[Eeglablist] Filter causality pop_eegfiltnew

Vito de Feo vito.defeo at zmnh.uni-hamburg.de
Tue Jan 28 08:05:28 PST 2014


Dear Tim,

thank you very much for all your emails and advises. I answer in detail to
your following email:

Quoting Tim Mullen <mullen.tim at gmail.com>:

> Dear Vito,       
>    Correct, SIFT can work on both continuous data or epoched multi-trial
> data and uses the third dimension (epochs) to determine which is the
> case. If you wish to operate on epoched single trials, you may
> concatenate the trials into a 2D matrix (chs x pnts*trials) In this
> case, bear in mind that if you're using a sliding window approach, you
> will need to discard the results within the period corresponding to 1/2
> of the window length at the beginning and end of each trial, since the
> sliding window will have overlapped two trials here. 
>
>
>  

I have a big problem with Granger Analysis (also using GCCA or BSMART)
because my data are not really easy to analyze. I record local field
potential from several brain regions of neonatal rats. I am analyzing the
casual flow from the Hippocampus (HIP) to the Prefrontal Cortex (PFC).
This flow should be unidirectional because the anatomic connections are
just in one direction. So my goal is to proof, using information flow
(Granger, Transder Entropy, etc.) that I have a direct connection (just
monosinaptic) from HIP to PFC. Obviously I have also a path from PFC to HIP
but this path is not direct and the information flow takes more time. First
of all I don't know if I can proof this kind of connectivity using Granger,
I would expect that the flow from HIP to PFC (where I have direct
connections) is higher than the flow from PFC to HIP (where I don't have
direct connections).

The signals are not easy to analyze because:

1) In this case we just measure the "spontaneous" activity that means
that we do not stimulate the animal and he is sleeping during the
recording (we use urethane anesthesia). So I don't have "naturally
epoched" data and I window the data (sliding window) also to improve the
stationarity. I have a window of 500ms and I use two approaches:
a) I consider each window as my entire trace and I calculate a MVAR model
for each window
b) I divide each window in 10 subwindows of 50ms and I consider each
subwindow as an epoche (as a realizzation of the same random process)
This is why asked about 
I found impossible to have one model for the entire recording (my recording
lasts 40 min). The best compromise beetween consistence and stability
(using Anil Seth toolbox) is to consider a maximum lenght of the trace of
500ms. With SIFT I consider the entire recording and this is probably why I
don't have stability and consistence. As I have understood fine SIFT create
one model for the entire trace (of 45 min) as is normal to do (but not in
my case I think). 

2) In the neonatal rats the activity is highly discontinuous and I
have bursts of activity interleaved by noise (the noise level is my
baseline and I have 5-6 bursts per minute of 0.5-8 sec each). In this way
the signal is very far to be stationary, also because the activity (the
bursts of activity) is very not stationary.

In this situation it is very difficult to have stationarity,
linearity, good model consistence, white residuals, etc... In addition
also when the criteria are satisfied (in few parts of the signal trace) I
cannot find high flow density. Also Steven Bressler that was here in
December and gave a look to my data told me that the problem is difficult.
Probably a linear Granger approach is not correct for this data.

Anyway I developed this method that gives some results. The method is based
on calculating a different model for each window during the sliding and on
the time shifting of the time series, as I wrote before.

You wrote: "For single-trial analysis, unless the number of variables
(channels/sources) you are modeling is low, you will likely need to use
regularized VAR estimators (ridge regression or sparse VAR) or a Kalman
filter approach (est_fitMVARKalman) to achieve a reasonable model fit."

I have just 2 channels (HIP and PFC). I can also have a third region that I
guess could be an intermidiate region for the flow from PFC to HIP (the
flow from HIP to PFC should be direct as I worte before).
Sorry, Tim, I don't know these methods: "ridge regression or sparse VAR) or
a Kalman filter approach (est_fitMVARKalman)". Where can I find some
explanation about these methods?
I just tried to differenciate but I don't have better results.

You wrote:
It appears to me that your data still has some stationarity or other issues
leading to model instability. I recommend trying the following:

1) clean your data to remove trials, channels/sources, or data subspaces
with large artifacts -- especially blinks and muscle artifacts, which are
not well-modeled by a stationarity VAR process
2) downsample the data to an appropriate sampling rate (256 Hz is typical
for EEG)
3) apply local detrending in SIFT pre-processing, if drift is present
4) decrease the window length to (e.g.) 0.5 sec to produce a more locally
stationarity signal.
5) increase the model order if needed (you might check the model order
selection procedure to help determine this)

I don't have artifacts (this is one of the few good things!) but the signal
often seem to contain artifacts because the local field potential activity
has often very big peaks (this a big difference beetween local field
potential and EEG or fMRI).
If I downsample to 800 Hz, if I downsample more the consistence is very low
and also the stationarity is worse.
I apply everytime detrending and the model order changes window by window
but the average is around 10.

If you have any suggestion I would appreciate a lot!

Vito

>  
>    Tim
>
>
>     On Sun, Jan 19, 2014 at 12:08 PM, Vito de Feo
> <vito.defeo at zmnh.uni-hamburg.de> wrote:
>
>> _Dear all,
>>
>> sorry if I write again. I am very interested to know better SIFT!
>>
>> I have understood that the stability test includes also the
>> stationarity test. So sorry for my prevoius question.
>> Now I am trying to understand if SIFT uses a multitrial approch or a
>> single trial approach. I guess that if the signal is not epoched it use
>> a single trial approach.
>>
>> Now I have problems with the consistency (I attach two pictures). The
>> consistency is very low as you can see. What can I do? Should I
>> increase the model order or should I decrease the moving windows
>> length? (now it is 1.5 s, I could decrese to 0.5 s beacuse the signal
>> is strongly not stationary).
>>
>> Thank you!_
>>
>>
>> _Vito 
>>
>> Quoting Andreas Widmann <widmann at uni-leipzig.de>:_
>>
>>         
>>
>>> _Dear all,_
>>>           _ _
>>>           _not directly related to your question and SIFT, but eegfilt
>>> is deprecated and I would recommend not using it any longer._
>>>           _ _
>>>           _Best,_
>>>           _Andreas_
>>>
>>> _Am 18.01.2014 um 15:47 schrieb "jfochoaster ."
<jfochoaster at gmail.com>:
>>>  _
>>>
>>>> _Hello all,
>>>>  _
>>>> _I'm following the SIFT tutorial, the section 6.5.1.3 is about
>>>> filtering, talk about eegfilt, about the zero-phase (acausal) filter
>>>>  _
>>>> _Is better forget this section of filtering and use the
>>>> recommendations in the past emails?
>>>>  _
>>>>               _Are these recommendation critical for the analysis?, I
>>>> mean, there is a lot of work about MVAR models in ECoG data_
>>>>               _ _
>>>> _Best wishes
>>>>  _
>>>> _John_
>>>>
>>>>
>>>>              _On Fri, Jan 17, 2014 at 11:05 PM, mullen.tim at gmail.com
>>>> <mullen.tim at gmail.com> wrote:_
>>>>
>>>>> _Oh thats interesting. I had not seen Anil's multitaper filter
>>>>> (might be fairly recent). But possibly it is exactly the same
>>>>> approach that is in Cleanline. If this is the method advocated by
>>>>> Mitra and Pesaran as in the Chronux toolbox then indeed its the
>>>>> same. And highly recommended._
>>>>>                _-----Original Message-----
>>>>> Date: Friday, January 17, 2014 1:21:30 pm
>>>>> To: mullen.tim at gmail.com
>>>>> Cc: trotta_gabriele at yahoo.com, drcoben at gmail.com,
>>>>> mmiyakoshi at ucsd.edu, widmann at uni-leipzig.de, eeglablist at sccn.ucsd.edu
>>>>> From: "Vito De Feo" <vito.defeo at zmnh.uni-hamburg.de>
>>>>> Subject: Re: [Eeglablist] Filter causality pop_eegfiltnew
>>>>>  _
>>>>> _Before using the Cleanline (that I used today for the first time) I
>>>>> did't use the notch filter, I used a multi taper filtering made by
>>>>> Anil Seth. I know that filtering is very bad for later VAR modeling,
>>>>> especially notch and high pass. Low pass is better (usually I use
>>>>> multi taper filtering to remove the noise lines and a low pass
>>>>> causal filter with cut off filtering of 100 Hz).
>>>>> Do you think is ok Tim?
>>>>> Best
>>>>> Vito
>>>>>
>>>>> Il giorno 17/gen/2014, alle ore 20:53, mullen.tim at gmail.com ha
>>>>> scritto:_
>>>>>
>>>>> _> Do not notch filter your data! This can be very bad for later VAR
>>>>> modeling -- and IMO bad in general. You can use an adaptive spectral
>>>>> regression method such as that in the Cleanline plugin for eeglab to
>>>>> remove line noise.
>>>>>>
>>>>>> See Barnett and Seth 2011 and Mitra and Pesaran 1999 for
>>>>>> theoretical discussions.
>>>>>>
>>>>>> Rob, there is no video of the SIFT workshop but the lecture pdfs
>>>>>> are online at the eeglab workshop page.
>>>>>>
>>>>>> Tim
>>>>>> -----Original Message-----
>>>>>> Date: Friday, January 17, 2014 10:18:32 am
>>>>>> To: "
>>>>>  _
>>>>>
>>>>> ________________________________________________
>>>>> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
>>>>> To unsubscribe, send an empty email to
>>>>> eeglablist-unsubscribe at sccn.ucsd.edu
>>>>> For digest mode, send an email with the subject "set digest mime" to
>>>>> eeglablist-request at sccn.ucsd.edu_
>>>>
>>>> _--
>>>> John Ochoa
>>>> Docente de Bioingeniería
>>>> Universidad de Antioquia_
>>
>>  
>>
>>        _--
>> Pflichtangaben gemäß Gesetz über elektronische Handelsregister und
>> Genossenschaftsregister sowie das Unternehmensregister (EHUG):
>>
>> Universitätsklinikum Hamburg-Eppendorf
>> Körperschaft des öffentlichen Rechts
>> Gerichtsstand: Hamburg
>>
>> Vorstandsmitglieder:
>> Prof. Dr. Christian Gerloff (Vertreter des Vorsitzenden)
>> Prof. Dr. Dr. Uwe Koch-Gromus
>> Joachim Prölß
>> Rainer Schoppik_
>
>      
> _--
> ---------  αντίληψη ----------- _



-- 
Pflichtangaben gemäß Gesetz über elektronische Handelsregister und Genossenschaftsregister sowie das Unternehmensregister (EHUG):

Universitätsklinikum Hamburg-Eppendorf
Körperschaft des öffentlichen Rechts
Gerichtsstand: Hamburg

Vorstandsmitglieder:
Prof. Dr. Christian Gerloff (Vertreter des Vorsitzenden)
Prof. Dr. Dr. Uwe Koch-Gromus
Joachim Prölß
Rainer Schoppik
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20140128/4d79797a/attachment-0001.html 


More information about the eeglablist mailing list