Can the built in matching pursuit function wmpalg be used to analyze signals that are much longer than dictionary elements?

I would like to perform a matching pursuit (MP) analysis of continuous neural data. Essentially the data is a sequence of oscillatory bursts with different frequencies, and I would like to use MP to automatically classify these bursts. I have chunks of data that are approximately 30min long (samp freq 2kHz). Each chunk contains well over 1 million data points.

The help section explains the following about the dictionary elements (mpdict) for wmpalg: "MPDICT is a N-by-P matrix where N is equal to the length of the input signal, Y"

If the input signal and the dictionary elements must be the same length then it looks like wmpalg can only be run on short segments of data because the dictionary elements are usually only about 2000 elements long. I don’t want to arbitrarily divide the data into small segments to run the analysis because I don’t want to risk cutting an oscillatory burst in half.

I am currently detecting the oscillatory bursts with a power threshold and so I could potentially pull out segments of data that are centered on these oscillatory bursts. Is this the best way to do this, or is there a better approach to using wmpalg with long streams of data?

Answer 1

[Matthew Aidekman]

bump? Any answers here? What's the deal?