I'm not familiar with how the statistics with RMA works.
So my question is: Can I apply RMA (safely) on each disease group separately (rma on case, rma on control), and combine them afterword? Or would an algorithm that accounts for disease label differences (i.e. qsmooth) be a better option?
RMA is still by far the most accepted method of preprocessing Affymetrix data. Personally, I would use RMA and not give it a second thought. qsmooth might also be good but I am not familiar with it.
The papers you cite have not received a high level of acceptance. I have not read them in detail but I suspect that they are rediscovering the well-known fact that global normalization assumes that the majority of genes are not DE, and some other strategy may be required if you want to detect global expression changes in one direction.
Applying RMA on each disease group separately sounds very dangerous to me.