Feature extraction from ERP instead of raw data for classification

Is it a right approach to extract time domain, feature domain including entropies from ERP instead of raw daw and feed them to a machine learning classifier.
What are the advantages and disadvantages of that approach?.
If anyone can provide me a paper, it would be of great help

I have only a little experience with decoding based on ERPs, but I would be concerned about the meaningfulness of entropy from an ERP. The data variance that generates entropy will be non-phase-locked over trials, and therefore my intuition is that most of the entropy that exists on single trials is not present in the ERP.

For relevant papers, you can try searching pubmed.gov or scholar.google.com