Since the trend might be different for each of your trials, it seems to me that averaging all the five datasets before regression analysis would not be a good strategy.
Furthermore, I fail to see why you would like to average the five trials. Instead it probably makes more sense to concatenate all the five datasets, create a new ideal.1D file five times longer as the original one, and then run 3dDeconvolve with the option '-concat' by specifying the starting points for all the five trials. In doing so, each trial is detrended separately, and more importantly the signal (including random noise) is modelled with all the datasets as a whole.
Gang