In this paper we present an extension of the recently proposed Stationary Subspace Analysis (SSA). This novel method solves the problem how to group signals from different conditions and/or subjects to find stationary subspaces. The original SSA approach does not offer a natural way to group data and therefore better define the non-stationarities of interest. This drawback is solved with group-wise SSA (gwSSA) and demonstrated with a simple but illustrative example: the classification of BCI data. If not treated correctly the BCI tasks are considered as non-stationarities in SSA, which complicates its use for classification purposes. We show how, by correctly defining groups, non-stationarities of interest can be extracted. In this paper, the application is in multi-class signals, where the groups are properly defined to even improve classification performance.