Automatic noise-level detection for extra-cellular micro-electrode recordings

Abstract

Extra-cellular neuro-recording signals used for functional mapping in deep brain stimulation (DBS) surgery and invasive brain computer interfaces, may suffer from poor signal to noise ratio. Therefore, a reliable automatic noise estimate is essential to extract spikes from recordings. We show that current methods are biased toward overestimation of noise-levels with increasing neuronal activity or artifacts. An improved and novel method is proposed that is based on an estimate of the mode of the distribution of the signal envelope. Our method makes use of the inherent characteristics of the noise distribution. For band-limited Gaussian noise the envelope of the signal is known to follow the Rayleigh distribution. The location of the peak of this distribution provides a reliable noise-level estimate. It is demonstrated that this new 'envelope' method gives superior performance both on simulated data, and on actual micro-electrode recordings made during the implantation surgery of DBS electrodes for the treatment of Parkinson's disease.

Neural network dynamics in Parkinson's disease

“…Spikes of probably 2 neurons For detailed method for spike sorting see Appendix C. In short, local maxima of the MER signal that exceed 4.5 times the noise-level, which was estimated with the 'envelope' method [51], were flagged as a spike event. Subsequently, the waveform of each detected spike was obtained from the MER by extracting a data window of 1.2 ms around the time of the spike event.…”

“…[51]. Middle: spike waveforms, extracted from the 18-s long MER fragment and aligned to their peak at time zero.…”

“…The noise-level could not be directly measured and had to be estimated from the whole MER signal (noise, spikes and artefacts). For noise-level estimation we used the mode-based envelope method as proposed by Dolan et al [51], under the assumption that the noise of the MER fragment was Gaussian distributed. This is a reasonable assumption, as the background noise signal is a superposition of distinct independent signals.…”

Selective stimulation of the subthalamic nucleus

“…To assess multi/single unit activity, each epoch was digitally filtered between 350 and 5000 Hz. Spike detection was performed using the envelope method (Dolan et al 2009). To obtain single unit activity, spike sorting was performed by computing the principal components, which were clustered using either K-means or the Gaussian mixture model and the expectation maximization algorithm (Lewicki 1998).…”

“…2. Artifacts in the spike signal were detected by first estimating the noise-level using the spike signal's envelope (Dolan et al 2009). The maximum value of the rectified spike signal was not allowed to exceed this value multiplied by a factor of 20.…”