Adsa Logo White Adsa Title White

A method to diagnose mid-infrared milk analyzer prediction equation performance.

M. Portnoy


A method to diagnose mid-infrared milk analyzer prediction equation performance.
M. Portnoy*, D. M. Barbano. Department of Food Science, Northeast Dairy Food Research Center, Cornell University Ithaca, NY.

Mid-infrared (MIR) milk analyzers are used for milk payment and product testing. Our objective was to determine if a modified milk calibration sample set could be used to diagnose and identify weaknesses in both partial least squares (PLS) based and traditional fixed-filter based predictions of milk component concentration. The modified milk calibration set (14 samples with a wide range of fat, protein, lactose and urea), is formulated in an orthogonal design and all-lab mean reference chemistry that, allows the identification of specific weaknesses in MIR prediction equations that are due to incorrect compensation for variation in the background milk matrix effects of fat, protein, and lactose concentration. In the case of traditional fixed-filter prediction models, the calibration equations can be adjusted based on the results of analysis of the modified milk set to improve instrument accuracy, while in the case of PLS models, specific model weaknesses can be identified and pointed out to the PLS model developer. For traditional filter models that predict fat, protein, and lactose, the sensitivity of predicted values to a mismatch of the intercorrection factor settings with the instrument optic system characteristics caused the standard deviation of the difference (SDD) between instrument prediction and reference chemistry to be larger (e.g., SDD of 0.004 vs 0.021 for fat when intercorrection factor for protein on fat B differs by 0.03), with systematic under or over estimation of the component being predicted at the ends of the concentration range of the interfering milk component. For PLS models, the inability of a PLS model to cancel out the background matrix variation effects of fat, protein, and lactose concentration on the parameter being predicted can be clearly identified and quantified. Based on this diagnostic data that can be produced by analysis of the modified milk samples, the population of milk sample spectra that need to be added to the PLS modeling population to improve the prediction accuracy of a PLS model measuring major milk components, or for prediction models for minor milk components (e.g., milk urea nitrogen or fatty acids) can be determined.

Keywords: mid-infrared, intercorrection factor, partial least squares models.

Biography: Matilde Portnoy is currently a PhD student in the Department of Food Science at Cornell University in Dr. David Barbano's laboratory, and just recently finished her MS degree at Cornell's Food Science Department as well. She is originally from Puerto Rico, where she obtained her BS degree in Chemistry in 2017 from University of Puerto Rico.