Assessing Phenotype Definitions for Algorithmic Fairness