Filtering directly the sampled signal or the raw signal after multiplying it by its gain and adding its offset?

[florian6973]

Hi,

I am working on extracellular signal filtering and I noticed that in Spikeinterface, this is directly the sampled signal (for instance int values from 0 to 2**16 -1) which is filtered with the bandpass function. However, in some other signal processing codes, it is often the raw signal, ie the sampled signal converted to mV by multiplying it by the channel gain and adding the channel offset, which is computed.

Therefore, I was wondering what the reasons are to perform filtering directly with the sampled signal, and not the converted signal please. Is it known to get better results for spike sorting then?

Thanks for your help!

[samuelgarcia]

Hi Florian.
This is to keep the same data size. float32 is the double than float32. (float16 do not exists)
When you filter a recording with integer mode you can still get the scaled traces to microV because the gain is propagated

rec = ... # a recordingf internally int16

rec_f = bandpass_filter(rec)
trace_int16 = rec_f.get_traces(return_scaled=False)
trace_float32_uV = rec_f.get_traces(return_scaled=True)

[florian6973]

Thank you very much for your explanation! I understand this choice better now. Maybe related to this topic, to extract the LFP, why is a low-pass filter not implemented in spikeinterface, unlike bandpass and highpass ones? Admittedly, we can bandpass the signal between 1 and 300 Hz, but it would seem easier to have a low-pass filter. Thanks!

[florian6973]

After some investigation, is it linked with the fact that recordings can be chunked and processed in parallel? For bandpass and highpass filters, it is possible to merge the results after taking a small margin for each chunk, but it is not possible for low pass? If so, what is the minimum low frequency which it is relevant to specify to get the LFP with a bandpass? Thanks!

[florian6973]

For lowpass filters, I tested with chunks of fs/5 and it seems to provide consistent results, after fixing the offset issue, so I am a little bit puzzled.

[alejoe91]

Indeed! The reason is that a low-pass filter would not support chunking. Ideally, it would need chunks of the same length as the recording. You can use a lowpass cutoff of 0.1Hz ;)

[florian6973]

I understand better, thank you very much for your reply!