Only way I see to do this is to either read line-by-line and parse each line to see whether to keep or not or on same idea to read in blocks of (say) 10000 lines or so and do the selection block-by-block.
Not knowing anything of the file format itself, it's possible one could preprocess it before reading with Matlab using a batch editor or regular expressions or the like. Of course, depending on just how big the file is (having 1M lines isn't necessarily out of hand depending on the length of each record), you might be able to read as character/cellstr array and do the cleanup in memory as character before conversion, too....too many possibilities depending on unknown details to say unequivocally what would be best approach. But, the first two will work, just time-consuming.
