That is a very non-trivial question -- essentially, it is the 'model order selection' problem. There has been a lot of work on this over the last 30+ years in statistics, with no universally accepted conclusion.
You might restrict yourself to (say) a 3 second resolution in the hemodynamic response, so if your total response duration is expected to be 15 s, use 5 sin functions. But this is just a guess.