The use of principal component analysis for the structural interpretation of mid-infrared spectra

Abstract

Abstract Perkins, J.H., Hasenoehrl, E.J. and Griffiths, P.R., 1992. The use of principal component analysis for the structural interpretation of mid-infrared spectra. Chemometrics and Intelligent Laboratory Systems , 15: 75–86. Principal component analysis (PCA) is shown to be a useful method to characterize spectra for interpretation in an expert system. A protocol is described for the generation of a numerical rule that converts a mid-infrared spectrum to a classification score that indicates whether the compound in question belongs to one structural class or another. The protocol consists of six steps: (1) selection of a training set of compounds that serves as an example and counter-example of the structure under investigation; (2) autoscaling the training set spectra; (3) feature weighting the training set spectra to emphasize those measurements (wavenumbers) that best separate the classes; (4) performing the PCA on the data set; (5) projecting a second larger training set onto the first few principal components; and (6) determining an optimal separation surface and logistic sigmoid function. The sigmoid function returns a classification score between 0 and 1 indicating the presence of the structure. The closer the score is to 1, the greater the certainty that the functional group or substructure is present. The protocol is illustrated for the generation of a rule for determining the presence of the nitro functionality. The derived rule has a no-choice rate (probability that a spectrum will not be classified) of 4% and an overall correct rate of 97% as determined from a validation set of 1000 spectra.