Retention of learning for adaptive filtering of evoked brain potentials

An in-depth analysis of the learning behavior of the least-mean-square adaptive algorithm was performed. Simulation studies show that the learned filter coefficients decay at a rate proportional to the convergence constant following the disappearance of the underlying signal. This conflicts with the need for rapid convergence when the signal changes, as is the case for the evoked potential. Studies involving mean-square error measurements show that significant improvements to weight retention can be obtained by using a gating algorithm which updates the weights only when the signal is present. A symmetric noncausal format for the weights was shown to extract the initial peak of the signal with the most consistency. After an investigation of several filter configurations, averaging was found to function approximately as well as each filter when attempting to estimate the single-response evoked potential. Yet, for the cases when detection of loss of the signal is most crucial, the adaptive filter offers more advantages than averaging. Results from applying the proposed algorithm to simulated evoked potential buried in prestimulus electroencephalogram are presented, and human visual evoked potential results are analyzed