search

gianba / SeizureDetection

3 stars 0 forks source link

Heuristic features #25

Closed gianba closed 10 years ago

gianba commented 10 years ago

two heuristic features added. currently they are commented out. maxAmplitudeChange is hopefully useful...

zadsas commented 10 years ago

Have you tried them out yet?

gianba commented 10 years ago

Well...:-) maxStepFeature didn't help and I think the early/late part doesn't make too much sense. That's why I commented it out. maxAmplitude is currently running, but on individual subjects I looked promising. I think it's okay to add pull features even if they are not perfect. We can always not use them or work further on feature selection. What we need is a better way to assess the quality. The current svm predictions are not stable in a sense, that we run a parameter optimization everytime, which is based on a 10-fold CV. This introduces some randomness, which lead to slightly different parameter settings and results everytime we run the whole thing. There are two things I'm currently looking at: Fixing the parameters on a pretty optimal set (which will result in reproducable results) and implementing a proper validation based on AUC on all subjects...