Comparison and predictability of epileptic seizures by a linear and nonlinear method. McSharry, P.E., Smith, L., Tarassenko, L. IEEE Transactions on Biomedical Engineering 50 (5): 628-633, May 2003
Abstract
This paper contrasts the performance of traditional linear statistics with nonlinear statistics for identifying epileptic seizures from changes in scalp EEG recordings. The conceptual foundation of modern nonlinear techniques is discussed and simplifications to improve both their power and clarify the clinical understanding of their inner workings are suggested. In particular a conceptually simple nonlinear statistic based on the evolution of the probability density function within a multi-dimensional state space is introduced, and compared with linear statistics. A synthetic recording is employed to illustrate that this nonlinear statistic is required to detect changes associated with an increase of the nonlinearities in the underlying dynamics. For the scalp recordings investigated here, this nonlinear statistic did not perform very differently from a linear statistic such as variance, except that it may generate fewer false positives.