Patient trained ICA-Fix

Dear all,

I am currently working on a project involving fMRI data from both healthy subjects and patients with epilepsy. I plan to clean the data using ICA-FIX, and in this case, I think there are three possible approaches:

  1. Train the classifier on data from healthy subjects only and apply it to all data.
  2. Train the classifier on all subjects (combined healthy and patient data) and apply it to all data.
  3. Train the classifier separately for each category (healthy/patient) and apply it within the respective category.

After some consideration, I believe option 3 might be the best, as it might better account for potential disease-specific alterations in both noise and signal. However, I am concerned about introducing bias by performing disease specific noise classification (e.g., amplifying noise components in the patient data that might not be detected but persist in the healthy dataset).

I haven’t found a clear consensus in the literature on this matter. Would you have any suggestions for how to proceed or have any relevant literature to recommend?

Best regards,

Thomas

Hi @Thomas1, I’m not an expert in these matters, but I suspect that option 2 is probably better - you want the model to capture features that are intrinsic to the data acquisition, rather than to a specific group; otherwise, as you suspected, you will run the risk of introducing a bias by cleaning the data for one of your groups better than the other.

Also, if you are not already doing so, can I recommend that you use the new Python-based FIX that is part of recent FSL, rather than the old MATLAB/R version, as the latter is an absolute nightmare to support (and the new version is faster and has better classification accuracy ) :slight_smile: