Estimation of single-trial evoked signals by local transform domain filtering

An algorithm for the estimation and reconstruction of event related signals corrupted by colored noise (e.g. evoked potentials, masked by EEG activity and additional sources of clutter and noise) is proposed. The local Karhunen-Loeve (LKL) basis, derived from the noisy signal local autocorrelation function, is used for optimal signal representation (in minimum mean squared error-MMSE-sense). The vector space of the noisy signal is decomposed by the LKL transform into the corresponding complementary orthogonal subspaces, i.e. the signal-plus-noise and the noise only. The event related signal is estimated from the signal-plus-noise subspace by modifying the corresponding LKL components with a Wiener-like gain function