Abstract
Introduction: Patients taking direct oral anticoagulants (DOACs), including direct anti-Xa inhibitors, have risk of major bleeding with substantial morbidity and mortality. Although Andexanet alfa is FDA-approved for reversal, its high cost and limited availability restrict its use in many centers. Off-label use of 4-factor prothrombin complex concentrate (4F-PCC) is an alternative. However, the data on safety and efficacy in Asian patients is scarce. Predicting outcomes in major DOAC-related bleeding is challenging due to patients' heterogeneity and complex interactions among comorbidities, treatments, and bleeding types. Machine learning can automatically capture non-linear relationships and high-order interactions of these parameters, offering improved risk stratification and outcome prediction compared with conventional regression analysis. This study aimed to (i) characterize features and outcomes of 4F-PCC–treated patients; (ii) identify predictors of clinical outcomes by supervised machine learning model.
Method: We retrospectively analyzed patients from multiple centers with emergency room services (9/2020–8/2023) aged ≥18 years on direct anti-Xa inhibitors who received 4-factor PCC for reversal in acute major bleeding. Clinical, laboratory, and treatment data were extracted from electronic medical records. The primary outcome was 30-day all-cause mortality. Secondary outcomes were in-hospital mortality and 30-day vascular mortality, defined as death related to the index bleeding episode and/or major adverse cardiac events (MACEs). The data of the patients' cohort was divided into training and testing set in 7:3 ratio. Data preprocessing included imputation, scaling, one-hot encoding, and normalization of DOAC dose by expressing individual doses as a proportion of standard therapeutic dose for each DOAC. Random forest models with randomized hyperparameter search and cross-validation were trained for all binary outcomes, and patients were stratified into high- vs low-risk groups based on predicted probabilities. SHAP values were used in evaluating model explainability, and the top three predictors for each outcome were validated using multivariate logistic regression. Survival between groups were analyzed by Kaplan–Meier and Cox regression analysis.
Results: A total of 168 patients taking factor Xa inhibitors who suffered acute major bleeding requiring 4F-PCC reversal were included. The median age was 80.5 years, and 64.3% were male. Overall, 74.4% presented with intracranial hemorrhage and the 30-day mortality rate was 32.7%. Random forest models showed good discrimination power for primary and secondary outcomes (Area-under-curve (AUC): 30-day all-cause mortality 0.99, in-hospital mortality 0.95, 30-day vascular mortality 0.99). The sole top predictor for 30-day all-cause mortality was GCS on presentation of intracranial hemorrhage. The top predictors for in-hospital mortality and 30-day vascular mortality were GCS on presentation and hemostatic efficacy. The model was able to classify “High-risk” and “Low-risk” patients with mortality of “High-risk” higher than that of “Low-risk” (30-day all-cause mortality: 92.7% vs 3.5%; in-hospital mortality:86.4% vs 8.8%; 30-day vascular mortality: 97.0% vs 2.2%). Logistic regression show independent associations for each outcome's top predictors: for 30-day all-cause mortality (GCS on presentation: OR 0.30 [95% CI 0.18–0.49]); for in-hospital mortality (GCS on presentation: OR 0.31 [95% CI 0.19–0.50]; poor hemostatic efficacy: OR 5.08 [95% CI 2.56–11.44]); and for 30-day vascular mortality (GCS on presentation: OR 0.27 [95% CI 0.16–0.48];good hemostatic efficacy: OR 0.17 [95% CI 0.06–0.52). Kaplan–Meier curves showed significantly lower survival in high-risk groups (log-rank p<0.001), with Cox regression hazard ratios HR 63.8 (95% CI 22.61–179.97) for 30-day all-cause mortality, HR 21.02 (95% CI 10.29–42.98) for in-hospital mortality and HR 161.48 (95% CI 46.30–563.20) for 30-day vascular mortality.Conclusion: 4F-PCC showed efficacy and safety in the reversal of direct anti-Xa inhibitor related major bleeding. Machine learning approach successfully identified predictors to classify patients with high and low risk for mortality outcomes. Further large-scale prospective studies are warranted to validate this risk stratification model and support therapeutic decision-making in patients presenting with major bleeding.
