Development and Validation of a Machine Learning Model to Predict Oral Anticoagulant Use in Stroke From Prothrombin Time–International Normalized Ratio and Activated Partial Thromboplastin Time

Figures

• 

  Figure 1. Model performance for predicting anticoagulant use in the validation cohort. ROC curves (A and B) and precision‐recall curves (C and D) for predicting VKA (A and C) and DOAC (B and D) use vs others in the validation cohort (n=5910). For VKA prediction, the AUROC was 0.879 (95% CI, 0.847–0.912) for random forest, 0.931 (95% CI, 0.908–0.952) for XGBoost, and 0.933 (95% CI, 0.908–0.955) for neural network; the AUPRC was 0.41 (95% CI, 0.33–0.50) for random forest, 0.50 (95% CI, 0.41–0.58) for XGBoost, and 0.52 (95% CI, 0.44–0.61) for neural network. For DOAC prediction, the AUROC was 0.763 (95% CI, 0.742–0.784) for random forest, 0.827 (95% CI, 0.808–0.845) for XGBoost, and 0.841 (95% CI, 0.822–0.858) for neural network; the AUPRC was 0.28 (95% CI, 0.24–0.31) for random forest, 0.37 (95% CI, 0.33–0.41) for XGBoost, and 0.39 (95% CI, 0.35–0.43) for neural network. The dashed diagonal line in the ROC curves represents the reference line of no discrimination (AUROC=0.5). AUROC indicates area under the ROC curve; AUPRC, area under the precision‐recall curve; DOAC, direct oral anticoagulant; ROC, receiver operating characteristic; VKA, vitamin K antagonist; and XGBoost, extreme gradient…

• 

  Figure 2. Calibration plots for predicting anticoagulant use in the validation cohort. Calibration plots for predicting VKA (A) and DOAC (B) use vs others in the validation cohort. Predicted probabilities were grouped into deciles, and the mean predicted probability was plotted against the observed proportion of events in each group. The dashed diagonal line represents perfect calibration (calibration in the large=0, slope=1). For VKA prediction, calibration in the large and slope were −0.38 and 0.47 for random forest, −0.39 and 0.82 for XGBoost, and −0.33 and 0.95 for neural network. For DOAC prediction, calibration in the large and slope were 0.44 and 0.28 for random forest, 0.40 and 0.83 for XGBoost, and 0.36 and 1.05 for neural network. DOAC indicates direct oral anticoagulant; VKA, vitamin K antagonist; and XGBoost, extreme gradient boosting.

• 

  Figure 3. Decision curve analysis for predicting anticoagulant use in the validation cohort. Decision curve analysis for predicting VKA (A) and DOAC (B) use vs others in the validation cohort. The net benefit is plotted against threshold probabilities ranging from 0 to 0.5. The gray dashed line represents the “treat all” strategy, and the black solid line represents the “treat none” strategy. DOAC indicates direct oral anticoagulant; VKA, vitamin K antagonist; and XGBoost, extreme gradient boosting.

• 

  Figure 4. Anticoagulant prediction heat map based on PT‐INR and APTT values. Heat map displaying the most likely predicted anticoagulant class across combinations of PT‐INR and APTT values using the neural network classifier. Each tile represents a region in the PT‐INR×APTT space, with the color indicating the class with the highest predicted probability: red for VKA, yellow for DOAC, and gray for none. The transparency reflects the average maximum predicted probability in each tile, with darker colors indicating higher prediction confidence. Numerical labels inside each tile represent the average maximum probability for the predicted class. APTT indicates activated partial thromboplastin time; DOAC, direct oral anticoagulant; PT‐INR, prothrombin time–international normalized ratio; and VKA, vitamin K antagonist.