Custom MEA with Spike Interface

[vijithan99]

I have been using a custom MEA with 8 x 8 config with 60 electrodes. I ran my raw recordings through SpyKing Circus on Spike Interface to get my units which then I used Phy to filter out the good units.

I have two issues that arise out of this.

1. My spike sorted units are showing that my spikes are right over the unit which I believe is the reason why I get 0 distance in the dist_to_max_chan in the decay_and_average_waveform function. This causes issues as I am unable to divide by 0 and so I solved this by manually setting my dist_to_max_chan array to a very small value. I am not sure if this is the correct fix for this issue however.
   

2. My next issue is when calculating quantiles I get a negative value as in line 126 of overlord.py where we calculate prior_match, the param['n_expected_matches'] is well above len(include_these_pairs) which makes the value negative and leads to a negative quantile value. Not sure how to solve this problem without clamping the prior_match to a value between 0 and 1.
   

Would appreciate some support on if I am on the right track or how I can fix these issues.

[EnnyvanBeest]

Hi,

UnitMatch was originally written for high-density electrodes. What is the spacing between your channels on your MEA? Do you typically record waveforms from the same neuron on multiple channels or just on one channel?

Your solution is sensible, but that probably means some of these distance features may be meaningless and potentially UnitMatch even performs suboptimal if you include these. You could try to run unitmatch in different ways and see if that works for your specific electrode type. For example, you can change the default parameter param.channel_radius and max_dist to a larger number to increase the number of electrodes included. Changing max_dist will also help with getter a larger len(include_these_pairs) to avoid negative values for prior_match. Have a look at other default_params as well to see if they make sense for your electrode.

Another option is to remove the distance based scores from the 'scores_to_include' (Line 110 of 'overlord.py), so that UnitMatch assigns probabilities based on waveform shapes rather than decay and distances between units.

Hope this helps you find a good configuration for this type of electrode. Good luck!

[vijithan99]

Hi,

Thanks for your reply!

The probe spacings are around 200 um which might be pretty high. I think sometimes neuron spikes do appear on multiple channels but usually its single unit. Do you know what are some good parameters for a 200 um spaced 8 x 8 electrode grid is? I will try the changes you have mentioned above. Another concern I have is that my protocol involves performing a theta burst stimulation and seeing how the waveforms change. This approach should be robust to detecting units that are changing over time right? Is there anything I should change if this is the case?

Thanks again!

[EnnyvanBeest]

If you mean indeed there's 200 micron in between each pair of electrodes, it's quite unlikely you measure the same unit on multiple channels. I would as advised above remove the paramters that rely on relative distance from 'scores_to_include', which are spatial_decay_score, centroid scores, and trajectory, keeping only amplitude and waveform information. That will be the only reliable source to compare waveforms with low spatial resolution I'm afraid. Hopefully it will still be enough information to get accurate matches! This has not been tested outside of Neuropixels properly, so please use the GUI to see if results are sensible.

to your other question, even if the waveform changes slightly, it is probably going to be more similar within or across neurons. However, this is an assumption. So either way, you will really need to use the GUI to see if these assumptions seem valid.

[lee-wanhee]

Hi, thank you for the helpful explanations above.

I am working with a Utah Array setup (8×8 grid, 400 µm spacing), and I have run into similar issues as the original questioner. Since neurons usually appear on only a single channel at this spacing, I understand the distance-based features (spatial decay, centroid, trajectory, etc.) should be removed.

1. After removing those spatial features, which features does UnitMatch still rely on (e.g., amplitude and waveform similarity)?
2. Do you think this reduced feature set is sufficient for low-density arrays like Utah arrays, or would you recommend alternative approaches?

Any comments or suggestions would be much appreciated!

[EnnyvanBeest]

After these features are removed UnitMatch will only rely on temporal waveform features, so indeed the similarity of the waveform and amplitude. It may be enough as your training data will also only rely on these features, but I don't know for sure as I have never tried myself.
I recommend carefully evaluating the within-session performance, and also check cross-session performance with secondary functional validation (e.g. use AUC values for ISI histogram correlations between matches and non-matches).
Most other methods would rely on similar temporal waveform features, or perhaps already use functional measures in addition for the matching itself rather than using it as validation.