Classification and identification of mass spectra of toxic compounds with an inductive rule-building expert system and information theory

An expert system for classifying and identifying low-resolution mass spectra of toxic and related compounds was developed with an expert shell program. The shell system used was an inexpensive, rule-building software package with an implementation of the ID3 algorithm. Seventy-eight target compounds were used to establish classes previously found by SIMCA class modeling. The six classes included nonhalobenzenes; chlorobenzenes; bromoalkanes and bromoalkenes; mono- and di-chloroalkanes and the analogous alkenes; tri-, tetra- and penta-chloroalkanes and the analogous alkenes; and unknowns. Identification modules for the target compounds were forward-chained to the classification modules. An expert system based on binary-encoded mass spectra, with 17 masses selected on the basis of information content, gave 97 and 86% classification accuracy for training and test spectra, respectively. Identification accuracy was 77 and 80%, respectively. An expert system was also developed which was based on ternary encoding of the mass spectra of 108 training compounds using 25 masses. Ternary encoding has many of the advantages of binary encoding, without the disadvantages. This latter system was tested with the spectra of thirty compounds found in field samples or potential air pollutants. The classification accuracy for training and test spectra was 99 and 97%, respectively. The identification accuracy was 96 and 93%, respectively. With proper precautions, the rule-building expert system can be very effective in spectral classification and identification problems.