Early prediction of the shelf-life of medium-heat whole milk powders using stepwise multiple regression and principal component analysis (original) (raw)

1997, International Dairy Journal

Fifteen medium-heat whole milk powders manufactured at the same plant under identical processing conditions from 15 different batches of raw milk were subjected to accelerated storage at 50°C and a water activity of 0.3 1 in air. With a view to the development of a new prediction method, 22 chemical and technological variables and 13 time-dependent changes over 2 days were measured at the beginning of the experiment. The development of oxidized flavour in the stored powders was subsequently assessed weekly for 6 weeks by a sensory panel. The shelf-life of each powder was determined as the time at which the mean flavour score crossed a threshold value (8 or 6 on a scale from 15 to 0). The relationship between shelf-life and the properties of the fresh powders, together with the time-dependent changes, was investigated by the use of stepwise multiple regression with maximum R2 improvement, and pseudo-optimal sets of regression equations were found which minimized the PRESS (predicted sum of squares) statistic. The initial time-dependent change of TBA-reactive substances was the variable which correlated best (r = -0.66) with the shelf-life at flavour threshold 8. followed by the yellowness Hunter b-value of the powder (r = 0.54) while the variables which alone correlated best with the shelf-life at flavour threshold 6 were the density of milk powder after 1250 tappings (r=O.58), TBA-reactive substances (r= -0.56), and the density of milk powder after 100 tappings (r=O.53). The variables selected by PRESS minimization for the regression equation of threshold 8 (R2 = 0.949) were the a-lactalbumin content, the time-dependent change in /?-lactoglobulin, the time-dependent change in TBA-reactive substances, the density after 1250 tappings, and the solubihty index. The variables selected for the regression equation of threshold 6 (R2=0.980) were the a-lactalbumin content, the time-dependent change in fllactoglobulin, the density after 1250 tappings, the total fat content, the time-dependent change in the redness Hunter a-value of the powder, and the initial moisture content. Good predictive abilities of the regression equations were indicated by small PRESS values, and the method appears ready for validation in industrial production. Differences between the milk powders were recognized by principal component analysis. One powder deviated clearly, but not extremely, from the other 14 powders on the first and second principal components, mainly because of a large initial amount of free radicals, a rapid generation of further free radicals, and a low initial concentration of fluorescent oxidation products. The first 7 principal components explained only 84% of the total variation, and no other groupings of the milk powders were apparent. ic? 1997 Elsevier Science Ltd. All rights reserved Keywords: medium-heat whole milk powder; sensory evaluation; shelf-life prediction; principal component analysis; multiple regression Guide for Personal Computers, Version 6 Edition. SAS Institute, Cary, NC. Stapelfeldt, H., Nielsen, B.R. and Skibsted, L.H. (1997) Effect of heat treatment, water activity and storage temperature on the oxidative stability of whole milk powder. International Dairy Journal 7, 299-307. Stapelfeldt, H. and Skibsted, L. H. (1994) Modification of fllactoglobuhn by aliphatic aldehydes in aqueous solution.