Obviously, you assumed that I was just making “my usual suggestion” when I am not.   Your algorithm is quite plainly at least the beginnings of the only feasible solution in this case ... especially given that the actual number of duplicates is known to be negligible.   (0.0022% ...)   If the data volumes are in the 2-billion range and you are certain that you could only be committing a Type-2 error with no possibility of Type-1, then you have solved the problem right now.   You produced a highly sensitive first-pass filter.   99.98% of the records have been excluded.   That’s as solved as you can get ... and what you already know about the expected distribution (that is it is overwhelmingly likely that the records are unique) makes it so.

In effect your algorithm is to calculate a digital-digest of the entire record and then to look for possible collisions among those digests.   The discriminating ability of the algorithm will be that of MD5 and nothing more or less.   Taking the MD5 result and using the left/right-most n bits would be statistically stronger than applying a separate reducing calculation to the entire result.   MD5 and its brethren scatter entropy fairly-uniformly across all of the bits at once as they are designed to do.   The distribution of values ought to be very flat, and any subset of the total bits in the vector, if not combined with others, will be equally flattened and therefore equally conducive to your hash.