Table 2 –

Model performance on the internal validation data.

Model	c-statistic	Average Precision	Sensitivity	Specificity	Positive Predictive Value	% Positive
Baseline	0.888 (0.881–0.894)	0.215 (0.197–0.235)	0.970 (0.963–0.977)	0.439 (0.436–0.442)	0.037 (0.035–0.038)	57.0% (56.7–57.3)
Logistic Regression	0.907 (0.900–0.913)	0.280 (0.260–0.300)	0.962 (0.954–0.970)	0.512 (0.509–0.515)	0.042 (0.040–0.043)	49.8% (49.5–50.1)
Decision Tree	0.916 (0.910–0.922)	0.298 (0.278–0.318)	0.963 (0.955–0.971)	0.618 (0.614–0.621)	0.053 (0.050–0.055)	39.5% (39.2–39.8)
Random Forest	0.913 (0.906–0.918)	0.247 (0.230–0.265)	0.962 (0.954–0.970)	0.593 (0.590–0.597)	0.050 (0.047–0.052)	41.9% (41.5–42.2)
Gradient Boosting Machine	0.924 (0.919–0.929)	0.292 (0.273–0.314)	0.963 (0.956–0.971)	0.651 (0.648–0.654)	0.058 (0.055–0.060)	36.2% (35.9–36.5)

Penalized logistic regression²⁴, decision tree²⁵, random forest²⁶, and gradient boosting machine²⁷ models were trained on the training data and evaluated on the internal validation data to predict transfusion on the day of surgery, using the procedure-specific transfusion rates observed in 2016–18. For comparison, a baseline model is also presented that used only the procedure-specific transfusion rate. c-statistic – area under the receiver operating characteristic curve. Average precision – area under the precision recall curve, indicative of model discrimination for the positive class. % Positive indicates percent of cases in the cohort for whom the model made a positive prediction, i.e., recommended a type and screen. All models were fixed with decision thresholds to achieve 96% sensitivity on the training data. 95% confidence intervals are shown in parentheses.