Figure 4.

Predictive importance of individual clinical features on the likelihood of being prescribed an oral anticoagulation. The highest-performing machine learning model, XGBoost, was used to determine the rank order of features associated with oral anticoagulation prescriptions. This feature importance was measured using the permutation importance metric. With the fully trained model, independent variables were randomly shuffled, removing the relationships learned by the machine learning model, and the decrease in model performance was assessed. The average decrease in performance across 5 independent runs, and the standard deviation of those runs (black error bars) is shown for each variable. AUROC = area under the receiver operating characteristic curve; BP = blood pressure; GFR = glomerular filtration rate; INR = international normalized ratio; LVEF = left ventricular ejection fraction.