We have read the study titled “Personalized Surgical Transfusion Risk Prediction Using Machine Learning to Guide Preoperative Type and Screen Orders” by Lou et al.1  and the accompanying editorial titled “Moving from ‘Surgeries’ to Patients: Progress and Pitfalls While Using Machine Learning to Personalize Transfusion Prediction” by Mathis et al2  The authors include 4 million surgical cases during a 3-yr period from the American College of Surgeons National Surgical Quality Improvement Program database. The authors used the American College of Surgeons National Surgical Quality Improvement Program database to develop a machine learning model that incorporates patient- and surgery-specific variables to predict transfusion risk and the associated need for preoperative type and screen. The authors hypothesize that their machine learning algorithm would outperform the traditional approach of relying primarily on historical surgery-specific transfusion rates and thus optimize resource allocation by decreasing blood bank waste. The machine learning algorithm recommends fewer preoperative type and screen orders.

The study presents in exceptional detail the methodologic approach to developing highly accurate algorithms to predict transfusion risk. Several authors have shown that race is an independent predictor of postoperative transfusion across surgical disciplines, associated with either higher or lower rates of transfusion.3,4  However, the work by Lou et al. did not include any mention of race or ethnicity. Experience from previous algorithms used to model resource allocation in health care demonstrate that omitting this information may lead to perpetuating bias that unfortunately exists within the United States healthcare system.5,6  Although the intent from healthcare providers is to provide the best possible care to their patients, determinants of health are closely linked to race and ethnicity and availability of resources in the United States. Therefore, artificial intelligence models aiming to personalize medicine can present pitfalls for those already with low resource availability, unwittingly withholding care in marginalized communities.5  On the other hand, it has been shown that including race or ethnicity in machine learning models may perpetuate bias, and therefore including race and ethnicity in artificial intelligence remains intensely debated.7,8 

Our primary question is the following: Why did the authors choose not to include race and ethnicity in their table 1 or in their prediction model? Was the absence of any demographic data in the Lou et al. article an intended or inadvertent omission? Given the potential impact of a patient’s race or ethnicity on clinician decision-making, and the ongoing controversy about the use of these variables in clinical prediction models,8  we believe that an explanation for the absence of this data would be helpful. Inclusion of ethnic and racial minorities in research is important, and transparency is key in the design of prediction models to improve societal health.