Adaptive stimulus artifact cancellation in biological signals using neural networks

The recording of somatosensory evoked potentials (SEP) is very important today in diagnostic and intraoperative procedures. Ensemble averaging improves the signal-to noise ratio (SNR) by reducing the random interference. Ensemble averaging does not reduce the stimulus artifact which tends to mask or at least distort the SEP. The artifact is a result of the relatively large voltage applied to the body in order to elicit a nervous response, and is thus synchronized with the SEP. Several adaptive cancellation techniques have been used to reduce the stimulus artifact, but these techniques have typically assumed linearity between the primary and reference channels. Neural networks offer the advantage of being able to model nonlinearities. A neural network structure called Pi-Sigma is presented and the resulting cancellation of stimulus artifact in SEP data is shown. The results are compared to cancellation obtained using linear filters and a nonlinear RLS filter