Is there some way, by observing the data, to automatically make educated guesses at what the appropriate metafeatures are? This is a difficult question, and it is the temporal classification equivalent of constructive induction in inductive logic programming - something of a holy grail. It may be possible, however it looks extremely difficult.
However, there are alternatives. For instance, it may be possible to have a good-sized library of metafeatures, and, given a new domain, somehow work out which are the appropriate metafeatures for it.