Reduction of stimulus artifact in somatosensory evoked potentials: segmented versus subthreshold training

A new approach to stimulus artifact cancellation is introduced, which attempts to model the process of stimulus artifact generation. This is done by training an estimator with multiple exemplars of the stimulus artifact at levels below the threshold of evoked response stimulation. Two estimators are formulated: one using a dynamic neural network and another using a linear estimator. The performance of these new approaches is compared to the segmented training approach, which has been previously demonstrated to be one of the most capable methods available. Performance assessment is carried out using a novel metric introduced in this paper, which focuses upon the relevant portion of the recorded waveform. The new cancellation schemes show distinct performance advantages over the segmented training approach.