Event-related potentials: a critical review of methods for single-trial detection

The analysis of ERP data has followed several lines over the last 20 years. The most prevalent method is simply to average ERPs for a given class of stimuli. The ERPs are compared for differences across classes of stimuli. Little other special data processing is used. The ERP comparisons are usually performed using visual examination of the wave-shapes. Sometimes statistics are calculated such as means, variances, and confidence limits. Linear filtering is used to reduce interference. Another approach is to model or analyze the ERP as a sequence of vectors or frames of data samples. These samples may be of the ERP time waveform or they may be of the frequency transform of the ERP waveform. The frames of data vary in length from the entire ERP waveform (500 to 1000 msec) to frames as short as ten sample points (100 msec). Recognition of an event in the ERP is achieved by computing a distance measure between parameter vectors for one class of stimuli and corresponding parameter vectors for another class of stimuli. Recognition is achieved by selecting the ERP with the lowest distance score. This approach is "pattern matching" and relies on two assumptions: adjacent frames of data are uncorrelated, and the variability of the data can be accounted for by the distance measured for all stimuli in the classes presented. Subject variability is generally not accounted for, other than to assume it is the same for all classes of stimuli. The data are clustered into a variety of reference patterns that represent particular manifestations of a particular stimulus. Another approach is "feature-based" recognition. The idea is to identify and automatically extract features of the data that can provide a characterization of stimuli. The features selected may be abstract. They are calculated from the data or transforms of the data.