I'm working on a project predicting seizures events using 10 minute EEG data recordings. This involves identifying the preictal (pre-seizure) phase from the interictal (normal/between seizure) phase, so binary classification.

There are 16 electrodes, therefore 16 channels in my data. I have read that calculating the cross-channel power density spectrum is a popular method for this kind of problem, but this would result in a high number of dimensions so I want to know if it is okay to use dimensionality reduction on cross-channel derived features and if so, which technique would be best.

I have tried PCA, but I read that features with maximised variance is not a good method for classifying preictal and interictal recordings. LDA would have been nice to use as this is supervised learning, but I only have two classes so I can't use this.

I'm interested if ICA would work in this instance, or would it just end up identifying the singular channel power densities? Or possibly random forests?

  • $\begingroup$ Two followup questions to understand your problem better: What speaks against simply trying a PCA and seeing if it helps/works? Why are the high number of dimensions a problem? Is it due to constraints of the modeling hardware, the chosen model, etc.? $\endgroup$ – Fnguyen Jul 9 at 16:50
  • $\begingroup$ Hi @Fnguyen, I'll try out PCA. Yes, I'm using Hierarchical Temporal Memory which can only handle a reasonably small number of dimensions $\endgroup$ – Dee Jul 11 at 13:39

Your Answer

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

Browse other questions tagged or ask your own question.