I did tetrode recording with 0.1 to 8000 Hz bandpass filter. Now I want to see whether I could get LFP with these data. What I did is to filter the data to 0.1-500Hz. Then I will downsample the data. My question is what I should do with the timestamps at the same time. How to make the timestamps consistent with downsampled neural activity data. Also, what about the timestamps of TTL signals? Thanks for your reply.
Hi Tianyang Ma,
In Matlab, you can downsample the data with the command 'interp1'. Here is an example (modified from the Matlab help for this command):
--------------------------------------------------------
X = 0:10; % original time vector
V = sin(X); % y data
Xq = linspace(X(1),X(end),100); % new time vector - this creates 100 linear data points from the first to the last index of the original time vector; you can change the number of data points in order to either up- or downsample your data.
Vq = interp1(X,V,Xq,'spline'); plot(X,V,'o',Xq,Vq); % interpolated data; there is different ways to interpolate data; here I used the 'spline' method but other methods exist (e.g. 'linear').
--------------------------------------------------------
Let me know in case anything is unclear. All the best,
Lukas
Hi Tianyang Ma,
why you want to downsample your data? You already low-pass filtered them. In doing so, you already discarded high frequency components, so in principle you can already analyse LFPs.
Best regards,
Erik
Hi Erik, because I at beginning recorded with tetrodes for spike analysis, so the sampling rates I set is 30 kHz, too high for LFP analysis. Do you think if I have artifacts
Dear Tianyang, your description already looks good.
With downsampling it is important to first low-pass filter the data, so that the upper band limit (here 500Hz) is well below the new Nyquist frequency. In your case you must not sample slower than 1000Hz, better keep a safety margin.
Filtering is a whole scientific field. In your case I'd recommend symmetric acausal filters. These do not lead to phase shifts and allow to keep the time stamps as they are. Causal filters are relevant when you want to do Granger causality measures or similar. But they might lead to phase shift close to the band stop frequency.Thus, unless Granger & Co. is your interest, I'd not bother with that.
Hope that helps, Peter
Hi,
I don't know your background in signal analysis, but I can speculating one possible source of misunderstanding: one thing is the sampling rate of your data, another one is the bandwidth. The Shannon-Nyquist sampling theorem requires that is you want to digitise an analog signal you need to provide a sampling frequency which is at least the double of the bandwidth (= max frequency contained in your signal) of your analog signal. Because, in principle, you don't know the bandwidth of your analog input signal the acquisition hardware system have (must have embedded on it) an analog low pass filter limiting it: this is the so called "ant-aliasing filter". After this filter, the signal can be digitised with a certain sampling rate. In practise, the Shannon-Nyquist requirement can be satisfied setting a sampling frequency which is not only 2 times the cut-off frequency of the antialiasing filter but at least 5-10 times more: this is due to the fact that it's impossible to have an anti-aliasing filter with infinite roll-off (too steep to be achieved).
Conversely, a sampling rate of 30kHz, does not mean you have an effective bandwidth of 30kHz: the real, aliasiaing-proof, bandwidth is always less than this. The design data of the antialiasing filter of the acquisition system should be asked to the manufacturer: it is very common to forget about this, but important detail. Even manufacturer sometimes don't have a prompt answer for this.
Saying all this, you always want to have your signal to be acquired at the highest possible sampling rate, because this best represents it with less artefact (e.g. aliasing). But once you have such acquired signal, you can reduce its bandwidth applying any type of filter you like without having to change the sampling frequency: e.g. you can put a 100Hz low pass filter and still have 30kHz sampling rate.
The cut off frequency of your low-pass filter is not connected with the sampling rate of your data: you only have to guarantee the Shannon-Nyquist condition.
Eventually, after low pass filtering your data, you can consider of reducing the sampling rate because you have too much data and you want to get rid of them (e.g. you don't want to store too much data).
How to properly filter your data can be very long to tell here in this post: one problem you want to avoid is the so called ringing. When you apply a filter, filter have a roll-off band: basically it does not attenuate completely the signal beyond the filter cut-off frequency. One way to increase the roll-off, in order to have a sharp frequency cut, is to use higher order filter: but their transfer function introduces modification in the signal which usually leads to false, artefacted, correlations in time (ringing). In order to avoid this you might use gaussians filter (because of its zero group delay) or a-causal filters (you filter with a kernel, then you time-reverse the kernel and you filter backwards in time: phase delays are this way compensated, regardless the transfer function of the filter you choose).
These are common methods, I hope I was able to shed some light on this - vey long - topic. There a lot of literature on this.
P.S. lot's of artefacts comes from the wrong choice of the filter!
Erik
P.P.S. My last answer was for Tianyang, Peter Koenig of course has a strong background in signal analysis :)
Lukas Wiedemann Hi, I have a question. Matlab downsample or resample or decimate which one is preferable for downsampling and why?
- Badges
- Science topic