CatGT: Tshift, CAR, Gfix

Tshift

Imec probes digitize the voltages on all of their channels during each sample period (~ 1/(30kHz)). However, the converting circuits (ADCs) are expensive in power and real estate, so there are only a few dozen on the probe and they are shared by the ~384 channels. The channels are organized into multiplex channel groups that are digitized together, consequently each group's actual sampling time is slightly offset from that of other groups.

CatGT automatically applies an operation we call tshift to undo the effects of multiplexing by temporally aligning channels to each other. This improves the results of operations that compare or combine different channels, such as global CAR filtering or whitening. The FFT-based correction method was proposed by Olivier Winter of the International Brain Laboratory.

Figure: Licking artifact (fast transient on all channels), 1200 uV full-scale:

Figure panels (from SpikeGLX FileViewer):

  • Left: Raw traces; 100 ms time span.
  • Middle: Raw traces zoomed-in; 1 ms time span.
  • Right: Aligned by CatGT tshift; 1 ms time span.

<BR/>

Artifact Corrections {CAR, Gfix}

CAR (CatGT -gblcar option)

Global common average referencing (CAR) calculates (for each sample) the average over all channels, and then subtracts that from each channel. This does a great job of cancelling artifacts that are indeed similar on all channels and that haven't saturated on any channels.

Zeroing (CatGT -gfix option)

Sometimes the hardware experiences a large electrical insult that affects all channels in a nonuniform way. This may arise from licking, chewing, head banging or laser stimulation events. In these cases CAR won't work as well. CatGT can detect these events and replace them with zeros.

This option is specified, for example, as follows: -gfix=0.40,0.10,0.02. This requires that a transient artifact must attain a peak amplitude of at least ||0.40 mV||, must reach its peak as least as rapidly as ||0.1mV/sample-tick||, and the transient is over when the amplitude settles back to ||0.02 mV|| or smaller.

These values are easy to read by inspecting examples with the SpikeGLX FileViewer.

Furthermore, at least 25% of the channels must show a transient with these characteristics. Such timepoints are flagged. After all other filtering has been applied, the data are edited to replace the artifact span with zeros in all channels.

It does make sense to use -gblcar and -gfix options together. Editing removes extreme artifacts completely, while CAR removes common background noise at all timepoints, not just at extrema.

Figure: Variety of artifacts spanning all channels, 240 uV and 100 ms full-scale:

Figure panels (from SpikeGLX FileViewer):

  • Left: Uncorrected.
  • Middle: -gblcar; Smaller artifacts removed.
  • Right: -gblcar -gfix; All artifacts removed.

<BR/>

What is global demux (gbldmx)?

The CatGT gbldmx option separates the channels into groups that are sampled concurrently so suffer no time shifts within the group. CAR is then performed on a group by group basis which has advantages and disadvantages compared to doing (tshift + gblcar).

Disadvantages:

  • Because the number of channels in a multiplex group is small the averages are noisier than we can achieve by tshifting all channels into one large group and applying CAR after that.

  • Outlier channels have a stronger biasing effect with smaller sample sizes.

  • The correction factors tend to be larger (small channel count in denominator) than with CAR, and that can produce overcorrection artifacts that look like spikes (often small inverted spikes).

When to use gbldmx:

  • The probe sampling rate is 30 kHz, so noise faster than 15 kHz is not adequately sampled, and therefore, not accurately propagated to other channel groups by the tshift operation. So for very fast noise artifacts, gbldmx may produce better cancellation within each subsampling group.

  • gbldmx usually outperforms gblcar when tshift is omitted. You may want to omit tshift in CatGT, using the -no_tshift option, if tshift will be applied by some later component in your analysis pipeline.

What is local CAR (loccar)?

CatGT offers the option to do local CAR over a ring of channels. The idea is that for each channel a disk is used to average the signal of neighbor channels, but an inner disk (containing a potential spike footprint) is excluded from the average. You get to specify the inner and outer disk radii in units of channels.

The downsides to this are:

  • Its hard to tune the disk sizes to the actual spike footprints, especially if the probe spans several brain regions with different characteristic neuron sizes, and considering that neurons are at a variety of distances from the probe.

  • Although loccar appears to have good noise reduction, more often than not the true spike amplitude is more severely attenuated because the footprint is larger than you expect.

When to use loccar:

  • If background is not uniform over the whole probe. If the probe spans dissimilar brain regions, you may do better averaging more locally. Try setting a fairly large outer radius to balance locality vs smoothness.

What about LFP?

Tshift is valuable for aligning LFP channels. CAR is not recommended when analyzing low frequencies: LFP varies slowly over the whole shank, making distant channels a bad reference for correcting noise.

fin