Non-Targeted Detection of Botanical Adulteration
Non-targeted detection of botanical adulterants is conceptually simple. Analyze authentic materials, build a model, establish acceptable statistical limits, and then determine if the unknown material lies within those limits. But, the devil is in the details. The definition for “authentic” should be as tight as possible, otherwise the variance of the model will increase and the model will be more susceptible to false positives. Collection of authentic samples will be dependent on the definition of “authentic”, the degree of natural variability, availability, and cost. All analyses are biased by the physical nature of the sample. Analysis of solid samples will be biased in favor of the macro components. Analysis of extracts can eliminate macro components but will be biased by the polarity of the solvent. If the adulterant is truly unknown, then multiple extraction and detection systems may be needed. Building a model is simplest if the adulterant is known. The model can then incorporate features of the authentic and the adulterating material. This allows prediction of adulterated spectra, use of sophisticated two-class or multi-class classification methods, and prediction of the limit of detectable adulteration. If the adulterant is unknown, modeling is limited to exploratory methods (PCA) and one-class classifiers (SIMCA). Whether a sample fits the model can be judged in two ways. The Hotelling T2 statistic indicates how well the sample fits within the model parameters. The Q statistic, also known as the squared prediction error, indicates how far the sample lies outside the model parameters. The simpler the model (based on only 1 or 2 principal components) the more useful the Q statistic. However, the limit of detectable adulteration cannot be determined until the adulterant is identified.