Why do mortality rates for nonvariceal upper gastrointestinal bleeding differ around the world ? A systematic review of cohort studies

1Translational Gastroenterology Unit, John Radcliffe Hospital; 2NHS Blood & Transplant, John Radcliffe Hospital, Oxford, United Kingdom; 3Division of Gastroenterology, McGill University Health Centre, Montreal, Quebec; 4Division of Epidemiology and Public Health & Nottingham Digestive Disease Centre, United Kingdom; 5Division of Clinical Epidemiology, McGill University Health Centre, Montreal, Quebec Correspondence: Dr Vipul Jairath, Translational Gastroenterology Unit and NHS Blood and Transplant, John Radcliffe Hospital, Headley Way, Oxford, United Kingdom. Telephone 44-1865-387906, fax 44-1865-387957, e-mail vipul.jairath@nhsbt.nhs.uk Received for publication September 20, 2011. Accepted December 20, 2011 Nonvariceal upper gastrointestinal bleeding (NVUGIB) is a common problem worldwide, with reports of its incidence ranging from 48 to 160 cases per 100,000 adults per year (1-3). NVUGIB is associated with considerable morbidity, mortality and economic impact, despite significant advances in its management over the past two decades. In the past two years, several publications have focused on the epidemiology and changing time trends of acute NVUGIB. Most of these studies have reported a reduction in its incidence and associated case fatality (henceforth referred to as ‘mortality’) rate (4-8), although some have reported no changes in mortality (9-11). The reductions in mortality observed over time have been largely attributed to combinations of therapeutic endoscopy, proton pump inhibitors, eradication of Helicobacter pylori, preventive strategies in individuals taking nonsteroidal anti-inflammatory drugs and advances in critical care. Because most patients do not die from uncontrolled bleeding (12,13), further improvements in outcomes will be challenging in the face of an aging population with the associated burden of comorbidities. Several recent studies and reviews have served to highlight variations in mortality rates reported for NVUGIB, ranging from 3% to 12%, as well as highlighting substantial differences in mortality among countries (9,14-18). The studies reporting these outcomes range from prospectively collected national registries for NVUGIB to retrospective analyses of large administrative databases. Such marked differences in outcomes are surprising because most of these studies originate from health care systems with access to modern-day, evidence-based standards and processes of care for managing NVUGIB. A range of factors can influence the survival of patients with NVUGIB including those related to patients, disease severity and health care systems. Comparisons among countries are fraught with difficulty due to the complexity and intricacy of factors influencing revieW


Methods search strategy
The MEDLINE, EMBASE and ISI Web of Knowledge databases were searched for nonrandomized studies reporting mortality rates following presentation with NVUGIB.Search terms were combined with a highly sensitive observational study filter and restricted to include adult, human studies in English spanning 1996 to January 2011 (search terms can be provided on request).Reports preceding this period were not sought because of the significant evolution of general supportive care, and advances in endoscopic and pharmacological therapies that occurred during that time period (19); therefore, it was hypothesized that the chosen time period would represent a more homogenous group of data to compare.The bibliographies of identified articles were also screened.

eligibility criteria
Cohort studies reporting mortality outcomes were searched for unselected patients presenting to hospital with acute NVUGIB.Only studies that reported outcomes for more than 1000 patients were included.For studies including both NVUGIB and acute variceal hemorrhage (AVH), only studies examining more than 1000 patients with NVUGIB were included.Studies published only as abstracts were excluded because they did not allow an adequate assessment of methodology.Citations identifed were screened in duplicate, with a kappa statistic of 0.65 (95% CI 0.60 to 0.70).Following the development of a data abstraction sheet, results were extracted by two reviewers (VJ, MM).Any disagreements were resolved by discussion and did not necessitate arbitration by a third person.Recorded information from each study included year of publication, study design, country of origin, nature of database analyzed, whether examining all-cause NVUGIB or a selected subgroup (eg, peptic ulcer bleeding only), age, new admission and/or inpatient bleeds, length of follow-up and mortality rate.Also assessed was the methodological quality of the study report, using criteria for the reporting of nonrandomized studies from the Cochrane Collaboration (20) and Strengthening The Reporting of Observational Studies in Epidemiology (STROBE) criteria for reporting of cohort studies (21).

data analysis
Because the main aim of the present study was to examine differences in mortality rates and to attempt to identify the sources responsible for the variations observed, and because of the heterogeneity of the studies, only descriptive individual study data, without an attempt at formal meta-analysis, are provided.Indeed, the underlying assumption was that the mortality rate attributable to NVUGIB was not uniform, but rather, that it is dependent on a number of factors that are raised and outlined.

What are the reported mortality rates following NVuGiB?
The studies originated from Europe, Asia and North America (Table 1).The sources of data ranged from prospective cohort registries for peptic ulcer bleeding from single centres of expertise (12) to large retrospective cohorts from nationwide administrative data sets (14).Ten studies used administrative databases and the remaining eight used registries specifically designed to assess 'real-life' outcomes following NVUGIB, resulting in a mixture of retrospective and prospective cohort studies.The mortality rates reported in these studies ranged from 1.1% in Japan to 11% in Denmark.Not only was there variation in reported mortality among countries, there was notable variation reported within countries (Table 1).
Why might reported mortality rates differ?source of case ascertainment: Definitions of NVUGIB should be clear and consistent throughout studies to enable comparability.Completeness of case ascertainment and degree of population coverage will influence reported mortality.Three studies from Italy in the present review conducted over the same time period reported mortalities ranging from 2.6% to 6.9% (17,28,31).One study used administrative data (28) and the other two used registry data (17,31).All three studies examined NVUGIB but definitions varied: two studies were reliant on retrospectively collected data coded by International Classification of Diseases, Ninth Revision (ICD-9)-CM (28,31) discharge diagnoses including variations in the codes selected within each study, whereas the third study was a prospectively collated registry requiring confirmation of upper gastrointestinal tract bleeding by clinical examination by admitting medical/nursing personnel (17).Administrative databases, not designed for research, have the advantage of size and population-based longitudinal follow up, but accuracy of case ascertainment is dependent on coding.In addition, they are often unable to adjust for key disease-specific prognostic variables.Positive predictive values (PPVs) of ICD-9-CM coding for peptic ulcers and gastrointestinal bleeding have been shown to be reasonably accurate, with PPVs ranging from 85% to 95% (34), but differ depending on the terminology used (35)(36)(37).The terms with the greatest PPVs (>90%) were for site-specific codes (eg, 'gastric' or 'duodenal' ulcer), but these were lower for lesion-specific codes (ie, 'peptic ulcer') and even lower for nonspecific terms such as 'melena' or 'hematemesis' (34).In addition, coding errors have been demonstrated in up to 7% of cases (34), which may result in thousands of miscoded cases in large datasets.This was fittingly highlighted in an article by Ahsberg et al (18), who stated "we found a sudden increase in hospitalisations for ulcer UNS (unspecified) after 1997, and at the same time-point, there was a large drop in hospitalisations for bleeding and perforated ulcers.These changes  Careful review of selection criteria, patient sampling and length of follow-up also demonstrates subtle but important differences that may influence reported outcomes.For example, exploratory data have suggested differing health care delivery patterns between outpatient bleeds and cases in whom hemorrhage started while already hospitalized for an unrelated condition (38).Therefore, inclusion of inpatient bleeds will influence reported mortality rates considering their markedly poorer prognosis (38), with some series reporting up to a fourfold increase in mortality for inpatients developing bleeding compared with new admissions (39).One of the three Italian studies included both new admissions and existing inpatients who developed upper gastrointestinal tract bleeding (17), whereas it was unclear in the other two studies whether inpatient bleeds were included in addition to new admissions (28,31).Three studies from the United States were included, with the study by Cooper et al (40) reporting almost double the mortality rate compared with the other two studies (14,22).There were subtle differences in all three studies: all used administrative data-sets and ICD-9-CM coding to identify cases; Cooper et al (40) used Medicare claims data whereas the other two studies used data from the Nationwide Inpatient Sample (NIS) (9,21).The study by Cooper reported 30-day mortality, in contrast to inhospital mortality reported by the other two studies.Unfortunately, without direct and uniform statistical adjustment for confounders among studies, one can only speculate whether such disparities can account for the twofold differences in reported mortality within countries.Patient-related factors: Age and presence of comorbid illnesses are highly influential prognostic factors following presentation with NVUGIB, and critical components of risk assessment scores.Age is a component of the Rockall score (41), and the presence of comorbid illnesses are key components of both the Rockall and Blatchford scores (42), both extensively validated prognostic scores following NVUGIB in a variety of global patient populations and clinically meaningful to the practicing clinician.Although we are not advocating a particular score, although one should be used routinely in clinical practice, there is increasing evidence that the Blatchford score is more useful than the Rockall score in predicting the need for endoscopic therapy, transfusion and identifying patients at low risk for discharge, but not superior to the complete Rockall score in predicting mortality (43,44).
For valid comparisons of mortality rates, risk adjustment according to such factors is essential because these important prognosticators of outcome are unevenly distributed across providers and variation in baseline status could make a major contribution to observed differences in mortality rates.Risk adjustment is a complex construct that can involve patient's sociodemographic factors (eg, age, sex and race), acute clinical stability, severity of primary disease, functional status and burden of comorbidity (45).As highlighted in Table 2, there was remarkable similarity in the central measures of age in the reviewed studies, ranging from 64 to 74 years.However, there was wide variation in the characterization of comorbid illnesses, with specific NVUGIB registries more likely to detail individual comorbid illnesses, whereas administrative data sets were more likely to present composite scores such as the Charslon index ( 46) -a measure that is less clinically meaningful to the practicing clinician and not well validated as a component of risk-adjustment scores for NVUGIB.The Charlson comorbidity index was originally designed as a measure of the risk of one-year mortality attributable to comorbidity in a longitudinal study of generalized hospitalized patients, subsequently adapted so that ICD-9 codes could be used to calculate the index with existing administrative data.The presenting international normalized ratio (INR) has also been suggested as a useful predictor of mortality and may be a more objectively measurable proxy of comorbidity (47).A more widespread use of validated prognostic scales (and their subsequent reporting in observational studies) has been recommended (48,49) but remains very low (<2% in a recent national study [50]).It is, therefore, disappointing -but not surprising -that only two studies presented the Rockall score (16,31) and none presented the Blatchford score.Given this wide variation in characterization of comorbidities, it is difficult to make meaningful adjustments of baseline risk, hampering any ability to compare mortality rates among these studies or riskadjust the reported information accordingly.Ethnicity and genetic factors may also play a role.Two studies involved Asian populations while the remainder assessed western populations (Table 1).Differences among these ethnic groups may affect outcomes based on varying rates of proton pump inhibitor metabolism, H pylori prevalence, gastric parietal cell mass and disease acuteness, which may further limit direct comparisons of mortality (49).disease-related factors: Adequate risk adjustment to facilitate comparison of mortality also requires disease-specific risk adjustment.For NVUGIB, this would include the ability to adjust for features of hemodynamic shock, presenting hemoglobin/biochemical parameters, endoscopic diagnosis and stigmata of bleeding, including stigmata, that would be predictive of rebleeding.Only four of 18 studies presented all of these features, with 10 of 18 presenting none; administrative data sets were less likely to present this information.The inability to characterize disease acuity is a major limitation of many of the studies we reviewed, which further limits any direct comparison of outcomes among studies.health care system-related factors: It is, of course, probable that there are important differences in standards and procedures of care for NVUGIB that may account for some of the differences reported in mortality rates.These include the extent of access to emergency and resuscitative care, to timely endoscopy and available operator skills in hemostatic procedures and, in some cases, availability of surgery or radiological procedures, all of which may influence case fatality.Recent data from the United Kingdom (UK) has highlighted deficiencies in the provision of timely endoscopy for NVUGIB, with only 50% of hospitals having formal provision of a 24 h endoscopy service and 42% of high-risk patients (Rockall score ≥5) waiting more than 24 h for their index endoscopy (16).This is in contrast to other national registries in which at least 70% of endoscopies were performed within 24 h of admission (1, 15,17,51).The actual national nature and representativeness of contributing sites may explain at least some of this variation.
The timeframe for performance of endoscopy may also play a role.Although there is some randomized evidence suggesting that timely endoscopy reduces transfusion requirements and duration of hospital stay (52), only more recent observational data have suggested an associated decrease in the need for surgery (24), although its impact on mortality remains uncertain and is an area for further research.The large, recently published UK audit barely failed to show an improvement in mortality associated with the availability of an after-hours endoscopy service (16).Several recent publications have also demonstrated increased mortality for patients admitted with NVUGIB on the weekend compared with a weekday (14,22,27), as has been reported for other acute medical conditions (53)(54)(55), which may have implications for processes and models of care in those countries, although other countries (France, Hong Kong, UK) have not reported such associations for NVUGIB (56)(57)(58).

Were studies reported well?
Clear reporting is necessary to enable an accurate assessment of the strengths/weaknesses in study design, conduct and analysis.The quality of reporting of randomized controlled trials has improved several years after the endorsement of the Consolidated Standards of Reporting Trials (CONSORT) statement (59,60) by researchers and journals.However, in observational research, important information is often missing or unclear (21), which has prompted the development of the STROBE consensus statements with the aim of helping authors and journals improve methods and quality of reporting of observational studies (21).In the present review, there were notable omissions when assessing quality of reporting (Table 3).The proportion of inversus outpatient onset of bleeding was unclear in 10 of 18 studies, and only one report provided a justification of sample size.Overall, 15 of 18 studies provided no information on missing data and how these were handled.Only four studies commented on processes for source data verification, while eight made no comments about the external validity of the reported data.Given the relatively large number of studies examining outcomes for NVUGIB, the development of a framework, similar to that recently developed to aid design and conduct of randomized controlled trials in NVUGIB (61), would help standardize reporting and enable more accurate comparisons of international data.

FutuRe diReCtioNs improving reporting
Mortality following NVUGIB from most data sources remains substantial and is influenced by a multitude of patient-, disease-and health care system-related factors (Table 4).The present systematic review of available data on NVUGIB mortality suggests that investigators should strive to provide detailed descriptions of prognostic patient factors, process of care and methodological information to better inform any attempts at national comparisons of outcomes in NVUGIB.More uniform and standardized reporting is needed to facilitate baseline risk adjustment.In this respect, it would be helpful for studies to present, where possible, validated and clinically meaningful risk scores such as the Rockall (41) or Glasgow Blatchford scale (62), which incorporate both background comorbidities and acute disease acuity.For administrative data, validation of composite indexes, such as Charlson (46) and Elixhauser (63) in the setting of NVUGIB, would be helpful.As a minimum, the criteria presented in Table 3 would aid clarity and quality of reporting.

improving outcomes
Reducing mortality further will be challenging in the face of an aging population and may depend on improved adherence to existing international recommendations (48) (which are currently poorly adhered to [64]), perhaps coupled to the emergence of important data on optimal resuscitation with blood components (65,66), and promising novel diagnostic (67,68) and therapeutic technologies (69).

Table 1 Studies detailed according to preselected reporting characteristics that may influence observed mortality rates author (reference), year; year(s) of patient admission Country Patient population Design Patients included in analysis, n New admissions only and/or inpatient bleeds length of follow-up Mortality rate
(12,17,26)uction of [the] ICD-10 classification sys-tem….".Careful attention to selection of well-validated coding is key to enable accurate interpretation of outcome data from such administrative databases.Use of specific registries for NVUGIB, although labour and cost intensive, enable rigorous selection of cases, usually by medical and nursing personnel, ideally prospective follow-up, adjustment for key disease-specific prognostic variables and auditing of case ascertainment(12,17,26), although they can be limited by size and inadequate population coverage.Admission *Sha Tin, New Territories, Hong Kong.AVH Acute variceal hemorrhage; GI Gastrointestinal; ICD International Classification of Diseases; NVUGIH Nonvariceal upper gastrointestinal hemorrhage; PCS Prospective cohort study; PUB Peptic ulcer bleeding; RCS Retrospective cohort study; UGIB Upper gastrointestinal bleeding; UGIH Upper gastrointestinal hemorrhage; UK United Kingdom; yrs Years of age coincided with