Author + information
- Margaret C. Fang, MD, MPH⁎ (, )
- Alan S. Go, MD and
- Daniel E. Singer, MD
- ↵⁎Department of Medicine, University of California, San Francisco, 503 Parnassus Avenue, Box 0131, San Francisco, California 94143
We did not test the HAS-BLED score in our report (1) because there was no internal validation in its original report (2). Risk prediction models are prone to overconfidence; for >25 years, the reporting standard has been internal validation through split-sample testing, bootstrapping, or other techniques (3). Further, the Euro Heart Survey, on which HAS-BLED was based, is problematic for follow-up studies of bleeds due to a relatively small number of outcome events, missing follow-up in >30% of patients, and incomplete information on longitudinal warfarin use (2). Incorporation of HAS-BLED into guidelines is not a substitute for good methodologic practice.
Dr. Olesen and colleagues touch on important issues concerning development of bleeding risk models (i.e., whether to include nonanticoagulated patients, adjust for international normalized ratio (INR) control, or restrict to new users). We continued the tradition of most previous bleed models in focusing solely on warfarin-treated patients. This mirrors the approach of stroke risk schemes (e.g., CHADS2) that predict risk solely in patients off anticoagulants and assume a fixed risk ratio in the alternative therapy category. We did not include INR values because we wanted the model to address the anticoagulation decision before starting therapy and to allow our score to apply to novel anticoagulants where INR testing is irrelevant. In ATRIA, the absolute increased risk of bleeding among new users of warfarin was small, and we included both new and long-term warfarin users to maximize power.
Dr. Olesen and colleagues are incorrect in asserting that we excluded patients with contraindications to warfarin therapy. Physicians are, of course, less likely to prescribe warfarin to patients at higher perceived bleeding risk, but this bias is present in all clinical studies of warfarin and should not necessarily affect our model's accuracy. The counterintuitive “protective effects” of certain univariate findings were based on very small numbers of events; these variables dropped out in the multivariable phase. We used univariate hazard ratios >1.5 as an objective quantitative basis to select variables for the multivariable phase of model building. This approach reduces false-positive findings. The variable age ≥75 years performed better than age ≥65 years in our model building. Additionally, we tested previous stroke, and it did not pass the bootstrapping threshold. Diagnosed hypertension did meet criteria. We were unable to find support for the assertion that controlled hypertension was not a bleeding risk factor in the paper cited by Dr. Olesen and colleagues.
Predicting a heterogeneous outcome such as bleeding is a daunting task. We based our model on the rich experience of a large community-based cohort and applied rigorous modeling techniques. We welcome testing of our model in other relevant datasets. Successful external validation plus incorporation into decision aids that thoughtfully balance bleed and stroke risk will dictate the clinical usefulness of our ATRIA bleed risk score.
Please note: Dr. Singer is a consultant for Bayer Health Care, Boehringer Ingelheim, Daiichi Sankyo, Johnson & Johnson, Merck and Co., Sanofi, and Pfizer. All other authors have reported that they have no relationships relevant to the contents of this paper to disclose.
- American College of Cardiology Foundation