Bagging/ augmentation/ under-over sampling to deal with imbalanced classes
Check on v3.0_0047.csv, need rest of subject ID. Can wait because there's no way it could be a seizure positive subject 3_17_0007 and 3_17_0012 both lack duration information. EEG v0 is therefore just totals of events.
Collect probability predictions and run fusion on these soft labels (or pass classifiers not predictions)