We have read with great interest the recent article by Lou et al.,  in which they used the American College of Surgeons National Surgical Quality Improvement Program participant use data file to expertly develop a transfusion prediction model with the goal of guiding type and screen ordering.

Lou et al. devised a clever method to broadly capture institution-specific transfusion information by redefining the procedure-specific transfusion risk on local institutional data in their external validation experiments. The choice to use this process to improve model performance speaks to the importance of institution-specific data and to the assumption that inclusion of granular institutional data results in superior prediction. One example of an institution-specific variable that may confer additional predictive power is surgeon identifier, as there is evidence of intersurgeon variability in transfusion requirements.  Also, it is unclear if anesthesiologist identifier is predictive of transfusion, which should be explored in greater detail. Widely externally valid approaches to modeling perioperative problems sacrifice data granularity that may be critical for practical implementation.