tell user about undetected spikes

[nsteinme]

One idea: run a naive threshold-crossing detection at the end and report times, channels, and amplitudes of anything that is left? have text or graphical output of the rate of such spikes relative to the rate of spikes that were detected?

[marius10p]

A lot more stuff gets detected now with the threshold in the final pass being so low (ops.Th(2) = 4). There should be very little left, but let's wait for user reports.

[czuba]

Just bumping this issue.

Some kind of post-sort thresholded output would be particularly useful for:
• catching undetected isolatable spikes (informing user parameterization issues, and possibly codebase issues)
• allowing some level of user-interpretable evidence of silently/automatically dropped clusters that the user may want/need to recover (...also useful for refining sort & merge parameters)
• providing an equivalence to classical multi-unit responses (e.g. hand-sorted outputs that are clearly neither noise, nor purely isolatable SU)

[susuchen66]

Hi we observe in about 5% large amplitude units (>150 uV) missing spikes while sorting different recording session, including long (>1hr) or short (5 mins) epochs. There could be more subtle undetected problems like this in other units. Lowering Thr1 and Thr2, minimal firing rate per channel, etc didn't help. Missing spikes are observed in the Matlab GUI Nick developed, Phy, and JRClust GUI. We used JRClust to sort same dataset and compared the result with KS2. We found large amplitude spikes dropped out over large stretches of session in KS2 but not in JRClust.

Comparison between JRClust output (left) and KS2 output (right) on the same unit:

[marius10p]

Closing for lack of activity.