Project: LFP extraction for quad base data

[mvdm]

Problem: LFP extraction currently fails for QB data

Spikeinterfaces has extract-LFP code and some documented issues/considerations-- @emma-gibbens to add

[emma-gibbens]

Please see 2026-neuropixels-hack/lfp_extraction for relevant files, or PR #6

Overview of changes:
To analyze data from a quad-base probe in spikeinterface, we have to use the latest releases of spikeinterface, neo, and probeinterface. Unfortunately, upgrading spikeinterface causes the LFP filtering to throw an error. The fix isn’t too arduous, but it does require some consideration and a decent amount of RAM.

As a brief overview, here’s how we currently pre-process our LFP data:

1. CatGT: Undoes the effects of multiplexing by temporally aligning channels to each other. We can also optionally split a multishank probe’s output into its respective shanks (not integrated yet into our pipeline).
2. Run Kilosort & Extract LFP pipeline (python, uses spikeinterface)
   1. Detect and remove bad channels
   2. Bandpass filter (1 Hz - 400 Hz)
   3. Resample (2500 Hz)
   4. Select channels to extract
      • Manually
      • Automatically (extract a channel every X um, usually every 100 um)
   5. Save the LFP times in npy format (.npy) - I think this is for later alignment across multiple streams
   6. Save LFP data in a .mat file in a way that’s consistent with mvdmlab data format.
   7. Key considerations for multishank probes (including quad-base): Bad channel detection, if using the default method, should happen independently for each shank.
3. TPrime: synchronizes data from multiple probes and the NIDAQ by aligning the sync pulse from each data stream to the sync pulse on imec0
   • Key consideration for quad-base: Each shank produces its own sync pulse. That said, I think they still are all part of the same data stream, so it is a little unclear to me if the sync pulses from the different shanks will need to be aligned with the sync pulse on a specified shank.
4. Sync LFP: Aligns LFP data from two different probes. Must be run even in one-probe recordings because there is a data structure reformatting that occurs.

Resources from spikeinterface:

1. Processing by shank: https://spikeinterface.readthedocs.io/en/stable/how_to/process_by_channel_group.html
2. Filtering LFP
   1. This pipeline has some basic code for it: https://spikeinterface.readthedocs.io/en/stable/how_to/drift_with_lfp.html
   2. https://spikeinterface.readthedocs.io/en/stable/forhowto/plot_extract_lfps.html
3. Detecting bad channels from LFP data
   1. The docstring for the function that does bad channel detection indicates that the neighborhood_r2 method is better for LFP use cases. We can discuss more if it makes sense to detect bad channels separately for spike-sorting data and LFP data. My tentative answer is no, but that’s because bad channels should theoretically be the same in both the LFP and spiking bands. That was not the case when I compared the neighborhood_r2 method with the coeherence+psd (default) method for my quad base test recording, but the quad base test recording wasn’t actually in the brain of a mouse, so it’s unclear if that’s representative of what would happen for a real recording.
   2. spikeinterface.preprocessing.detect_bad_channels: https://spikeinterface.readthedocs.io/en/stable/api.html#spikeinterface.preprocessing.detect_bad_channels

[CodyCBakerPhD]

Amazing stuff @emma-gibbens

Thank you for all the details; I will try to parse this all before the hackathon so I can get something setup ASAP