Application of pattern recognition and feature extraction techniques to volatile constituent metabolic profiles obtained by capillary gas chromatography.

The applicability of threshold logic units, a form of nonparametric pattern recognition, to the processing of metabolic profile data obtained by high-efficiency glass capillary column gas chromatography has been investigated. The test data included profiles of the volatile constituents of urine from normal individuals and from individuals with diabetes mellitus. A feature extraction algorithm allowed for dimensionality reduction and indicated the constituents most important in the normal versus pathological distinction. With an optimum number of dimensions, a normal versus pathological prediction rate of 93.75% was achieved. Gas chromatography-mass spectrometry was utilized to identify important profile constituents.