Thanks.

*"Using the preintervention cohort, a multivariate logistic regression model was constructed to generate variables that predicted the need for intraoperative or postoperative blood product transfusions. This predictive model was then applied to the postintervention cohort to identify a subgroup of patients who were predicted to receive transfusion, but did not. This subgroup is defined as a “misclassified” population that was predicted to receive transfusion, did not actually receive transfusions, and did not have any difference in clinical outcomes. The area under the receiver-operating characteristics curve was used to assess the accuracy of the models.*

Using the postintervention cohort, a decision tree was built to forecast the likelihood of perioperative blood transfusion. Specifically, three separate multivariate logistic regression models—one for demographics (D), one for preoperative comorbidities (C), and one for operative factors (O)—were built and used to generate scores D, C, and O for each patient. Then the three scores were included in the final combined model. Three models were fitted to allow the use of this score at different time point in the course of patient care. For example, operative information is not available in the preoperative setting, but using the preoperative category of the score would provide the likelihood of being transfused. To simplify the computation of the scoring system, the regression coefficients were uniformly rescaled to make the maximum total score 200. The area under the receiver-operating characteristics curves of the models generated by the simplified scoring systems were identical to those derived from the original regression coefficients. Calibration was tested using the Hosmer-Lemeshow goodness-of-fit test. A classification and regression tree was used to define cutoffs for each component using the new scores. Each node split decision in the tree was chosen from the possible cutoffs for all components according to Gini’s coefficient impurity measure. The node and depth of tree were set manually."

Using the postintervention cohort, a decision tree was built to forecast the likelihood of perioperative blood transfusion. Specifically, three separate multivariate logistic regression models—one for demographics (D), one for preoperative comorbidities (C), and one for operative factors (O)—were built and used to generate scores D, C, and O for each patient. Then the three scores were included in the final combined model. Three models were fitted to allow the use of this score at different time point in the course of patient care. For example, operative information is not available in the preoperative setting, but using the preoperative category of the score would provide the likelihood of being transfused. To simplify the computation of the scoring system, the regression coefficients were uniformly rescaled to make the maximum total score 200. The area under the receiver-operating characteristics curves of the models generated by the simplified scoring systems were identical to those derived from the original regression coefficients. Calibration was tested using the Hosmer-Lemeshow goodness-of-fit test. A classification and regression tree was used to define cutoffs for each component using the new scores. Each node split decision in the tree was chosen from the possible cutoffs for all components according to Gini’s coefficient impurity measure. The node and depth of tree were set manually."