Kuijer PMM, Hutten BA, Prins MH, Büller HR. Prediction of the Risk of Bleeding During Anticoagulant Treatment for Venous Thromboembolism. Arch Intern Med. 1999;159(5):457–460. doi:10.1001/archinte.159.5.457
To construct and validate the bleeding risk prediction score, which is based on variables identified in the literature that can be easily obtained before the institution of anticoagulant therapy, in a large independent cohort of patients who were treated with anticoagulant therapy for established venous thromboembolism to allow for quantitative assessment of the risks and benefits of the therapy and to adapt the patient's management accordingly.
We constructed a bleeding risk prediction score, based on variables and their odds ratios identified in the literature, which can be easily obtained before the institution of anticoagulant therapy (score=[1.6×age]+[1.3×sex]+[2.2×malignancy]). Subsequently, we evaluated the score in a test group of 241 patients treated with anticoagulant therapy for venous thromboembolism to determine the optimal cutoff points for the prediction of hemorrhagic complications, using receiver operating characteristic curve analysis. We then validated this score in an independent cohort of 780 patients. A score of 3 or more points, 1 to 3 points, or 0 points represented a high, intermediate, or low bleeding risk, respectively.
The score in about one fifth of the patients in the test group was classified as predicting high risk for bleeding complications. The risk of all bleeding complications was 26% in this group and the risk of major bleeding complications was 14%. The area under the curve was 0.75 (95% confidence interval, 0.64-0.84) and 0.82 (95% confidence interval, 0.66-0.98) for all bleeding complications and major bleeding complications, respectively. When validated, there was a moderate loss of predictive power of the score, but the categorization of the patients by the score remained clinically useful; 20% of the patients were classified as high risk, and the bleeding rate was 17% for all bleeding complications and 7% for major bleeding complications compared with 4% and 1%, respectively, in those categorized as low risk.
With the use of 3 easily obtainable, clinical variables in a prediction model, it is possible to identify a subgroup of patients at the start of anticoagulant therapy who have a high risk of developing hemorrhagic complications. Further studies should address whether additional measures to prevent bleeding decrease the bleeding incidence without compromising efficacy.
CURRENTLY, the range of indications for the use of anticoagulant therapy is steadily growing. There is now evidence from clinical trials that anticoagulant therapy is effective in the primary and secondary prevention of thromboembolic complications in patients with atrial fibrillation, prosthetic heart valves, cardioembolic ischemic cerebral disease, myocardial infarction, and venous thromboembolism (VTE).1
However, bleeding complications during anticoagulant treatment are not uncommon and can be life threatening. These hemorrhagic complications usually occur suddenly, are difficult to predict, and limit the therapeutic benefit of this treatment. The reported incidences of bleeding complications associated with a therapeutic dose of oral anticoagulant therapy vary from 10 to 17 per 100 patient-years for all bleeding complications and from 2 to 5 per 100 patient-years for major bleeding complications.2- 9 Clearly, these figures greatly depend on the characteristics and underlying diseases of the patients. There is also evidence that this risk is associated with the achieved and targeted international normalized ratio (INR).4,9,10
In clinical practice, it would be useful to estimate the bleeding risk for an individual patient before the start of anticoagulant treatment. This would allow the treating physician to make a quantitative assessment of the risks and benefits of the therapy and adapt the patient's management accordingly. Additional measures to prevent bleeding should be considered in patients with an increased risk for bleeding.
In previous studies,2- 9 several clinical and laboratory variables have been identified that are associated with an increased bleeding risk. These include increasing age, sex, the presence of hepatic, cardiac, or malignant diseases, coumarin type, a previous bleeding tendency, and the concomitant use of other drugs, affecting hemostasis. Some investigators6,11- 15 quantitatively combined these variables in risk indexes to predict bleeding. The major disadvantage of these indexes is their limited clinical utility, since they often use variables that are unknown at the start of therapy (such as achieved INR or worsening of liver function), are not usually recorded (such as body surface area), or are ill-defined (such as serious cardiac illness).
In the present study, we constructed a bleeding risk prediction score, based on variables identified in the literature, that can be easily obtained before the institution of anticoagulant therapy. Subsequently, we validated this model in a large independent cohort of patients treated with anticoagulant therapy for established VTE.
Data for the present analysis were derived from the database of the Columbus Investigators16 study, the results of which were published recently. Briefly, in this study a total of 1021 consecutive patients with objectively confirmed VTE were randomly allocated to receive an initial treatment with either subcutaneous low-molecular-weight heparin, reviparin sodium (Clivarin; Knoll AG, Ludwigshafen, Germany), 175 anti-factor Xa U/kg twice daily, or continuous intravenous unfractionated heparin with dose adjustments according to activated partial thromboplastin time. Both treatment groups received oral anticoagulant therapy (warfarin sodium or coumarin), which started on the first or second day of treatment, for at least 3 months, with a targeted therapeutic INR of 2.0 to 3.0. Study outcomes evaluated during a 3-month follow-up period were the incidence of symptomatic recurrent VTE, which was confirmed by objective investigations, and the incidence of bleeding complications. All outcomes were assessed by an independent adjudication committee that was unaware of treatment allocation. Bleeding was defined as major if it was clinically overt and associated with a decline in hemoglobin concentration of at least 20 g/L, if there was a need for transfusion of 2 U or more of red blood cells, if it was retroperitoneal or intracranial, or if it warranted permanent discontinuation of treatment. During the 3-month observation period, 93 clinically important bleeding events occurred, 28 of which were classified as major. About half of the bleeding episodes occurred in the first 14 days of treatment, while the other hemorrhages were distributed over the remaining 10 weeks of follow-up. The overall incidence of bleeding complications was 9% in the patients treated with low-molecular-weight heparin (3.1% major bleeding episodes) and 9.2% in the patients treated with unfractionated heparin (2.3% major bleeding episodes). Hence, the 2 treatment regimens are considered to be equivalent with respect to safety.16 Therefore, we combined the data of the 2 treatment groups for the present evaluation of the prediction of hemorrhage. To assess the intensity of oral anticoagulation, the time spent in certain INR ranges (<2.0, 2.0-3.0, and >3.0) was calculated for each patient during the 3-month period.
We performed a critical review of the literature for variables and models used to predict bleeding during anticoagulant therapy.1- 15 Based on this information, a score was constructed with the variables and their odds ratios, which were repeatedly associated in the various studies with an increased bleeding risk and can be obtained at the start of anticoagulant treatment. These variables, the odds ratios, and the resulting score are detailed in Table 1.
The evaluation of the score was performed in 2 phases. In the first phase, the constructed score was evaluated in a subset of 241 patients (the test group) to determine the optimal cutoff points for the prediction of all hemorrhagic complications as well as major bleeding complications using receiver operating characteristic curves.17 Three groups were defined: those at low, moderate, or high risk of bleeding complications during the 3-month period of anticoagulant treatment. The 241 patients in the test group, 24% of the entire study population, were all consecutive patients included in 5 representative study centers participating in the Columbus Investigators16 trial. In these patients, additional potential risk factors for bleeding were collected at baseline, such as the presence of atrial fibrillation, heart failure, hypertension, liver dysfunction, recent surgery, stroke or trauma, previous gastrointestinal tract bleeding, and known alcohol abuse. Using logistic regression, variables with a P value less than .05 in the univariate analysis were candidates for the score. Finally, the data from the test group were used to investigate whether the score could be optimized or further simplified without loss of accuracy. In the second phase, the score with the defined cutoff points was validated in the remaining 780 patients in the trial. In this phase, both the original and the simplified score were evaluated.
In Table 2 the clinical characteristics at baseline and the incidences of the recurrent VTE and bleeding complications during a 3-month follow-up period in the test and validation groups are detailed. Half of the patients were male, the mean age was approximately 60 years, and about a quarter of the patients had primary pulmonary embolism. The 2 study groups were comparable at baseline with respect to clinical characteristics, and, during follow-up, similar outcome rates were observed.
In the test group of 241 patients, the bleeding risk prediction score resulted in an area under the curve of 0.62 (95% confidence interval [CI], 0.50-0.75) for all bleeding complications and 0.72 (95% CI, 0.52-0.92) for major bleeding complications. Based on the receiver operating characteristic curves, the high-risk group was defined as having a score higher than 6.25, and the low-risk group was defined as having a score of 3.75 or lower. In the high-risk group, the incidence of all bleeding complications was 15% (7 patients) and major bleeding episodes occurred in 7% (3 patients). In contrast, in the low-risk group, these incidences were 5% (2) and 0% (0), respectively. When the variables of body surface area and coumarin type were removed from the model, the resulting area under the curve improved and became 0.75 (95% CI, 0.64-0.84) for all hemorrhagic complications and 0.82 (95% CI, 0.66-0.98) for major bleeding episodes. None of the additional risk factors collected at baseline showed significant results in logistic regression analysis. Hence, all subsequent analyses were performed with the simplified score, consisting of the variables age, sex, and known malignancy (Table 1). The high-risk patients were defined as those with a score of 3 or more points, while patients with a score of 0 points were considered low risk. A score of 1 to 3 points indicated an intermediate risk. The resulting classification of the patients in the test group in the low-, moderate-, and high-risk groups and the incidences of bleeding episodes using this simplified score are shown in Table 3. In the low-risk group (42 patients [17%] in the test group), no bleeding complications occurred. In contrast, of those patients classified as high risk, 13 patients (26%; 95% CI, 14.6-40.4) experienced a bleeding complication, while in 7 (14%) of them (95% CI, 5.8-26.7) a major bleeding episode occurred. Of the 241 patients in the test group, 51 (21%) with VTE who began anticoagulant therapy were classified as high risk. Of all 22 bleeding episodes during the 3-month observation period, 13 (60%) occurred in the high-risk patients. Similarly, 7 (78%) of 9 major bleeding episodes were seen in this patient category.
In the validation group of 780 patients, the application of the bleeding risk prediction score again identified approximately 20% of the patients as high risk (Table 4). The 3-month incidence of all bleeding complications was 17% (95% CI, 11.3-23.4), while the frequency of major bleeding was 7% (95% CI, 3.2-11.9) in these patients. In those patients designated low risk (170 [22%]), 6 patients (4%) experienced a bleeding complication, of which 1 (1%) was major. The intensity of anticoagulant therapy among the 3 risk categories, expressed as time spent in various INR ranges, was comparable between the 3 groups in all patients in the study by the Columbus Investigators (data not shown).
For every therapy instituted, it would be ideal if the risk of complications associated with this treatment could be reliably predicted in the individual patient. The findings of the present study indicate that with 3 simple, easily obtainable clinical variables, such as age, sex, and the presence of malignancy, it is possible to identify a subgroup of patients at the start of anticoagulant therapy who have a high risk of developing hemorrhagic complications.
In patients treated for VTE, our model categorized approximately one fifth of the patients to be at high risk for bleeding complications. In this group, the incidence of all bleeding complications during anticoagulant therapy was 17%, of which more than one third met the criteria for major bleeding. These rates were about 6 to 7 times higher than the rates observed in those identified as low risk. This difference in bleeding incidence was not due to differences in the intensity of anticoagulant treatment among the various risk categories.
Several aspects of our investigation warrant comment. As documented previously for other prediction rules, we also observed a moderate decrease in the predictive power of the prediction score when validated. This illustrates the importance of proper independent evaluation of derived prediction rules in a large group of similar patients. Nevertheless, the categorization by the score remained clinically useful since the absolute risk of bleeding, in particular major bleeding, is substantially higher in the high-risk group than the low-risk group.
To construct our bleeding risk prediction score, we selected variables that have been previously associated with an increased risk of bleeding complications. The advantage of the present score is that these variables have been combined to allow for a quantitative assessment of the bleeding risk at an individual level at the start of anticoagulant treatment.
We studied patients with VTE who received an initial 5-day course of heparin, concomitant with the use oral anticoagulant therapy, which was continued for 3 months. Hence, application of the prediction score for other groups of patients treated with anticoagulant therapy should be done with caution, although it is unlikely that the indication for anticoagulant treatment per se can be considered a risk factor for hemorrhagic complications.
What are the possible implications of the study findings? If the validity of the bleeding risk prediction score can be confirmed, subsequent study should address whether additional measures to prevent bleeding complications in those designated as high risk, such as aiming for a lower target INR, closer laboratory monitoring, and careful use of concomitant medication, decrease the incidence of bleeding complications without compromising the efficacy. This is particularly important since these patients are the same patients with an increased risk for recurrent episodes of VTE, in whom adequate anticoagulation therapy is indicated.
We conclude that a simple quantitative combination of the patient's age, sex, and the presence of malignancy allows for a clinically meaningful bleeding risk categorization of the patient who begins to use anticoagulant treatment for VTE.
Accepted for publication July 14, 1998.
Harry R. Büller, MD, PhD, is an established investigator of the Dutch Heart Foundation.
Reprints: Philomeen M. M. Kuijer, MD, PhD, Center for Haemostasis, Thrombosis, Atherosclerosis, and Inflammation Research, Academic Medical Center, Meibergdreef 9, 1105 AZ Amsterdam, the Netherlands (e-mail: email@example.com).