0
$\begingroup$

I am using MATLAB to evaluate power spectral density estimates of half second EEG signals, using modified covariance method. Can anyone suggest me how to select the AR model order for this process? Is there any function in MATLAB which can be used for selecting the best order using any model order selection techniques?

Thanks a lot for your help.

$\endgroup$
1
$\begingroup$

Do you want to select the model order adaptively based on the content of the input EEG, or choose the AR order as a design parameter for a system that will always be the same?. In the second case (not adaptive), maybe if you have some sample EEGs you could try a a lot of models from order = 0, to some number N (large enough but not so that running all the models takes forever). And Graph the Error between the EEG signal and white noise filtered by the model. You should arribe at a graph or Error vs order. At some value of N the error only marginally decreases. That is your model order.


EDIT

Matlab offers AIC and others.

$\endgroup$
  • $\begingroup$ Thanks for your reply bone. Yes I am looking for a non adaptive system with a fixed model order. I have heard about some model selection techniques like AIC, BIC, FPE. Does MATLAB provides any function to implement any of these techniques to select best model order? I have manually tried various model orders 5-13 to plot power spectral density plots and for orders 9-11, I am getting pretty smooth plots with 2-3 peaks. But I want to try some model order selection technique to see what order does it suggest. $\endgroup$ – Hemang Shrivastava Oct 29 '15 at 16:21
  • $\begingroup$ I don't know If matlab offers this, maybe. But try the process i described manually, and select the model order "by eye". Lets say you have an EEG, first compute the "true" power spectral density (FFT squared). Then compute 10 different power spectrums, generated from modeling with orders 1 to 10, and taking some sample white noise sequence and filtering with the 10 FIR filters. Then compute the 10 mean squared errors between the true spectrum and the estimated based on the model. Plot them, and see what happens (Maybe post the results here as well just to see what you were working with) $\endgroup$ – bone Oct 29 '15 at 17:41

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.