Application of singular value decomposition to topographic analysis of flash-evoked potentials

Singular value decomposition is a robust numerical method for decomposing a matrix of multichannel EEG or EP data into a sharply reduced set of features with corresponding waveform, amplitude, and spatial vectors. In 19 normal subjects aged 19 to 40 years, the three largest features computed by the SVD algorithm accounted for 93-98 percent of the total variance of the averaged flash-evoked potential. There was good separation of major brain areas as well as clustering of related electrode sites. Orthogonal rotation of the three spatial vectors is essential to see clustering of brain areas across subjects. Three-dimensional display showed the regular presence of orthonormal occipital, frontopolar, and vertex spatial vectors. Since the spatial feature vectors cluster tightly and yet are orthonormal, statistical comparison of patients with normal control groups will be facilitated.