Performance of the HEMORR₂HAGES, ATRIA, and HAS-BLED Bleeding Risk–Prediction Scores in Patients With Atrial Fibrillation Undergoing Anticoagulation

Objectives

The objective of this study was to compare the predictive performance of bleeding risk–estimation tools in a cohort of patients with atrial fibrillation (AF) undergoing anticoagulation.

Background

Three bleeding risk–prediction schemes have been derived for and validated in patients with AF: HEMORR₂HAGES (Hepatic or Renal Disease, Ethanol Abuse, Malignancy, Older Age, Reduced Platelet Count or Function, Re-Bleeding, Hypertension, Anemia, Genetic Factors, Excessive Fall Risk and Stroke), ATRIA (Anticoagulation and Risk Factors in Atrial Fibrillation), and HAS-BLED (Hypertension, Abnormal Renal/Liver Function, Stroke, Bleeding History or Predisposition, Labile International Normalized Ratio, Elderly, Drugs/Alcohol). Τhe relative predictive values of these bleeding scores have not previously been compared.

Methods

We analyzed the dataset from the AMADEUS (Evaluating the Use of SR34006 Compared to Warfarin or Acenocoumarol in Patients With Atrial Fibrillation) trial, a multicenter, randomized, open-label noninferiority study that compared fixed-dose idraparinux with adjustable-dose oral vitamin K antagonist therapy in patients with AF. The principal safety outcome was any clinically relevant bleeding event, which was a composite of major bleeding plus clinically relevant nonmajor bleeding.

Results

The HAS-BLED score performed best in predicting any clinically relevant bleeding, reflected both in net reclassification improvement (10.3% and 13% improvement compared with HEMORR₂HAGES and ATRIA, respectively) and receiver-operating characteristic (ROC) analyses (c-indexes: 0.60 vs. 0.55 and 0.50 for HAS-BLED vs. HEMORR₂AGES and ATRIA, respectively). Using decision-curve analysis, the HAS-BLED score demonstrated superior performance compared with ATRIA and HEMORR₂HAGES at any threshold probability for clinically relevant bleeding. HAS-BLED was the only score that demonstrated a significant predictive performance for intracranial hemorrhage (c-index: 0.75; p = 0.03). An ATRIA score >3 was not significantly associated with the risk for any clinically relevant bleeding on Cox regression or on ROC analysis (c-index: 0.50; p = 0.87).

Conclusions

All 3 tested bleeding risk–prediction scores demonstrated only modest performance in predicting any clinically relevant bleeding, although the HAS-BLED score performed better than the HEMORR₂HAGES and ATRIA scores, as reflected by ROC analysis, reclassification analysis, and decision-curve analysis. Only HAS-BLED demonstrated a significant predictive performance for intracranial hemorrhage. Given its simplicity, the HAS-BLED score may be an attractive method for the estimation of oral anticoagulant–related bleeding risk for use in clinical practice, supporting recommendations in international guidelines.