CONNECTIONIST HYPERPRISM NEURAL-NETWORK FOR THE ANALYSIS OF ION MOBILITY SPECTRA - AN EMPIRICAL-EVALUATION

Spectra from ion mobility spectrometry (IMS) consist of low resolution information arising from complex gas-phase ion-molecular reactions. Such spectra lack specific structural information and have been considered inappropriate for analysis by traditional library search and pattern recognition methods. A data set of 292 mobility spectra from 12 organic compounds was treated using a connectionist hyperprism classification (CHC) neural network. Advantages of the CHC network for such spectra included high speed, linear training, and flexibility to adapt training parameters, to minimize false negative identifications. The principal disadvantages were a sensitivity to unnecessary spectral information in the training vectors and a high false positive ratio. However, sensitivity exhibited by the CHC neural network assisted in identifying dispensable spectral information and in suggesting the importance of spectral features.