Beyond averaging: the use of discriminant functions to recognize event related potentials elicited by single auditory stimuli.

A test of the stepwise discriminant analysis (SWDA) procedure for assessing single-trial event related potentials (ERPs) is presented. Discriminant functions (DFs) were built from a data base composed of single-trial ERPs from sixteen subjects who were presented trains of loud and soft tones. Loud tones occurred randomly on 10% of the trials. Subjects either counted the rare--loud stimuli or solved a hidden-word puzzle. Various DFs at three electrode sites (Fz, Cz and Pz) were obtained to assess the feasibility of performing pairwise discriminations between the various combinations of events which are defined by this procedure. For the pair of events which yielded the most striking differences between their average ERP waveforms it was possible to classify correctly, an average of 84% of the events using information from one electrode site, and 89% of the events if information from multiple electrode sites was used. A "subject-independent" DF was developed from these data and applied to data obtained from seven new subjects. This subject-independent function proved to be sufficiently generalized to classify correctly 81% of the trials. The nature of classification errors by this procedure is discussed.