Trends in Acute Non-Variceal Upper Gastrointestinal Bleeding in Dialysis Patients Ju-Yeh Yang, MD, MS,1,2 Tsung-Chun Lee, MD,4 Maria E. Montez-Rath, PhD,1 Jane Paik, PhD,3 Glenn M. Chertow, MD, MPH,1 Manisha Desai, PhD,3 Wolfgang C. Winkelmayer, MD, ScD1 From the Divisions of Nephrology 1 and General Medical Disciplines,3 Stanford University School of Medicine, Palo Alto, CA; Division of Nephrology and Department of Internal Medicine, Far Eastern Memorial Hospital, Taipei, Taiwan;2 Department of Internal Medicine, National Taiwan University Hospital and National Taiwan University College of Medicine, Taipei, Taiwan.4 Correspondence to: Wolfgang C. Winkelmayer, MD, ScD, Stanford University School of Medicine, Division of Nephrology, 780 Welch Rd., Suite 106, Palo Alto, CA 94304, U.S.A. Email: firstname.lastname@example.org
Word Count: 3,597 (manuscript); 250 (abstract)
Impaired kidney function is a risk factor for upper gastrointestinal bleeding (UGIB), which associates with poor outcomes. The burden of UGIB and its impact on patients with ESRD are not well described. Using data from the United States Renal Data System, we quantified the occurrence rate and associated 30-day mortality of acute, non-variceal UGIB in patients on dialysis using medical claims and previously validated algorithms where available. Overall, 948,345 patients contributed 2,296,323 patient-years for study. The occurrence rates for UGIB were 57 and 328 episodes per 1000 person-years using stringent and lenient definitions of acute, non-variceal UGIB, respectively. Unadjusted occurrence rates remained flat (stringent) or increased (lenient) from 1997 to 2008; adjustment for socio-demographics and comorbidities, however, resulted in a significant decline for both definitions (linear approximation, 2.7% and 1.5% per year, respectively; P<0.001). In more recent years, patients had higher hematocrit levels prior to UGIB episodes and were more likely to receive blood transfusions during an episode. Overall 30-day mortality was 11.8%, which declined significantly over time (relative declines of 2.3% or 2.8% per year for the stringent and lenient definitions, repsectively). In summary, despite declining trends worldwide, crude rates of acute, non-variceal UGIB among patients on dialysis have not decreased during the past 10 years. Although UGIB-related 30-day mortality declined, perhaps reflecting improvements in medical care, the burden remains substantial to the ESRD population.
More than 600,000 U.S. patients have advanced chronic kidney disease requiring maintenance dialysis or kidney transplantation. The majority of these patients are covered through Medicare’s End-Stage Renal Disease (ESRD) program, which spends more than $27 billion annually for the care of these patients.1 Previous studies have indicated that patients requiring dialysis are particularly prone to developing upper gastrointestinal bleeding with subsequent poor outcomes.2-4 In the general (non-ESRD) population, the incidence of, and mortality associated with, acute non-variceal upper gastrointestinal bleeding (ANVUGIB) has been declining over time.5-11 It is unknown, however, whether the ESRD population has experienced similar changes in rates or complications of ANVUGIB and its outcomes. One may posit that patients on hemodialysis have unique risks for upper gastrointestinal bleeding (UGIB) owing to repeated exposure to anticoagulants. Residual uremia might also render patients resistant to the benefits of certain therapeutics. Thus, the beneficial trends towards lower incidence and better outcomes of UGIB observed in the general population may not have materialized in the ESRD population
We used the national registry of ESRD patients in the U.S. to study the occurrence of ANVUGIB and its outcomes in patients undergoing dialysis. We tested the null hypotheses that the occurrence of, and mortality from, upper gastrointestinal hemorrhage over the past decade did not change among U.S. patients on dialysis.
From 1998 to 2007, 948,345 unique patients were represented in our study, contributing 2,296,323 person-years of observation. Characteristics of the study population in calendar years 1998 and 2007 are shown in Table 1. During the decade of observation, mean age, the proportion of men, median dialysis vintage, and prevalences of certain comorbid conditions increased.
Using the stringent algorithm for the ascertainment of ANVUGIB, a total of 101,561 patients experienced at least one upper gastrointestinal bleeding event during observed period, with a total of 131,022 episodes (57 per 1000 person-years). Using the more lenient algorithm, 380,343 patients experienced 753,508 events (328 per 1000 person-years). Trends in occurrence rates over time are shown in Figure 1. The proportion of hospitalized episodes remained unchanged around 85-90% using either definition (Figure 2). The proportion of bleeding episodes related to PUD declined from 38.7% to 31.2% for the stringent definition and from 19.3% to 11.1% for the lenient definition of ANVUGIB (Figure 2). From unadjusted analyses, the occurrence rate of ANVUGIB using the stringent criteria was essentially unchanged from 1998 to 2007 (decline per year: 0.2%; 95% CI: 0% - 0.4%; P=0.12; Table 2). However, after adjusting for age, sex, race, dialysis vintage, modality, previous kidney transplant, history of ANVUGIB, Medicaid coverage, and comorbidities, the occurrence of ANVUGIB declined by 2.7% per year (95% CI: 2.5% to 2.9%; P<0.001). Using the lenient criteria , the adjusted rate of ANVUGIB declined by 1.5% (95% CI: 1.4% to 1.6%) per year. Overall 30-day mortality rates after ANVUGIB were 11.8% (stringent criterion) and 9.6% (lenient criterion), respectively. Mortality was slightly higher for first episodes: 12.5% and 10.0% for stringent and lenient criterion episodes, respectively. Mortality declined from 12.3% in 1998 to 10.5% in 2007 (stringent criterion) and from 10.2% in 1998 to 8.1% in 2007 (lenient criterion; Figure 4). Formal tests for trend showed that the adjusted 30-day mortality odds from ANVUGIB decreased by 3.3% per year (relative change; 95% CI: 2.8% to 3.8%) for the stringent criterion and by 3.6% (95% CI: 3.4% to 3.8%) for the lenient criterion, respectively (Table 2 and Figure 5).
To explore correlates of declining mortality trends, we studied patterns of hematocrit values prior to ANVUGIB events as well as use of blood transfusions during these episodes. Mean hematocrit and the proportion of patients receiving blood transfusions increased significantly over time (Figure 6).
From a comprehensive database detailing the experience of U.S. dialysis patients over a decade, we found occurrence rates of ANVUGIB almost two orders of magnitude higher than in the general population, where rates between 50 and 150 events per 100,000 person-years had been recently reported.7-11 In contrast to the general population, where occurrence of ANVUGIB and associated mortality have declined over the past decade, we observed an increase in occurrence rates when adjusted for demographic factors and comorbidities. Of note, 30-day mortality declined during the same period, consistent with reductions in non-ESRD populations.6-7, 10
Relatively little is known about the occurrence and outcomes of UGIB in patients undergoing dialysis. In an analysis of the USRDS Dialysis Morbidity and Mortality Study (DMMS), Wasse et al. reported a rate of 23 UGIB events per 1000 person years, less than half of our conservative estimate.12 Of note, the DMMS was oversampled for patients onperitoneal dialysis and included person-time spent after kidney transplantation. Patients on peritoneal dialysis and kidney transplant recipients had lower risk of UGIB than patients on hemodialysis.12 Differences in reported rates need to be considered in light of different approaches to ascertaining and studying UGIB.
For example, some studies included only bleeding episodes that were deemed to be related to peptic ulcer disease (PUD). In the general population, only 40-60% of all UGIB are due to PUD;10-20% are attributed to variceal bleeding in the general population.7-10 The proportion of PUD-related and variceal bleeding episodes were reported to be approximately 60% and 10%, respectively, in a cross-sectional survey of 727 patients with CKD.13 Since we were interested in the overall impact of ANVUGIB in patients on dialysis, we used a more extensive selection of ICD-9 diagnosis codes to identify all ANVUGIB episodes, rather than including only those related to PUD. In our cohort, 35.2% of ANVUGIB episodes appeared PUD-related.
Most previous published studies only considered the first episode and/or hospitalized cases with primary discharge diagnosis indicating UGIB. However, UGIB are acute events which can be fully resolved and later recur. Previous studies have indicated that patients on dialysis are at particular risk of recurrence of UGIB. While UGIB frequently complicates patients already hospitalized, the management of episodes that begin in the ambulatory setting does not necessarily mandate admission. Almost 40% of UGIB in Medicare patients were managed in the outpatient setting 14 and 20% of these UGIB recurred.15 In our cohort, 28.3% of episodes were recurrent and approximately 10% episodes were managed in the outpatient setting. We would have missed many events had we not meticulously investigated outpatient claims and identified recurrent episodes.
Another cause of such high UGIB occurrence rates might be explained by the special characteristics of dialysis population itself. The dialysis population is enriched by individuals exhibiting several well-known risk factors of ANVUGIB: older age, anticoagulant use, platelet dysfunction, cardiovascular disease and disability.12 Further, patients on dialysis are exposed to heightened medical surveillance due to their regular interactions with health care providers and routine laboratory surveillance, where management of anemia has been an important focus of care related to quality metrics and reimbursement.16 In response to the ESRD Prospective Payment System (the “bundle”) instituted in January 2011, it is likely that vigilance for gastrointestinal bleed will remain, as there is an 8% higher reimbursement for dialysis services in patients carrying a diagnosis of gastrointestinal bleed.
While UGIB has declined over the past decades worldwide,7-11 we found consistently high occurrence rates of ANVUGIB in the dialysis population from 1998 to 2007. The most probable explanation for the non-decline in occurrence rates is the changing character of the dialysis population.
Advanced age, male sex, and certain medications such as antiplatelet or anticoagulant agents are well-known risk factors of ANVUGIB.17-19 A subtle but important change in the dialysis population is the shift away from peritoneal and toward hemodialysis, where the latter obligates regular, intermittent use of anticoagulants. Only after adjusting for multiple comorbid conditions that increase the risk of ANVUGIB were the adjusted rates found to be declining, albeit slightly. (Figure 2)
The prevalence rates of several comorbidities, including diabetes, hypertension, heart failure, arrhythmia, valvular heart disease, peripheral vascular disease, chronic obstructive pulmonary disease and past UGIB history, all increased over time (Table 1). It is important to note, however, that more aggressive coding practices may have contributed to these apparent increases,20 a type of information bias. Thus, adjustment for these comorbidities may have actually led to overestimation of any true reductions in ANVUGIB occurrence over time.
The decline of UGIB in general population is mainly due to decline of PUD-related bleeding, especially in younger populations.7-8 In the past decades, there have been many advances in prevention and treatment of PUD, such as introduction of H2-receptor antagonists, proton pump inhibitors, and eradication regimens for Helicobacter pylori. Whether these treatments are as effective in ESRD as in the general population for PUD-related or non-PUD-related ANVUGIB, is unknown.21-25
Mortality associated with UGIB ranged between 5-15% in the general population, dependent on the specific population studied, specific definition of UGIB and definitions of mortality (in-hospital vs. all). The 30-day mortality rate in our report was within the same range, although mortality following outpatient ANVUGIB episodes was much lower than those among hospitalized cases (7.3% vs. 13.6%), again comparable to the Medicare population.14 Earlier studies reported no change in UGIB-related mortality rate over time, probably due to the co-occurring trends in age and comorbidities.8-9, 11 However, similar to our findings in patients on dialysis, recent reports demonstrated an improvement of outcomes even though UGIB patients became more vulnerable over time in terms of age and comorbidity.6-7, 10 The declining trends of ANVUGIB-related 30-day mortality rates were demonstrated both in PUD-related and non-PUD-related bleeding episodes.
Despite an increasingly sicker dialysis population, the ANVUGIB-associated 30-day mortality rate was found to have declined by 2-3% per year, which may reflect success of one or more therapeutic strategies or improvements in the general care for this population. There were at least two obvious changes during past decade – 1) the evolution of endoscopic hemostatic techniques and 2) a trend toward higher hemoglobin concentrations. Combined endoscopic modalities for hemostasis were shown to improve bleeding outcomes.26 Patients on dialysis might particularly benefit from these advancements since they were viewed as a high risk group for surgery. With effective nonsurgical therapies, these patients may have better chances to survive from even massive ANVUGIB without the need for operative intervention. Similarly, more aggressive treatment with ESAs and intravenous iron and subsequently, higher hemoglobin targets, have contributed to an increased “hematocrit reserve”, which might also have translated into the observed better outcomes (i.e., providing a larger “margin of error” in the face of ANVUGIB); more liberal use of transfusion and increasing use of proton pump inhibitors may have contributed to the decline in case fatality. Finally, it is possible that more subtle cases of ANVUGIB may have been noticed and coded in more recent years (also known as “code creep”), although the proportion of hospitalized episodes having remained stable at approximately 90% during the study period argues against this possibility.
To the best of our knowledge, this study is the most comprehensive analysis of ANVUGIB in the dialysis population to date. We applied both stringent and lenient algorithms to approximate the upper and lower ranges of ANVUGIB occurrence rates and took outpatient, inpatient, and recurrent episodes into consideration. We ascertained multiple covariates in addition to basic demographic data and primary outcomes for adjustment, including comorbidities and transfusion events.
Despite these strengths, our study also carries certain limitations, mostly from using administrative billing claims for our research. The accuracy of the coding for our primary outcome is the most prominent concern. The algorithm we used to identify ANVUGIB events came from a validated list of ICD-9 diagnosis codes proposed by Cooper et al.27 and modified by Targowinik et al.11 Cooper et al. reported good sensitivity and positive predictive values (85-95%) of diagnosis coding for source of hemorrhage and upper endoscopy among 882 patients with upper GI hemorrhage. Targowinik et al. modified the algorithm by adding nonspecific GIB diagnostic codes and procedure codes to exclude variceal bledding, lower GI bleeding or chronic GI bleeding in an attempt to increase sensitivity and positive predictive value. Since the latter codes did not cover most non-ulcer related GIB, we augmented the algorithm by adding several lesion specific diagnostic codes, as detailed above. For the stringent algorithm, 100% of patients had endoscopy with three days of event date, which again indicates high reliability of this algorithm; 40% of patients had endoscopy within three days of an assumed event date.
Another inherent limitation from administrative claims is the absence of other clinically relevant data including most laboratory results, biometric values, endoscopic findings, or detailed medications, which are all unavailable. Finally, there were some missing data for certain variables which we imputed using standard methods,29-30 which assume that missingness occurred at random.
Finally, the primary goal of this study was to describe secular trends in the occurrence and outcomes of ANVUGIB, which guided us in the specific set up of our cohort. We did not focus on studying risk factors of ANVUGIB, as this would have required a different cohort assembly (inception cohort) and analytical approach.
In summary, we found an exceptionally high occurrence rate of ANVUGIB among U.S. dialysis patients, at least an order of magnitude higher than in the general population. In contrast to trends in the general population towards lower rates of these events, occurrence rates among patients on dialysis did not change materially between 1998 and 2007. While 30-day mortality of ANVUGIB declined from 12.3% to 10.5% in the recent decade, further investigation is needed to identify care pathways that may serve to reduce the burden of ANVUGIB in the dialysis population.
We used the United States Renal Data System (USRDS) database for this study, with data available from 1996-2008. We obtained information on the outcome of interest, ANVUGIB, as well as presence of several comorbidities from Medicare claims files. We obtained demographic data and selected clinical information from USRDS Standard Analysis Files, including the Medical Evidence Report (CMS-2728). To use clinician-assigned codes as often as possible, we used International Classification of Diseases, 9th revision (ICD-9-CM) codes; we excluded from consideration diagnostic codes from Part B clinical laboratory, diagnostic imaging and durable medical equipment claims.31
We included all patients eligible for and covered by Medicare who underwent maintenance dialysis at some point between January 1, 1998 and December 31, 2007. Patients under the age of 18 years or whose kidney function recovered were excluded. To ensure complete ascertainment of events of interest, we restricted all analyses to those periods during which a patient was covered by Medicare (Part A and B) as primary payer of medical expenses. Patients over 65 were eligible for Medicare benefits on the basis of their age. For patients younger than 65, primary Medicare coverage became active 36 months after the start of dialysis.
We modified the methods proposed by Targownik et al. 11 to define ANVUGIB events using two criteria: a specific or stringent criterion and a more inclusive or lenient criterion. Using the stringent criterion we defined ANVUGIB events as those corresponding to a group of ICD-9 diagnostic codes, which specify the identified cause of the bleeding, e.g. the ICD-9 codes for gastroesophageal laceration-hemorrhage (530.7) or for gastric ulcer with hemorrhage (531.0). Using the lenient criterion, we defined ANVUGIB events by additionally including incidents with corresponding diagnosis codes that did not specify the lesions responsible for the bleeding but coincided with bleeding-related medical procedure codes. Additionally for both definitions we included several important lesion-specific events not considered by Targownik et al.  (e.g., 537.83 for angiodysplasia of stomach and duodenum with hemorrhage; 537.84 for Dieulafoy lesion (hemorrhagic) of stomach and duodenum). This modification was based on a systematic review of all potentially relevant ICD-9 codes and consultation with an experienced gastroenterologist (TCL). The stringent algorithms and codes used are listed in Appendix 1. We excluded all claims with diagnosis codes of “variceal bleeding” on the bleeding date in our algorithm.
Given trends toward outpatient management of a variety of health conditions, including gastrointestinal disorders over the past decade,14 we made use of outpatient claims information when ascertaining ANVUGIB events. Such data are important to investigate recurrent events of interest, which may have been less severe than a first event and would be missed if we had considered only events requiring hospitalization. On the one hand, one needs to be careful when defining an event from outpatient claims only, as a single diagnostic code might not reliably represent a true event. On the other hand, a single true event could be represented by multiple reimbursement claims that spanned several days. The former error would reduce specificity of event ascertainment and the latter would lead to overestimation of true event rates. To reduce these potential errors, we defined a single ANVUGIB event from outpatient claims based on the presence of two outpatient claims within a 7-day time period. Alternatively, one outpatient claim was considered sufficient if it was accompanied by a claim for esophagogastroduodenoscopy performed on the same date.14 Further, if the onset dates of two events were within a 30-day period, these two events were counted as a single episode to avoid overestimating the occurrence rate of ANVUGIB.14
The second outcome of interest, which only applies to those who experienced an ANVUGIB event, is short term (30-day) mortality following an episode of ANVUGIB, defined as having died within 30 days of the ANVUGIB diagnosis date regardless of discharge status.
From the USRDS, we ascertained basic demographic variables, such as age, gender, race, and primary kidney disease causing ESRD. Dialysis vintage, dialysis modality, transplant status, and Medicaid eligibility as a representation of lower socioeconomic status were defined as time-dependent variables. We systematically surveyed ICD-9-CM diagnostic and procedure codes and referred to previous studies to derive a comprehensive table of codes for relevant co-morbid conditions (Appendix 2).32-35 The initial status of co-morbidity was based on information from the Medical Evidence Report forms and then updated according to claims data. Corresponding codes had to appear at least twice in outpatient claims or be present in any inpatient claims to be counted.31 For all comorbidities and past histories of gastrointestinal bleed, we searched claims back to 1.1.1996, i.e., two years before the start date of the observation period. We matched the date of gastrointestinal bleed events to institutional claims to obtain the most recently recorded hematocrit value, Transfusion events were identified via ICD-9 procedure codes (99.03, 99.04) and HCPCS codes (P9010, P9011, P9016, P9021, P9022, P9038, P9039, P9040, and 36430) from claims data.36
To address whether the occurrence of ANVUGIB events changed over time, we used Poisson regression to model the number of ANVUGIB episodes each year with adjustment for the length of time observed via an offset term. Logistic regression models were used to evaluate changes in 30-day mortality following ANVUGIB events over time. For each outcome of interest, we fitted three models that adjusted for different sets of covariates. In Model 1, we adjusted for age, gender and race. Model 2 included covariates in Model 1 and additionally adjusted for Medicaid coverage, dialysis vintage, and modality. Model 3 (full model) included covariates in Model 2 and additionally adjusted for history of kidney transplantation, prior ANVUGIB, all comorbidities (Appendix 2), alcohol dependence, tobacco use, drug dependence, inability to transfer, inability to ambulate, and baseline Quetélet’s (body mass) index (BMI) for analysis of ANVUGIB events. For mortality analyses, two additional event-related variables were included: whether the episode was peptic ulcer disease-related and whether it involved a hospitalization or not. The observation of a single patient may have been interrupted by ineligible payor histories (with Medicare Part A and B not being the primary payor), dialysis modality switches, transplantation, or calendar years. This means that a patient can have more than one observation per year and several observations between years. As each subject may contribute more than one outcome, we expect correlation among observations within subject. The generalized estimating equations (GEE) method with an exchangeable working correlation structure was applied to adjust for the expected dependence of observations.37 Validity of results using this method, however, relies on an assumption that the data are missing completely at random (MCAR).38
The handling of missing data
Some of the key covariates were missing for a proportion of the subjects, where missingness ranged from less than 1% (Race) to 16% (BMI). Examination of the patterns of missingness revealed the data were not MCAR. We therefore utilized multiple imputation techniques 29-30 to handle the missing data. More specifically, standard multiple imputation techniques were used to yield 10 imputed data sets to which we subsequently fit the GEE Poisson regression (or GEE logistic regression depending on the outcome), and applied the combination rules described by Little and Rubin 39 to provide parameter estimates that address our questions of interest. Multiple imputation relies on the assumption that the data are missing at random (MAR) or that missingness is related to observed variables only. This assumption is reasonable in our context, but would be violated if missingness were related to the missing values themselves (for example, if we were more likely to be missing BMI for those with higher or lower BMI values). One challenge posed by implementing the imputation related to the dependence of observations. While methods exist for imputation in the presence of correlated data, it can be computationally demanding to format a data set with a large number of time-varying covariates as in our case. We therefore assumed independence of observations when performing multiple imputation, where the imputation model included all variables used in the scientific model of interest. All analyses were performed using the SAS 9.2 software (www.SAS.com).This work was approved by the Institutional Review Board of Stanford University School of Medicine.