Predicting Risks of Intracranial Bleeding in Older Adults with Atrial Fibrillation to Aid Prescribing Decisions

Background: High risk of intracranial hemorrhage (ICH) due to falls is among the leading reasons for withholding anticoagulation in patients with Atrial fibrillation (AF) when it is indicated by clinical guidelines. However, such risk is often assessed arbitrarily and the most relevant claims-based algorithm for objective assessment (Homer fall risk score; Am J Med 2017) was not developed for ICH in the AF population.

Objectives: To develop an ICH prediction tool and compare its performance vs. Homer fall risk score in older adults taking oral anticoagulant (OAC) for predicting risk of ICH in patients with AF.

Methods: We used the US Medicare claims data to identify new users of OAC (warfarin, dabigatran, rivaroxaban, or apixaban) aged ≥ 65 years with a diagnosis for non-valvular AF in the preceding 365 days from 10/19/2010 to 12/31/2017. We used patients from the Midwestern US as the validation set and used patients from all other US regions as the training set for the model development. We used Lasso regression with 10-fold cross-validation to select among 93 predictors based on literature to predict 1-year risk of ICH. We compared the prediction performance with vs. without accounting for censoring due to death by inverse probability censoring weighting (IPCW) based on 95 covariates. We compared Homer score vs. new AF ICH risk score in terms of Area under ROC curve (AUC) and reclassification, which measures performance of different scores in classifying patients into multiple risk groups. We defined the risk category as low (less than the average risk), intermediate (1 to < 2 times average risk), or high (≥2 times average risk).

Results: Our study cohort included a total of 897,998 (mean [sd] age 76.1 [8.9] years and female 51.3 %) patients (672,312 in training and 225,686 in validation set). Based on 4,178 ICH events (0.6%) in the training set, the new ICH risk score included 50 variables and has an AUC of 0.656 in the training and 0.651 in the validation set. After IPCW to account for competing risks due to death, the AUC further improved to 0.682 and 673 in the training and validation set, which was superior to the corresponding Homer's AUC of 0.622 and 0.623, respectively (p>0.001). In the validation set, our AF ICH risk score reclassified 21.0, 44.2, and 58.0% of low, intermediate, and high-risk patients determined by Homer score with a net reclassification improvement of 7.28% (P < 0.001).

Conclusions: When comparing effects of OAC in patients with AF, our claims-based ICH risk score can be helpful in both clinical (to aid prescribing decisions) and research (to stratify patients for treatment effect heterogeneity assessment) contexts.