To date, there have been few studies evaluating outcomes of patients with atrial fibrillation (AF) who have experienced gastrointestinal (GI) hemorrhages. We examined short- and long-term mortality of major GI hemorrhage in patients with AF on and off warfarin in recent clinical care. We evaluated this association in the large Anticoagulation and Risk Factors in Atrial fibrillation (ATRIA) and ATRIA-Cardiovascular Research Network (CVRN) California community-based cohorts of patients with AF (study years 1996 to 2003 and 2006 to 2009, respectively), where all events were clinician adjudicated. We used proportional hazards regression with propensity score adjustment to estimate the short- (30 days) and long-term (>30 days for 1 year) mortality rate ratio for patients using warfarin compared with those who were not using warfarin at the time of GI hemorrhage. In the 414 ATRIA participants with major GI hemorrhage, 54% were taking warfarin at the time of the hemorrhage; in the 361 ATRIA-CVRN participants with major GI hemorrhage, 58% were taking warfarin. Warfarin use at the time of GI hemorrhage was not associated with 30-day mortality in the ATRIA cohort but was associated with significantly reduced 30-day mortality in the ATRIA-CVRN cohort (adjusted mortality rate ratio [95% confidence interval], ATRIA 0.97 [0.54 to 1.74]; ATRIA-CVRN 0.38 [0.17 to 0.83]). There was a modest suggestion of lower mortality on warfarin after 30 days in both cohorts. In conclusion, our study demonstrates that GI hemorrhages on warfarin are certainly no worse and may be less life threatening than those occurring off warfarin. These findings are in stark contrast to the deleterious effect of warfarin on mortality from intracranial hemorrhage and add another factor favoring anticoagulation in clinical decision making for patients with AF.
Highlights
- •
Gastrointestinal hemorrhages on warfarin may be less life threatening than those occurring off warfarin.
- •
These findings are in stark contrast to the deleterious effect of warfarin on mortality from intracranial hemorrhage.
- •
These findings add to the positive net clinical benefit of warfarin therapy for most subgroups of patients with atrial fibrillation.
To date, there have been few studies evaluating outcomes of patients with atrial fibrillation (AF) who have experienced gastrointestinal (GI) hemorrhages, the most common extracranial hemorrhage site. Although intracranial hemorrhages are more severe, extracranial hemorrhages are more common and can be fatal. In warfarin users, intracranial hemorrhages are much more likely to result in death than GI hemorrhages. Although warfarin exposure significantly worsens outcomes from intracranial hemorrhage, it is not known whether anticoagulation similarly affects mortality from GI hemorrhage. This is a particularly relevant question for patients with AF as it may further inform the decision to use warfarin therapy. These considerations apply as well to the novel anticoagulants for which GI hemorrhage remains an important concern. The present study aimed to assess the association between warfarin use at the time of GI hemorrhage and the rate of short-term (30 days) and long-term (1 year of follow-up after surviving 30 days) mortality in patients with AF in clinical care. We studied 2 large community-based cohorts with comprehensive follow-up and clinical adjudication of hemorrhages.
Methods
Patients included in these analyses were members of 2 cohort studies of patients with AF from Kaiser Permanente (KP) Northern and Southern California, 2 large integrated health care delivery systems. The Anticoagulation and Risk Factors in Atrial fibrillation (ATRIA) cohort includes 13,559 adults aged 18 years and older with diagnosed nonvalvular AF who received care with KP Northern California. The ATRIA-Cardiovascular Research Network (CVRN) cohort includes 33,247 adults aged 18 years and older with incident nonvalvular AF who were enrolled in either KP Northern or KP Southern California. In each sample, cohort members were identified by searching electronic inpatient, outpatient, and electrocardiographic databases for physician-assigned International Classification of Diseases, Ninth Revision, Clinical Modification , diagnostic codes of AF (427.31, 427.32). Patients from the ATRIA cohort were identified between July 1996 and December 1997 and followed through September 2003, and patients from the ATRIA-CVRN cohort were identified between January 2006 and June 2009, with follow-up through June 2009. Because we were interested in nontransient nonvalvular AF, we excluded patients with AF with diagnosed mitral stenosis (ATRIA only), valvular repair or replacement (ATRIA only), transient postoperative AF, or concurrent hyperthyroidism. Unlike the ATRIA cohort, the ATRIA-CVRN cohort did not exclude patients with mitral stenosis or a history of a valve replacement; such patients accounted for only 1.5% of the ATRIA-CVRN cohort.
We focused on patients who were hospitalized for major GI hemorrhage during the observation period of the full ATRIA and ATRIA-CVRN studies. GI hemorrhages were identified by searching comprehensive hospitalization and billing databases for primary discharge diagnoses ( Supplementary Material A ). Major GI hemorrhages were defined as fatal (death during hospitalization) or requiring transfusion of 2 or more units of packed red blood cells. GI hemorrhages not leading to hospitalization were excluded. Because KP is an insurer and a health care provider, cohort members admitted to facilities outside the network were still identified by our search strategy. Each potential GI hemorrhagic event was reviewed by an Outcomes Review Committee composed of physicians. This resulted in identification of 712 patients in the ATRIA cohort who experienced a hospitalization for a validated GI hemorrhage, with 414 (58%) classified as major hemorrhages. In the ATRIA-CVRN cohort, 601 patients experienced a validated GI hemorrhage, with 361 (60%) classified as major hemorrhages.
Warfarin use at the time of the GI hemorrhagic event was assessed through manual review of the admission medical record. Patients were considered to be using warfarin if the admission medical record indicated that they were taking warfarin within 5 days before admission for GI hemorrhage. We recorded the international normalized ratio (INR) value at admission for each patient before reversal of warfarin effect. For the ATRIA cohort, we recorded the number of units of blood that were transfused and whether fresh frozen plasma (FFP) was used. For the ATRIA-CVRN cohort, we recorded whether the number of transfused units of blood was ≥2 and did not record use of FFP. Deaths were determined through reviewing medical charts, health plan databases, Social Security Administration vital status file, and the comprehensive California State death certificate registry.
Patient characteristics were obtained from administrative databases. Low educational attainment status (<high school graduation) and low-income status (annual household income < $35,000) were assigned from geocoding based on the US Census block group data. Current smoking status was based on patient self-report. History of GI hemorrhage, fall, and co-morbid conditions, including ischemic stroke, intracranial hemorrhage, diabetes mellitus, coronary disease, hypertension, chronic heart failure, chronic liver disease, cancer (excluding nonmelanomatous skin cancer), dementia, lung disease (ATRIA-CVRN only), and transient ischemic attack (ATRIA-CVRN only), were collected from clinical inpatient and ambulatory databases using validated algorithms and were assessed using data during the 5 years before the date of hospitalization of GI hemorrhage. Kidney dysfunction (defined as estimated glomerular filtration rate < 30 mL/min/1.73 m 2 ) was calculated from the most recent laboratory values before the GI hemorrhage using the Chronic Kidney Disease Epidemiology Collaboration formula. Patients without an estimated glomerular filtration rate value within the 6 months before the GI hemorrhage were considered to have normal renal function. The most recent outpatient hemoglobin (Hgb) value from 30 days before to 1-year before the GI hemorrhage was used to identify patients with anemia (women Hgb < 13.5 g/L; men Hgb < 12.0 g/L). Patients without a hemoglobin value during this time period were considered not to have anemia. Use of aspirin at the time of GI hemorrhage was determined from manual review of the admission medical record. Use of other antiplatelet agents, statins, or other lipid-lowering agents, angiotensin-converting enzyme inhibitors, and angiotensin II receptor antagonists (angiotensin receptor blockers) at the time of GI hemorrhage were determined from information based on dispensings from health plan ambulatory pharmacy databases. For all patients, the CHADS 2 , CHA 2 DS 2 -VASc, ATRIA stroke, HAS-BLED, and ATRIA hemorrhage risk scores were calculated.
For the analysis of warfarin use at the time of GI hemorrhage and 30-day mortality, the index date was the date of admission to the hospital for GI hemorrhage. Follow-up began the day after admission and ended at the first occurrence of any of the following: death, censoring because of withdrawal from the KP network, or after 30 days. For the long-term mortality analysis, the index date was defined as the 31st day after hospital admission for GI hemorrhage (to exclude immortal person-time specific to this time period ) and ended at the first occurrence of any of the following: death, censoring because of withdrawal from the KP network or 1 year after the index date (days 31 to 396). To examine which patients resumed use of or initiated use of warfarin, we looked for a new prescription of warfarin during 1 year of follow-up.
We found heterogeneity in the crude association of warfarin use at the time of GI hemorrhage and mortality between the 2 cohorts (ATRIA and ATRIA-CVRN) and, therefore, decided to examine the cohorts separately rather than as a combined data set ( Supplementary Material B ). We conducted descriptive and stratified analyses to examine the distribution of potential confounders in those using and not using warfarin. Cox proportional hazards regression was used to estimate the crude and adjusted mortality rate ratio (mRR) for the association between warfarin use at the time of GI hemorrhage and mortality (SAS, version 9.2; SAS Institute, Cary, North Carolina). Additionally, we examined models stratified by age group (<85 and ≥85 years).
To control for baseline confounding, we used Cox proportional hazard regression models adjusted for propensity score. Propensity scores were generated from logistic regression models to estimate the predicted probability of using warfarin at the time of GI hemorrhage versus not using warfarin. Propensity scores were generated for 4 separate analyses: (1) ATRIA sample: 30-day mortality, (2) ATRIA-CVRN sample: 30-day mortality, (3) ATRIA sample: long-term mortality, and (4) ATRIA-CVRN sample: long-term mortality.
For each of these 4 analyses, we fit a preliminary proportional hazards model predicting death that included all potential confounders to determine the strength of the association of each covariable with either short- or long-term mortality. To avoid including nonconfounding predictors of warfarin use in the propensity score model, only those variables with a hazard ratio (or 1/hazard ratio) of at least 1.2 from the preliminary outcome model were used to create the propensity scores. After generating the propensity scores, we trimmed 2.5% of patients on both extremes of the propensity score distribution to improve comparability of the warfarin-use and non–warfarin-use groups. These observations were outside the primary area of overlap of the propensity scores and increase residual confounding. The Cox proportional hazards models were adjusted for propensity score quintile ( Supplementary Material C ). We also performed additional analyses where the propensity score stratification was more finely divided (8 and 10 strata, respectively).
Results
In the 414 patients in ATRIA with major GI hemorrhages, 54% were taking warfarin. Major GI hemorrhages included 205 (49%) upper GI bleeds, 94 (23%) lower GI bleeds, and 115 (28%) where the location was undetermined by the time of hospital discharge. At the time of GI hemorrhage, patients using warfarin were younger, much less likely to be using aspirin, less likely to have diagnosed dementia or a history of GI hemorrhage, and were more likely to have a history of ischemic stroke and to be using lipid-lowering agents and angiotensin-converting enzyme inhibitors, compared with patients not using warfarin. There was no statistically significant difference in CHADS 2 , CHA 2 DS 2 -VASc, or ATRIA stroke risk scores between patients using or not using warfarin at the time of their bleeding event, whereas warfarin users had clearly lower HAS-BLED and ATRIA hemorrhage risk scores ( Table 1 ). Characteristics of warfarin users and nonusers within propensity score quintiles were quite similar ( Supplementary Material C ). Overall, 398 (96%) patients were transfused with a median number of 3.0 (interquartile range [IQR]) units of blood. In warfarin users, the median INR at presentation was 3.0 (IQR). FFP was used in 56% of warfarin users.
| ATRIA | ATRIA-CVRN | |||||
|---|---|---|---|---|---|---|
| Warfarin use at GI hemorrhage (n = 223) | No Warfarin at GI hemorrhage (n = 191) | P-Value | Warfarin use at GI hemorrhage (n = 210) | No Warfarin at GI hemorrhage (n = 151) | P-Value | |
| Variable | ||||||
| Age (years), mean (SD) | 77.4 (7.1) | 79.9 (9.3) | 0.002 | 76.0 (8.0) | 77.9 (9.1) | 0.03 |
| Asian/Pacific Islander | 5.8% | 6.3% | 0.78 | 9.5% | 9.3% | 0.33 |
| Black | 5.4% | 7.9% | 8.1% | 13.3% | ||
| Hispanic ethnicity | 7.2% | 6.8% | 15.7% | 9.9% | ||
| Other/Unknown | 1.4% | 0.5% | 1.0% | 0.7% | ||
| White | 80.3% | 78.5% | 65.7% | 66.9% | ||
| Women | 43.5% | 45.0% | 0.75 | 40.5% | 48.3% | 0.14 |
| Low educational attainment | 5.8% | 3.7% | 0.31 | 27.1% | 29.1% | 0.68 |
| Low annual household income | 9.9% | 7.9% | 0.47 | 15.7% | 13.9% | 0.63 |
| Current Smoker | 4.9% | 3.1% | 0.63 | 4.8% | 7.3% | 0.57 |
| Unknown | 14.8% | 14.1% | 6.2% | 5.3% | ||
| History of mechanical falls | 10.8% | 11.5% | 0.81 | 12.4% | 21.2% | 0.02 |
| Prior GI hemorrhage | 13.9% | 20.9% | 0.06 | 14.8% | 20.5% | 0.15 |
| Prior ischemic stroke | 15.7% | 12.0% | 0.30 | 9.5% | 8.0% | 0.60 |
| Diabetes mellitus | 30.5% | 31.4% | 0.84 | 41.4% | 45.0% | 0.49 |
| Coronary disease | 45.3% | 43.5% | 0.71 | 42.4% | 39.1% | 0.52 |
| Hypertension | 66.8% | 70.7% | 0.40 | 91.4% | 94.7% | 0.24 |
| Chronic heart failure | 57.4% | 53.9% | 0.48 | 48.6% | 49.7% | 0.84 |
| Dementia | 7.6% | 13.1% | 0.07 | 3.3% | 7.3% | 0.09 |
| Chronic liver disease | 4.0% | 4.2% | 0.94 | 3.3% | 5.3% | 0.36 |
| Kidney dysfunction (eGFR < 30 ml/min/1.73 m 2 ) | 21.1% | 21.5% | 0.92 | 12.4% | 13.9% | 0.67 |
| Cancer † | 22.0% | 22.5% | 0.90 | 22.9% | 23.2% | 0.94 |
| Anemia ‡ | 47.1% | 49.2% | 0.67 | 55.2% | 65.6% | 0.05 |
| Lung disease | n/a | n/a | 47.6% | 44.4% | 0.54 | |
| Transient ischemic attack | n/a | n/a | 7.1% | 6.0% | 0.66 | |
| Medications at Admission | ||||||
| Aspirin | 13.9% | 56.5% | <0.001 | 38.6% | 63.6% | <0.001 |
| Other antiplatelet therapies | 0.0% | 1.6% | 0.10 | 10.5% | 12.6% | 0.53 |
| Lipid-lowering agents | 12.1% | 6.3% | 0.04 | 63.3% | 51.0% | 0.02 |
| ACE inhibitors | 44.4% | 35.1% | 0.05 | 38.6% | 40.4% | 0.73 |
| Angiotensin II receptor blockers | 4.0% | 4.7% | 0.74 | 11.0% | 10.6% | 0.91 |
| Stroke/Bleeding Risk | ||||||
| CHADS 2 , mean (SD) | 2.6 (1.1) | 2.6 (1.2) | 0.93 | 2.6 (2.5) | 2.7 (2.5) | 0.46 |
| CHA 2 DS 2 -VASc, mean (SD) | 4.4 (1.4) | 4.5 (1.6) | 0.89 | 4.4 (1.5) | 4.6 (1.6) | 0.28 |
| ATRIA stroke risk score, mean (SD) | 7.6 (2.3) | 7.7 (2.5) | 0.67 | 7.1 (2.3) | 7.5 (2.4) | 0.07 |
| HAS-BLED, mean (SD) ∗ | 2.3 (1.0) | 2.8 (1.0) | <0.001 | 2.6 (0.9) | 3.0 (0.9) | <0.001 |
| ATRIA hemorrhage risk score, mean (SD) | 4.3 (2.4) | 4.6 (2.3) | 0.12 | 4.4 (2.0) | 5.0 (2.1) | 0.01 |
∗ The HAS-BLED score presented here has a maximum of 7 points since data on alcohol usage was not available and time in therapeutic range was not included. If adding 1 point for time in therapeutic range < 60%, there is no significant difference between those using and not using warfarin.
† All cancer excluding non-melanomatous skin cancer.
‡ Female: hemoglobin < 13.5 g/L; Male: hemoglobin < 12.0 g/L.
In the 361 patients in ATRIA-CVRN with major GI hemorrhages, 58% were taking warfarin. The distribution of anatomical sites was very similar to that of the ATRIA cohort with 176 (49%) upper GI bleeds, 99 (27%) lower GI bleeds, and 86 (24%) undetermined. Similar to those in the ATRIA cohort, patients with GI hemorrhage in the ATRIA-CVRN using warfarin were younger, less likely to be women, less likely to have a history of falls, GI hemorrhage, or dementia, and much less likely to be using aspirin. In the more recent ATRIA-CVRN cohort, the use of lipid-lowering agents was far more common and patients with GI bleed on warfarin were prescribed such agents more frequently than those not taking warfarin. Patients with GI bleed on warfarin were less likely to have preexisting anemia. As with the ATRIA experience, there was no statistically significant difference in CHADS 2 , CHA 2 DS 2 -VASc, or ATRIA stroke scores between patients using or not using warfarin at the time of their bleeding event, whereas warfarin users had lower HAS-BLED and ATRIA hemorrhage risk scores ( Table 1 ). Comparison of the characteristics of warfarin users and nonusers in each of the 5 propensity score quintiles showed similar distributions for covariates included in the propensity score model ( Supplementary Material C ). Overall, 357 (99%) patients were transfused with ≥2.0 (IQR) units of blood. In warfarin users, the median INR at presentation was 3.2 (IQR).
In the ATRIA cohort, warfarin use at the time of GI hemorrhage was not associated with 30-day mortality in unadjusted or adjusted analyses. In contrast, in ATRIA-CVRN, warfarin use at the time of GI hemorrhage was associated with a much lower 30-day mortality rate ( Table 2 ). Warfarin use appeared to be more strongly associated with lower 30-day mortality in the oldest group of patients. In the ATRIA cohort, the adjusted mRR (95% confidence interval) was 0.75 (0.28 to 2.03) for patients ≥85 years (n = 94) versus 1.16 (0.56 to 2.41) for those <85 years (n = 320). In ATRIA-CVRN, warfarin appeared to be associated with reduced mortality in all ages, but, as in ATRIA, the largest observed association again was seen in older patients (adjusted mRR [95% confidence interval] for ≥85 years [n = 62] 0.15 [0.02 to 1.27]; for <85 years [n = 299] 0.50 [0.20 to 1.23]). A substantial number of patients resumed using warfarin in the year after GI hemorrhage (ATRIA 25%; ATRIA-CVRN 52%), whereas few nonwarfarin users began using warfarin during this time period (ATRIA 1%; ATRIA-CVRN 6%). After excluding deaths and follow-up time during the first 30 days after major GI hemorrhage, warfarin use was associated with a small, but not statistically significant, reduction in mortality over 1 year of follow-up after GI hemorrhage in both ATRIA and ATRIA-CVRN ( Table 2 and Figure 1 ).
