Abstract Title: Prototypical Machine Learning Model for Predicting Postoperative Blood Transfusion

Background Postoperative bleeding is a serious surgical complication that can result in re-exploration, transfusion, and mortality. Currently, there is no standardized, quantitative method to screen patients undergoing major surgery to identify patients who are at increased risk of postoperative bleeding. While it has been shown that certain independent risk factors such as age, smoking, medication use, and bleeding disorders play a large role in perioperative bleeding, no prior literature has explored the application of machine learning (ML) models to assess an individual’s risk quantitatively based on co-morbidities and medications. We aim to create an ML model to screen patients undergoing major procedures and assess their risk for significant postoperative bleeding requiring blood transfusion within 30 days.

Methods Using the MIMIC-IV dataset available through PhysioNet, we identified 199,277 records of major procedures as defined by the Agency for Healthcare Research and Quality (AHRQ). For each record, patient medical history, demographics, and procedure information were recorded, along with the outcome of blood product administration within 30 days of surgery. A comprehensive list of features can be found in Table 1 to the right. 80% of the records were used to train a Gradient Boosting Machine model, and 20% were used for validation. Based on the receiver operating characteristic (ROC), an optimal threshold of ~0.15 was calculated to differentiate between negative and positive predictions. Performance metrics were reported from the model’s predictions of validation set outcomes.

Results The final model predicted blood product administration with a sensitivity of 80.82% and specificity of 86.26%. The positive predictive value is 45.26%, and the negative predictive value is 96.97%. The area under the receiver operating characteristic (AUROC) was 0.91. The top five most predictive features, in order, were surgical procedure type, BMI, age, comorbid coagulative disease, and comorbid kidney or ureter disease.

Conclusion Our results suggest that an ML model can be utilized to screen patients undergoing major procedures to determine if they are at risk of postoperative bleeding requiring blood transfusion. The model’s predictions can be used to guide postoperative management and follow-up of high-risk patients. Further iterations will modify model inputs and parameters to improve performance and validate the model using another dataset from other institutions.