I think they take the 5'/3' difference and test that as a separate exon. In other words they split an exon into two smaller sub-exons and testing each of them separately. Do however note that this only works for quite large changes as you need reads (30-100nt) that maps uniquely to the regions you test against.

If you want to find smaller features a better approach is to predict these changes from transcript/isoform level data. You can use cufflinks/cuffdiff to quantify and predict isoform switches and afterwards use tools such as spliceR (which directly supports cufflinks/cuffdiff data) to analyze all 6 types of alternative splicing.