferring calibration models between instruments for fat composition traits has been demonstrated previously (Grelet et al, 2015;Bonfatti et al, 2017a). Grelet et al (2017) also demonstrated the effectiveness of standardization for reducing prediction errors for traits with lower quality calibration models such as methane emissions and cheese yield.…”

“…A third source of noise is the variation between instruments or within instruments across time. This variation, which exists even between instruments of the same brand, can result in prediction errors and bias and is particularly problematic when applying calibration models developed on one instrument across a historical database of spectra collected on different instruments (Grelet et al, 2015;Bonfatti et al, 2017a).…”

“…With the growing number of traits predicted from spectra, many of which do not have reference samples, there is increased interest in standardizing individual FTIR spectra wavenumbers directly (Grelet et al, 2015Bonfatti et al, 2017a). Recent publications by Grelet et al (2015) and Bonfatti et al (2017a) present methods for standardizing individual wavenumber absorbance values. Both strategies involve assigning a primary instrument and standardizing spectra from other (secondary) instruments to align them to the spectral response of the primary instrument.…”

“…Grelet et al (2015) presented a piecewise direct standardization (PDS) approach based on the method described by Wang et al (1991). Bonfatti et al (2017a) presented a retroactive approach (retroactive percentile standardization, RPS), in which percentiles of spectral responses for each wavenumber are used to map the relationships between primary and secondary instruments.…”

Standardization of milk mid-infrared spectrometers for the transfer and use of multiple models

“…However, the developed method is valid only once a slave instrument has been integrated in the network and has analyzed the standardization samples. Recently, a study aimed to standardize spectra over instruments and over time using historical data sets as a basis (Bonfatti et al, 2017a). This method allows the harmonizing of historical databases and the use a posteriori of models when instruments have not been standardized.…”