Frequency domain discriminant analysis of the electroencephalogram

A frequency domain generalization of the classical quadratic discriminant function was applied to the problem of classifying alpha-band multichannel electroencephalogram recordings in three task conditions. The data consisted of 41-channel recordings obtained in eyes closed, verbal, and spatial task conditions. Classifier performance was measured by deriving a decision rule from a training sample of 42 recordings and then applying the obtained rule to a test sample of 46 recordings. The proportion of correct classification was .93 in the training sample and .85 in the test sample. The classifier performed better when based on the complete cross-spectral matrix than when restricted to power spectrum variables. Classification based on a subset of 16 leads reduced the overall proportion of correct classification to .79 in the training sample and to .70 in the test sample.