Use of pyrolysis mass spectrometry with supervised learning for the assessment of the adulteration of milk of different species

Binary mixtures of 0-20% cows' milk with ewes' milk, 0-20% cows' milk with goats' milk, and 0-5% cows' milk with goats' milk were subjected to pyrolysis mass spectrometry (PyMS). For analysis of the pyrolysis mass spectra so as to determine the percentage adulteration of either caprine or ovine milk with bovine milk, partial least-squares regression (PES), principal components regression (PCR) and fully interconnected feed-forward artificial neural networks (ANNs) were studied. In the latter case, the weights were modified by using the standard back-propagation algorithm, and the nodes used a sigmoidal squashing function. It was found that: each of the methods could he used to provide calibration models which gave excellent predictions for the percentage adulteration with cows' milk to <1% for samples, with an accuracy of +/-0.5%, on which they had not been trained. Scaling the individual nodes on the input layer of ANNs significantly decreased the time taken for the ANNs to learn, compared with scaling across the whole mass range; however in one case this approach resulted in poor generalization for the estimates of percentage cows' milk in ewes' milk. To assess whether the calibration models had learned the differences between the milk species or the differences due to the different fat content of in each of the milk types, we also analyzed pure milk samples varying in fat content by PyMS. Cluster analysis showed unequivocally that the major variation between the different milk species was not due to variable fat content. Since and biological material earn be pyrolyzed in this way, the combination of PyMS with supervised learning constitutes a rapid. powerful, and novel approach to the quantitative assessment of food adulteration generally.