search

comojin1994 / m-shallowconvnet

Rethinking CNN Architecture for Enhancing Decoding Performance of Motor Imagery-based EEG Signals
31 stars 4 forks source link

BCI Competition IV 2b results #1

Closed sansiro77 closed 2 years ago

sansiro77 commented 2 years ago

I guess the accuracy for each subject is from training on the dataset of 3 T files and evaluating the dataset of 2 E files. How about cross-subject results, or specifically, leave-one-out-cross-validation results? And by the way, do you know the related SOTA results? Thank you.

comojin1994 commented 2 years ago

Hi @sansiro77,

Yes, I evaluated the performance of BCI competition 4 2b using two evaluation files which were the original test data training with three training files.

However, I didn't evaluate the subject independent tasks. Also, I didn't research the SOTA results in subject independent tasks.

sansiro77 commented 2 years ago

I also wonder why you choose tmin, tmax = 0., 3.. I get the dataset from MOABB, so the data format may be a little different. However, according to the original description file, the imagery period is from 4s to 7s without feedback or from 3.5s to 7.5s with feedback. Did I miss something?

comojin1994 commented 2 years ago

@sansiro77

The format of MOABB and original data is different.

In my case, tmin, tmax = 0., 3. indicates the 3 to 6 sec.. The reason why I chose an interval like this is that many previous studies commonly used the data containing cue time. I think the features of the motor imagery were also contained before the imagery period like MRCP. For this reason, I contained the cue time data.

The reason why set the tmax as 3 is that I optimized the interval, which showed the best performance.

However, in my custom data, I verified that this model also worked well when conducted only in the imagery period.

sansiro77 commented 2 years ago

Great, I see. Thank you very much!