The Effect of Pyrolysis and Instrument Parameters on the Results of Various Supervised Learning Techniques

Abstract

Four supervised pattern-recognition procedures—linear discriminant analysis, soft-independent models of class analogy, a univariate expert system, and a multivariate expert system—were applied to pyrolysis-mass spectrometry data from five bacterial species. A test set was generated by individually modifying the tuning parameters of the mass spectrometer followed by collection of one spectrum for each bacteria. Five replicate pyrolysis-mass spectra of each bacteria were used as a training set for the classification of the test set. The multivariate expert system consistently scored over 92% correctly identified, while the other procedures resulted in less than 80% correctly identified. Under these harsh conditions, the multivariate expert system was not affected seriously by the mass spectrometer tuning changes.