Chronically implanted electrode arrays have enabled a range of
advances, particularly in neural prosthetics. Initial human clinical
trials are underway for prototype neural prosthetic systems that aim
to help disabled patients interact with the world. However,
characterization of the stability of the electrode arrays and their
neural recordings is limited by current experimental protocols, which
only provide a few hours of continuous data each day. The HermesB
system, an autonomous, long-duration neural recording system for
freely behaving non-human primates, enables the collection of
continuous multi-day broadband neural datasets (Gilja et al. in this
volume). Datasets recorded with HermesB from a Cyberkinetics array
implanted in an adult Rhesus monkey show significant variability in
waveform amplitude, up to 30% relative to the mean, observed across 54
hours. This variability occurs over multiple timescales and may be
influenced by a number of factors brain , changes in intracranial
pressure (ICP), or other homeostatic factors. For 5 minute blocks the
variability is highly correlated to mean firing rate (mFR);
correlation coefficients of -0.21 and 0.69 were found for two units,
with similar characteristics observed for units on other channels. A
spectral analysis of waveform amplitude over time shows significant
power at ~1 cycle/day; this modulation is also seen in the mFR, which
is a good proxy of general activity. Abrupt changes in waveform
amplitude (~ 20%) were observed which were highly correlated in time
with large acceleration events (defined by a 3g+ movement of the
head). These amplitude changes decay over a time period of minutes and
may correspond to abrupt array movement. These preliminary results
suggest that recorded spike waveforms cannot be considered stable
across any timescale. For basic science applications, when attempting
to compare neural responses day-to-day, the waveform variations would
imply it is essential to verify the identity of each neuron. When
recording from individual neurons for use in neural prostheses,
adaptive spike sorting algorithms capable of contending with abrupt
waveform change and slow drift are most likely needed.