Risk Factors for Intracranial Hemorrhage in Outpatients Taking Warfarin

Methods

Case-Patient Identification and Eligibility

Using a discharge log of consecutive admissions to the Massachusetts General Hospital during the period from 1 January 1981 through 31 December 1991, we identified 1881 patients with a principal diagnosis of intracranial hemorrhage (ICD-9 codes 430, 431, 432.0, 432.1, and 432.9). Forty-one (2%) medical records for these patients could not be located. In the remaining 1840 case-patients, warfarin use was determined from review of the neurologist, neurosurgery resident, and attending staff physician admission notes. Intracranial hemorrhage was verified by computed tomographic (CT) scanning, lumbar puncture, or postmortem examination in all but one case-patient. In this latter case-patient, the diagnosis was based on clinical grounds because the patient died before diagnostic studies were done and no autopsy was performed.

To be eligible for the study, patients had to be at least 18 years old and taking warfarin as an outpatient. Patients with an anatomic abnormality or underlying bleeding diathesis predisposing to intracranial hemorrhage, regardless of anticoagulant therapy, were not included. Hemorrhages sustained as a result of major head trauma (with skull fracture and loss of consciousness) also were not eligible.

Of 131 patients with intracranial hemorrhage identified, 10 were excluded as ineligible: Four patients had subarachnoid hemorrhage resulting from angiographically identified intracranial aneurysms; 2 had hemorrhage into primary or metastatic tumors; 2 had acute subdural hemorrhage after major head trauma; 1 bled after multiple craniotomies and radiation therapy for a recurrent craniopharyngioma; and another patient had aplastic anemia.

Controls: Source and Matching

Controls were selected from the registry of the Massachusetts General Hospital's anticoagulant therapy unit. During the study period, this unit managed warfarin dosing for approximately 8000 patients referred from all hospital clinical services. The most common indications for anticoagulation were atrial fibrillation, previous stroke, presence of prosthetic heart valves, and venous thromboembolism. Approximately 40% of the patients managed by the anticoagulant therapy unit have prothrombin time tests done at the Massachusetts General Hospital. An additional 20% have their prothrombin times measured at one large commercial laboratory (via home phlebotomy services). The remaining 40% have their tests done at various local laboratories.

Each case-patient was matched to three randomly selected controls taking warfarin at the time of the case-patient's intracranial hemorrhage. Matching was accomplished by first identifying all patients managed by the anticoagulant therapy unit at the date of hospital admission for the given case-patient. Each potential control was assigned a random number. The three controls with the lowest random numbers were selected. No control was used more than once. Three controls subsequently became case-patients. Matching was done to control for any change in background risk over the 11-year study period (for example, from changes in prothrombin time targets or technique of measuring the prothrombin time).

Data Collected

Clinical features of case-patients and controls were extracted primarily from hospital records, with supplementation in a few instances from physician office records. Variables were entered on a predesigned data form and included indication for anticoagulant therapy; race; sex; age; prothrombin time ratio (PTR); duration of warfarin therapy; history of diagnosed hypertension, stroke, transient ischemic attack, diabetes mellitus, myocardial infarction, atrial fibrillation, and congestive heart failure; and medications. Clinical features of the intracranial hemorrhages were also recorded. Clinical data for the controls were current at the admission date of the matched case-patient with intracranial hemorrhage.

The prothrombin time ratio was expressed as the ratio of the patient's value divided by the simultaneously reported control value. For case-patients, we used the PTR on admission, or, if available, the PTR closest to the reported onset of symptoms. For controls, the PTR closest to that of the case-patient admission date was recorded from the anticoagulant therapy unit database. Because values for the international sensitivity index (ISI) for thromboplastins were not universally reported before 1988, we analyzed our data using the PTR rather than the international normalized ratio (INR). Since 1988, Massachusetts General Hospital's hematology laboratory has used Simplastin Automated (Organon Teknika Corporation, Durham, North Carolina), with values for the ISI ranging from 1.9 to 2.0. During the study period, this company (previously General Diagnostics) supplied the thromboplastin and assisted our laboratory in selecting lots of comparable sensitivity. Studies were routinely done to minimize year-to-year variation in thromboplastin sensitivity. Since 1988, the most frequently used commercial laboratory has used thromboplastins with ISI values ranging from 1.9 to 2.1. The PTR was missing for four case-patients; the data for case-patients and controls were otherwise complete.

Selected Relevant Definitions

Hypertension was defined as probable if the patient had such a diagnosis listed in the medical record and as definite if the patient was receiving antihypertensive medication. When hypertension was diagnosed, it was classified as definite 97% of the time for case-patients and 93% of the time for controls. All diagnoses of hypertension are used in the analyses.

Patients with documented carotid or vertebrobasilar disease and those with a history of previous stroke were defined as having cerebrovascular disease. Diagnoses of carotid or vertebrobasilar disease or both were confirmed by angiography in 71% of case-patients and 93% of controls, by Doppler studies in 17% of case-patients and 2% of controls, and solely by assessment by a neurologist in 12% of case-patients and 5% of controls.

Statistical Analysis

Case-patients and controls were compared using chi-square tests and the Fisher exact test, where appropriate, for categorical variables and using the Student t-test for continuous variables. Univariate odds ratios were calculated using unmatched and matched techniques. The matched odds ratios were provided by the Mantel-Haenszel [12] summary statistic across matched sets. The unmatched and matched techniques provided very similar estimates of odds ratios. We report the unmatched results. Confidence intervals for odds ratios were calculated using the Taylor series method [13]. The test of trend was done using the Cochran-Mantel-Haenszel test [14]. Logistic regression models assessed the independent effect of multiple clinical features and the significance of interaction terms. Conditional logistic models [15] accounting for matching provided estimates similar to those of the unmatched logistic regression analyses. The estimates from the unmatched analyses are reported. Data were recorded in R:BASE (Microrim, Bellevue, Washington) from the paper data forms. Statistical analyses were done using SAS (SAS Institute Inc., Cary, North Carolina) and GLIM (Numerical Algorithms Group Limited, Oxford, United Kingdom).

Results

Clinical Course of Case-Patients

During the 11-year study period, 121 patients with intracranial hemorrhage were eligible; 77 hemorrhages were intracerebral and 44 were subdural (Table 1). Three of the patients with subdural hemorrhage had a history of trivial head trauma; the others had no known antecedent head trauma. For the patients with intracerebral hemorrhages, headache was the most common presenting feature (53%), followed by nausea and vomiting (40%) and unresponsiveness (36%). Seventy-eight percent of the case-patients presented to the emergency department within 24 hours of the onset of symptoms and 87% within 48 hours. In contrast, only 36% of the patients with subdural bleeding came to medical attention within this same period. Forty-six percent of patients with intracerebral bleeding died, and 17% survived with major neurologic deficits that prevented subsequent independent living.

The most common presenting complaint in those patients with a subdural hemorrhage was also headache (68%), followed by confusion (39%) and ataxia (32%). Twenty percent of patients with subdural bleeding died before discharge, and 9% had major neurologic sequelae.

Comparison of Case-Patients and Controls

The case-patients and controls did not differ significantly in sex; race; or past medical history of diagnosed hypertension, diabetes mellitus, or congestive heart failure (Table 2). The PTR was the dominant risk factor for both intracerebral and subdural hemorrhage, with the risk for both increasing dramatically for PTR values above 2.0. The relation between the PTR and the relative odds of intracranial hemorrhage is shown in Figure 1. Age also achieved significance as a risk factor for both intracerebral and subdural hemorrhage in the univariate analysis, although it was more important as a risk factor for subdural hemorrhage (Table 2). The mean age of patients with intracerebral hemorrhage was 69 years (range, 28 to 92 years) compared with a mean age of 64 years (range, 18 to 92 years) for controls (P = 0.003). For patients with subdural hemorrhage, the mean age was 73 years (range, 52 to 88 years) compared with 64 years (range, 20 to 88 years) for controls (P = 0.001). Additional risk factors for intracerebral bleeding included a history of cerebrovascular disease (odds ratio, 2.9; 95% CI, 1.7 to 4.9) and the presence of a prosthetic heart valve (odds ratio, 2.0; CI, 1.1 to 3.8).

For subdural hemorrhage, other variables achieving significance in the univariate analyses included a history of atrial fibrillation (odds ratio, 2.9; CI, 1.4 to 5.8), a history of myocardial infarction (odds ratio, 0.3; CI, 0.1 to 0.7), and duration of warfarin therapy more than 5 years (odds ratio, 2.8; CI, 1.3 to 6.2).

Multivariate Analyses

In multiple logistic models the PTR remained the dominant independent risk factor for both intracerebral and subdural hemorrhage (Table 3). For subdural hemorrhage, a term representing the square of the PTR was significant. As a result, the estimated odds of subdural hemorrhage increased 7.6-fold as the PTR increased from 2.0 to 2.5. For intracerebral hemorrhage, the risk rose fourfold for each unit increase in the PTR. (The corresponding odds ratio for a 0.5 increase in the PTR was 2.1.) Age was the only other independent risk factor for subdural hemorrhage (odds ratio, 2.0 per decade; CI, 1.3 to 3.1).

For intracerebral hemorrhage age was of borderline statistical significance after controlling for the PTR and the other independent risk factors: history of cerebrovascular disease (odds ratio, 3.1; CI, 1.7 to 5.6) and presence of a prosthetic heart valve (odds ratio, 2.8; CI, 1.3 to 5.8). The independent variables had no significant correlations and no significant two-way interactions.

Combined Analysis

When all 121 case-patients of both types were combined, independent risk factors for intracranial hemorrhage were PTR (odds ratio, 4.5 per unit increase; CI, 2.6 to 8.0), history of cerebrovascular disease (odds ratio, 2.3; CI, 1.4 to 3.7), presence of a prosthetic heart valve (odds ratio, 2.1; CI, 1.2 to 3.8) and age (odds ratio, 1.4 per decade; CI, 1.2 to 1.7).

Analysis of Case-Patients and Controls Managed by the Anticoagulant Therapy Unit

Fifty-five case-patients with intracranial hemorrhage had had their warfarin therapy managed by the anticoagulant therapy unit. Analyses restricted to these case-patients and their respective controls represent a case–control study nested within the defined anticoagulant therapy unit population. The effects of previously identified independent risk factors for intracranial hemorrhage were similar in this subset of our study population: prothrombin time ratio (odds ratio, 8.6 per unit increase; CI, 3.3 to 22.6), cerebrovascular disease (odds ratio, 2.0; CI, 1.0 to 4.1), prosthetic heart valve (odds ratio, 1.6; CI, 0.6 to 3.9), and age (odds ratio, 1.5 per decade; CI, 1.1 to 2.0).

Discussion

Recent randomized trials have greatly expanded the number of persons with strong indications for long-term anticoagulant therapy. In particular, anticoagulants largely reverse the increased risk for stroke among patients with atrial fibrillation, a group that constitutes 4% of persons older than 60 years and 10% of persons older than 80 years [16]. The decision to use anticoagulants hinges on the balance between the decreased risk for thromboembolism and the increased risk for hemorrhage. Intracranial hemorrhage assumes great importance in these considerations because it is the only complication that is comparable in severity to the thromboembolic events that anticoagulants prevent. The anticoagulant decision would clearly benefit from powerful studies identifying risk factors for intracranial hemorrhage. Low absolute event rates have hindered previous prospective studies. Estimates of the absolute rate of anticoagulant-related intracranial hemorrhage range from 0.3% per year observed in the first five trials in atrial fibrillation [2-6] to approximately 2% per year reported from an observational study [9] as well as from a randomized trial studying patients older than 75 years [17]. Rates of 0.5% to 0.6% per year were found in two large observational studies from anticoagulant therapy units [10, 11]. Unfortunately, rates of intracranial hemorrhage as low as 1% to 2% per year can reverse the decision to use anticoagulants [1].

We chose a case–control design to overcome the problem of the low absolute event rate of intracranial hemorrhage. Our study's 121 case-patients represent roughly the equivalent of 20 000 person-years of prospective follow-up. Large population-based studies have estimated that anticoagulants raise the risk for intracranial hemorrhage tenfold [18-20], implying that nearly all this risk is attributable to anticoagulant therapy itself and dictating that our controls should also have been taking anticoagulants.

Our large number of case-patients allowed for separate risk assessment for intracerebral and subdural hemorrhage. The PTR was the dominant independent risk factor for both. The relation between intensity of anticoagulant therapy and the risk for major bleeding is well known [21] but has not been well quantified for intracranial hemorrhage. We found that for each 0.5 increase in PTR, the risk for intracerebral bleeding doubled. For subdural hemorrhage, the risk was unchanged between PTRs ranging from 1.0 to 2.0 but rose dramatically for those above 2.0. Our risk estimates relate to the PTRs that actually occurred in the study patients, regardless of their target PTR. It is likely that the intensity of anticoagulation is an even greater risk factor for intracranial hemorrhage than our data indicate because the PTR is not well standardized [22, 23]. One can anticipate that the association between INR and intracranial hemorrhage would be stronger and might reveal a monotonic relation with risk even at the low end of “therapeutic” intensity of anticoagulation (for example, INR, 2.0 to 3.0). Even in our data set, most of the intracranial hemorrhages occurred at PTRs below 2.0. Since 1988, the two laboratories serving most of our study patients have recorded ISI values for their thromboplastins. This ISI range has been 1.9 to 2.1. As a result, our PTR threshold of 2.0 corresponds to an INR range of 3.7 to 4.3.

The high mortality rate associated with anticoagulant-related intracranial hemorrhage has been attributed to the increased volume of the hematoma, reported to be approximately double that occurring with spontaneous intracerebral hemorrhage [17, 24]. Radberg and colleagues [24] also reported an increased hematoma volume and mortality rate with higher INR values.

Previous studies have reported conflicting conclusions about the role of age [9, 10, 25]. For example, Fihn and colleagues [10] showed a relative risk of 0.8 for persons older than 65 years with an upper 95% CI limit estimate of 1.0, whereas Landefeld and Goldman [9] found a highly significant (P < 0.001) relative risk of 3.2. The report of the Second Stroke Prevention in Atrial Fibrillation (SPAF-II) Trial has noted an intracranial hemorrhage rate of 1.8% per year for patients receiving anticoagulation who were older than 75 years compared with 0.5% per year for patients receiving anticoagulation who were 75 years or younger [17]. Age was a strong independent risk factor for subdural hemorrhage in our study, nearly doubling the risk with each 10-year increase, perhaps reflecting the increased fragility of bridging veins and the effect of cerebral atrophy [26].

For intracerebral hemorrhage, age was of borderline statistical significance in the multiple regression model. The incidence rates for intracerebral hemorrhage in patients not taking warfarin clearly rise with age [27], an effect attributed to age-related pathologic changes [28, 29]. It is not so clear that age increases the risk for intracerebral hemorrhage among patients taking anticoagulants, where other features, in particular, the intensity of anticoagulation, assume greater importance.

Cerebrovascular disease and prosthetic heart valves were the other identified independent risk factors for intracerebral bleeding. Levine and Hirsh [30] have previously noted that major bleeding in patients taking warfarin was especially likely in patients with ischemic cerebrovascular disease and was frequently intracerebral in origin, and Landefeld and Goldman [9] have reported a univariate relative risk of 6.6 for intracranial hemorrhage in patients with a history of past or current stroke.

It is unclear why patients with prosthetic heart valves should be at increased risk for intracerebral hemorrhage. The effect of prosthetic heart valves appears to be independent of higher intensity of anticoagulation. It also seems unlikely that these intracerebral hemorrhages were initially embolic infarctions because such hemorrhagic transformations are unusual and occur mainly with large ischemic strokes [31]. Several recent randomized, controlled trials in patients with prosthetic heart valves have shown decreased rates of major bleeding with less intense anticoagulation without loss of antithrombotic efficacy [32, 33]. Our data further support the use of lower intensities of anticoagulation in patients with prosthetic valves.

Only 4% of the case-patients with intracranial hemorrhage had a history of head trauma (including the two case-patients with major head trauma who were excluded). The rarity of antecedent trauma in anticoagulant-related cases of chronic subdural hemorrhages has been previously documented [34, 35]. Acute nontraumatic subdural hemorrhage seems almost exclusively dependent on anticoagulant therapy [34].

We did not find any independent association between duration of anticoagulant therapy and intracranial hemorrhage, nor did we find any risk associated with sex or treated hypertension. Because patients on warfarin therapy are highly selected and closely followed, a cohort of truly untreated hypertensive patients who are receiving warfarin therapy may not exist.

Potential methodologic limitations of this study merit discussion. In a case–control study, controls should be a representative sample of the population that gives rise to the case-patients [36]. Ideally, controls for this study would have been a random selection of patients taking warfarin who would have come to our hospital because of a bleeding complication. Such a population is difficult to define unambiguously, and no registry of such patients exists. However, the patients receiving care from our anticoagulant therapy unit do constitute a comparable cohort. Our substudy of case-patients arising from the anticoagulant therapy unit and their controls constitutes an unbiased case–control study nested within the anticoagulant therapy unit cohort. The fact that our results were largely the same in this subgroup argues strongly for the lack of substantial selection bias overall.

Our case–control design does not directly provide the absolute risk for intracranial hemorrhage associated with each risk factor. However, the average absolute risk observed in recent prospective studies can be coupled with our findings (including the distribution of risk factors in our control group) to provide estimates of the absolute risk in any risk factor category. For example, using the risk estimate of van der Meer and colleagues [11]—0.6% per year—one can estimate that a PTR greater than 2.0 conveys an important absolute risk for intracranial hemorrhage of 2% per year. Finally, we note that although case–control studies can add efficiency and power to the study of infrequent events, they may fail to identify important but uncommon risk factors.

Recent randomized trials have established the efficacy of long-term warfarin therapy in preventing thromboembolic complications of several common chronic conditions. Nonetheless, warfarin remains a demanding and risky therapy and should be used only in those groups of patients where the expected benefit clearly exceeds the risk. The randomized trials coupled with observational epidemiologic studies have helped to identify groups at high risk for embolic events in certain conditions [37]. Much less information is available on risk factors for intracranial hemorrhage. Our study shows the importance of the effect of increasing age on risk for subdural hemorrhage with warfarin and of previous cerebrovascular disease and of prosthetic heart valves on the risk for intracerebral hemorrhage. Most importantly, our study shows the dramatic increase in risk for intracranial hemorrhage with increasing PTR, especially PTRs greater than 2.0, underlining the need for careful control of anticoagulation at the lowest effective intensity, particularly among the many older patients for whom warfarin is now indicated.