[FieldTrip] Kurtosis topography at channel level

[Frank Neugebauer]

Hi Rui,

*Question 1:* why and how does the filter affect the kurtosis?

As kurtosis is not linear, that is kurtosis(data+noise)=/=kurtosis(data)+kurtosis(noise), it is a little hard to say how exactly changing your data will change the output kurtosis.
If you filter Gaussian noise the signal should form stronger outlier and thus the data should yield higher kurtosis.
As far as I understand kurtosis, the frequency itself does not directly influence the kurtosis output.


*Question 2:* Why the kurtosis topographies at channel level from two
> dataset case1.ds (figure 1 (no filter), figure 2 (filtered)) and
> case1_cHPI_raw_trans_sss.fif (figure 3 (no filter) and figure 4 (filtered))
> are not consistent? Because these two datasets are measured from the same
> patient, the kurtosis topography should be consistent. Do you have any idea
> about this?


As far as I understand, epileptic activity is hard to measure because you don't know when it will occur and different measurements are not inconsistent, but just show different activity at a different time and circumstance (such as different sensor noise, interference, but also the patient's condition).
In the introduction to the case they say,
>> Corticography also showed interictal discharges in the frontal lobe, though seizures were of parietal origin. <<. Maybe this is the explanation? The kurtosis points to the frontal lobe, too.

I hope this was a little helpful,
best regards,
Frank


Rui Li wrote on 2016-07-21:
> Dear FieldTrip users,

> Recently, I am working on the case 1 dataset of epilepsy tutorial. The
> first patient got the MEG recording from Neuromag and CTF, respectively. I
> encountered some problems when I tried to depict the Kurtosis topography
> at channel - level.


> *Question 1:* why and how does the filter affect the kurtosis?

> Figure 1 and Figure 2 are kurtosis topographies at channel level without
> band pass filter and with band pass filter, respectively. As we can see,
> these two figures are very different. Therefore, I am wondering why and how
> does the frequency filter affect the kurtosis?

> The figure 1 is generated by the following program;

> %% preporcessing the channel level data

> dataset = 'case1.ds';

> cfg = [];

> cfg.dataset = dataset;

> % cfg.hpfilter = 'yes';

> % cfg.hpfreq = 10;

> % cfg.lpfilter = 'yes';

> % cfg.lpfreq = 70;

> cfg.channel = {'MEG'};

> data = ft_preprocessing(cfg);



> %% compute channel-level kurtosis



> datak = [];

> datak.label = data.label;

> datak.dimord = 'chan';

> datak.kurtosis = kurtosis(data.trial{1}')';



> cfg = [];

> cfg.comment = 'computed channel-level kurtosis';

> datak = ft_annotate(cfg, datak);



> %% plot kurtosis topography at channel-level



> cfg = [];

> cfg.layout = 'CTF275.lay';

> cfg.parameter = 'kurtosis';



> figure;

> ft_topoplotER(cfg, datak);


> ​

> Figure 1 kurtosis topography + no filter + case1.ds

> If the band pass filter is included in the pre-processing, the kurtosis
> topography is figure 2; the pre-processing matlab program is

> %% preporcessing the channel level data

> dataset = 'case1.ds';

> cfg = [];

> cfg.dataset = dataset;

> cfg.hpfilter = 'yes';

> cfg.hpfreq = 10;

> cfg.lpfilter = 'yes';

> cfg.lpfreq = 70;

> cfg.channel = {'MEG'};

> data = ft_preprocessing(cfg);



> Figure 2 kurtosis topography + [10Hz 70Hz] filter + case1.ds

> *Question 2:* Why the kurtosis topographies at channel level from two
> dataset case1.ds (figure 1 (no filter), figure 2 (filtered)) and
> case1_cHPI_raw_trans_sss.fif (figure 3 (no filter) and figure 4 (filtered))
> are not consistent? Because these two datasets are measured from the same
> patient, the kurtosis topography should be consistent. Do you have any idea
> about this?


> ​

> Figure 3 kurtosis topography + no filter + case1_cHPI_raw_trans_sss.fif


> ​

> Figure 4 kurtosis topography + [10Hz 70Hz] filter +
> case1_cHPI_raw_trans_sss.fif

> Regards,

> Rui.