Calssification for many subjects (patients)

✉️ Mailing List Archive (read-only)
1

External Email - Use Caution

Hi all,
I'm new to the EEG World. I have EEG-measurement from about 10 patients. I
did CSP (Common Spatial Patterns) classification for each subjects and I
got an accuracy 70% for some of them and got 60% for the rest. Then I did
classification for all the subjects together (to have more data) and got
90% accuracy.
I was always thinking that the EEG is different from subject to subject and
it's like fingerprints. Brain Regions are same but Connectivity differs.
Electrodes localization are same but not exactly on the same positions for
all subjects.
Can I say that this 90% accuracy is the accuracy for every patient on the
set I have and in the future? Or should I say no, every patient has it's
own result?

Also what about transfer learning? I heard that this is a way to learn a
model on one patient and then take it and re-learn it again on another
patient and that in this case we get better results? Is there transfer
learning package in MNE-Python?

Regards,
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.nmr.mgh.harvard.edu/pipermail/mne_analysis/attachments/20190223/3b4bd207/attachment-0001.html

2

External Email - Use Caution

Hi Julian,

I?ve recently written a paper that addresses transfer learning between subjects. The idea is that a (linear) classifier aims to extract the signal of interest from the rest of the EEG signal. Hence, the model is influenced by both. With transfer learning, the assumption is that the signal of interest is similar between subjects (otherwise it would be impossible). However, we can relax the assumption that the rest of the EEG signal (noise and other brain activity) is similar across subjects, and just transplant the signal of interest only. It?s written up here (includes a link to a Python package that implements the technique):

Post-hoc modification of linear models: combining machine learning with domain information to make solid inferences from noisy data
https://www.biorxiv.org/content/10.1101/518662v1

And I use the technique in this paper as well:

Exploring the Organization of Semantic Memory through Unsupervised Analysis of Event-related Potentials
https://research.aalto.fi/files/27884560/jocn_a_01211.pdf

Maybe that helps.

Best,
Marijn.