Detection of human nerve signals using higher-order statistics

Afferent, whole nerve signals recorded using an implanted nerve-cuff electrode were analyzed using three detectors based on the 1st, 2nd and 3rd order statistical properties of the signals. Results based on standard rectified, bin-integrated (1st order statistical) processing are compared with two algorithms based upon a singular value decomposition (SVD) of the signal's 2nd and 3rd order correlation (cumulant) matrices. Due to the very low signal levels obtainable from nerve-cuff electrodes and the high levels of interference from adjacent muscles, the overall signal-to-noise ratio (SNR) is very poor. In addition, the noise level is non-stationary. The inherent properties of the 3rd order statistics of these signals yield a detector that performs better than the other two.