Use of a multivariate moving window PCA for the untargeted detection of contaminants in agro-food products, as exemplified by the detection of melamine levels in milk using vibrational spectroscopy

In this study, the concept of a Local moving window along the wavelength range in vibrational spectroscopic data was used to build reduced PCA models for characterizing agro-food products and detecting the presence of unusual ingredients or contaminants in an untargeted way. For each selected wavelength window in a locally reduced calibration set, a PCA analysis was performed and score residuals were extracted and used as to define thresholds to be applied to the spectral score residuals of the sample being investigated. When a residual at a certain wavenumber exceeded defined thresholds, the sample was suspected of being abnormal, indicating the possible presence of unusual ingredients and allowing non-targeted analysis. The method was applied to liquid UHT milk samples spiked with varying levels of melamine. Samples spiked at levels higher than 100 ppm were easily detected using this method, which would not have been possible using classical techniques such as Mahalanobis distance, usually applied as an outlier detection method.