This is intuitively appealing, but do biometrics really boil down to an exact number that we could hash like this? (Genuine questions; I don't know.) It seems more likely to me that biometric measurements would be considered to "match" when they're within particular tolerances. This is an operation you can perform on the original measurements, but not on hashes of those.
There is a cryptographic tool called a "fuzzy extractor" that solves this problem (c.f. Fuzzy Extractors:
How to Generate Strong Keys from Biometrics
and Other Noisy Data, by Dodis, Reyzen and Smith [0]). At enrollment time, you compute some (non-sensitive) data P = Enroll(biometric). Then every time you compute Recover(P, biometric') you will get the same (high-entropy) output, as long as biometric' is close enough to biometric.
This is intuitively appealing, but do biometrics really boil down to an exact number that we could hash like this? (Genuine questions; I don't know.) It seems more likely to me that biometric measurements would be considered to "match" when they're within particular tolerances. This is an operation you can perform on the original measurements, but not on hashes of those.