Classifier-Directed Signal Processing in Brain Research

Abstract

Because of the difficulty of extracting useful information from brain electrical or magnetic field measurements, sensitive analytic methods are often required. "Open-loop" techniques for the choice of signal features and the testing of statistical hypotheses are often not sufficient for such problems. The sensitivity of analyses can be increased by "closed-loop" analyses which use feedback from the hypothesis testing to optimize the feature extraction and/or primary analysis to achieve maximal classification accuracy for a pattern recognition analysis which attempts to separate experimental or ciinical conditions. Signal processing algorithms whose parameters are set to maximize the strength of consequent inferences as measured by classifier performance could be called classifier-directed methods. This paper reviews the application of classifier-directed methodologies to waveform detection and categorical classification problems in brain research. Pattern recognition methods are shown to be a convenient way of incorporating expert knowledge in a statistical framework with minimal assumptions about the statistics of the desired or undesired components.