Classification of single-trial ERP sub-types: application of globally optimal vector quantization using simulated annealing

Abstract

Examination of the single trials which are traditionally averaged to form late-component ERPs reveals a number of different sub-types of response. This study introduces an automated and robust approach to objectively classify these ERP sub-types. Auditory oddball ERP (target tones) data were examined in 25 normal subjects. Globally optimal vector quantization using simulated annealing (the “Metropolis algorithm”) was employed to determine the natural groupings of the single-trial responses that constitute the average ERP. No prior assumptions about the ERP patterns were imposed. This is the first study to employ a cluster analysis technique with globally optimal properties in ERP research. We demonstrate that, due to the presence of many different undesirable local minima, a globally optimal solution is crucial if the classification of the single-trial ERPs is to reflect their real structure. The results of this study showed that only around 40% of single trials had a morphology which resembled the averaged ERP wave form. The remaining single trials had a response morphology which was different from the average, in terms of the amplitude and latency of the components. Single-trial ERP response sub-types may provide fundamental complementary functional information to the ERP average.