A Clinical Risk Prediction Model for Barrett Esophagus

Esophageal cancer is the sixth leading cause of cancer-related mortality worldwide; in 2008, an estimated 482,000 new cases and 407,000 cancer-associated deaths were predicted (1). In the United Kingdom, the United States, and most other Western countries, esophageal adenocarcinoma is the predominant histologic subtype of esophageal cancer (2). From 1975 to 2001, the rate of increase in incidence of esophageal adenocarcinoma was among the highest for any cancer in the Western countries, with incidence increasing by a factor of 6 in the United States (3). In 2007, the United Kingdom reported the world's highest rate of esophageal adenocarcinoma incidence (approximately 5/100,000; ref. 4). Survival for esophageal adenocarcinoma is poor, as most patients with esophageal adenocarcinoma present with metastatic disease and less than 20% survive for 5 years (5).

Almost all cases of esophageal adenocarcinoma are thought to arise from underlying Barrett esophagus, progressing through low- to high-grade dysplasia (6); however the risk of progressing from Barrett esophagus to esophageal adenocarcinoma is unclear. Initial studies suggested that the absolute rate of progression was between 5 and 10 per 1,000 person-years (7); however, a recent study in Denmark reported a much lower estimate of 1.2 per 1,000 person-years (8). It is widely accepted that gastroesophageal reflux is the principal underlying cause of Barrett esophagus (9); however, a number of large population-based studies have identified other factors associated with Barrett esophagus, including central obesity (10, 11) and smoking (12).

The identification of patients with Barrett esophagus holds the greatest potential, at least in theory, for reducing mortality from esophageal adenocarcinoma. The challenge remains, however, to determine which patients presenting with gastroesophageal reflux symptoms are more likely to have underlying, undiagnosed Barrett esophagus based on their history and symptoms. Current guidelines from both the British Society of Gastroenterology (13) and the American Gastroenterological Association (14) do not recommend endoscopic Barrett esophagus screening of the general gastroesophageal reflux population. The American Gastroenterological Association do recommend endoscopic Barrett esophagus screening for those with multiple risk factors for esophageal adenocarcinoma (including age 50 years or older, male sex, white race, chronic gastroesophageal reflux, hiatal hernia, and elevated body mass index). However, these guidelines have been developed on the basis of expert opinion and weak evidence and have not been validated (14, 15).

There is growing interest in developing risk prediction models to assist practitioners in identifying patients who may benefit most from investigations and interventions. Models have been developed and used for a variety of conditions ranging from cardiovascular diseases (16) to cancers of the breast (17–20), lung (21–23), prostate (24) and bladder (25), and to melanoma (26). These models have been shown to be clinically useful and reliable at the population level, providing a cost-effective approach to disease prevention and treatment. While models have been developed to predict Barrett esophagus in symptomatic patients, they have focused on only a restricted list of factors (e.g., age, sex, and upper gastrointestinal symptoms) and have not considered phenotypic and environmental factors which may explain a high proportion of Barrett esophagus cases not attributable to gastroesophageal reflux (27, 28). If a comprehensive risk model could be developed, it could have the potential to influence clinical decision-making in the care of patients with gastroesophageal reflux.

We aimed to develop and validate a risk prediction model for quantifying the probability of Barrett esophagus in patients with frequent gastroesophageal reflux symptoms and to show the potential use of this model in the clinical setting for selection of patients for endoscopic screening for the presence of Barrett esophagus.

To derive the prediction model, we used data from participants in the Study of Digestive Health, a population-based study of Barrett esophagus and reflux-related conditions conducted in Brisbane, Australia. The study population has been described in detail previously (12). We then validated the model in a separate case–control study of Barrett esophagus conducted in western Washington State (29).

In brief, eligible cases for the Study of Digestive Health were residents of metropolitan Brisbane aged 18 to 79 years with a new (incident) histologically confirmed diagnosis of Barrett esophagus (for nondysplastic Barrett esophagus cases) or dysplasia (for dysplastic Barrett esophagus cases) between 1 February, 2003 and 30 June, 2006. Barrett esophagus was defined as the presence of specialized intestinal metaplasia (that is, columnar epithelium with goblet cells) in an esophageal biopsy taken from the tubular esophagus by upper gastrointestinal endoscopy, irrespective of the length of involvement. Patients with Barrett esophagus were prospectively identified from the private and public pathology laboratories servicing metropolitan Brisbane. Of 1,714 patients with presumptive Barrett esophagus, we gained permission to contact 1,096 (64% response rate). Of these, 614 were ineligible (487 were “prevalent cases” with a previous Barrett esophagus diagnosis, 86 had only intestinal metaplasia of the gastroesophageal junction, 30 invalid address, 6 too old, 5 other), and 89 were excluded from the study (3 too ill, 5 unable to complete an English language questionnaire, 5 unable to be contacted, and 76 failed to return a completed questionnaire). Thus, 393 Barrett esophagus cases (285 nondysplastic, 108 dysplastic) completed the study.

Two separate control groups were recruited for the Study of Digestive Health: “inflammation controls” (i.e., patients who underwent endoscopy but for whom the histology report identified only esophageal inflammatory changes consistent with gastroesophageal reflux and no other pathologic or macroscopic changes including no evidence of Barrett esophagus), and population controls. For the purposes of developing risk prediction models to discriminate patients with Barrett esophagus from patients with gastroesophageal reflux disease without Barrett esophagus, the present analyses compared the Barrett esophagus cases to the inflammation controls only. In total, 706 of 1,354 patients approached as inflammation controls gave permission to be contacted (52% response rate). Of these, 57 refused to participate, 317 were ineligible (304 previous diagnosis of inflammation of the esophagus, 11 invalid address, 1 too old, 1 other), 19 were excluded (6 uncontactable, 1 psychological problems, 1 too ill, 3 unable to complete an English language questionnaire, 8 other), and 313 completed the study.

Both cases and inflammation controls were ineligible for the study if they had a previous diagnosis of Barrett esophagus or cancer. Approval to undertake the Study of Digestive Health was obtained from the human research ethics committees of the Queensland Institute of Medical Research (Queensland, Australia) and all participating hospitals, and written informed consent was obtained from all participants.

Candidate predictor variables were selected a priori from the literature and practitioner input and included: age (years); highest level of education (school only, technical college/diploma, university); body mass index (BMI) 1 year before diagnosis (<25, 25–29.9, ≥30 kg/m²); smoking status (never smoker, ex-smoker, current smoker); cumulative smoking exposure (never smoker, 0–29.9, ≥30 pack-years); smoking duration (never smoker, <15, 15–24, 25–34, ≥35 years); average lifetime alcohol consumption (nondrinker, <1, 1–6, 7–20, ≥21 drinks/wk); frequency of use of acid suppressant medications (including proton pump inhibitors and H₂-receptor antagonists) in the past 5 years (never, ever); frequency of use of aspirin and other nonsteroidal anti-inflammatory drugs (NSAID) in the past 5 years (never, less than weekly, at least weekly); physical activity levels (low, medium, high); average fruit (<2, ≥2 serves/d), and vegetable (<3, ≥3 serves/d) consumption; and number of comorbidities [categories defined by Charlson and colleagues (30)]. A standardized health and lifestyle questionnaire was used to collect detailed information on these variables for each participant. Most items in the questionnaire showed excellent repeatability after 4 months (31). Furthermore, we conducted a follow-up interview with the Barrett esophagus cases for up to 7 years after diagnosis and found similar self-reports of key characteristics (κ = 0.65–0.80) suggesting very high reproducibility for these measures. We imputed data for the small proportion of participants with missing values. We compared the model with imputed data with a complete case analysis and found similar model coefficients, but more precise estimates with imputed data.

The prediction model was externally validated using data from a community-based case–control study of Barrett esophagus conducted in western Washington State (29). Barrett esophagus cases were defined as residents aged 20 to 80 years newly diagnosed with Barrett esophagus (i.e., specialized intestinal metaplasia in an esophageal biopsy). Of the 208 patients diagnosed with Barrett esophagus, 193 (92.8%) were successfully interviewed. We subsequently excluded 18 cases, which were simultaneously diagnosed with esophageal adenocarcinoma (n = 2) and/or dysplasia (n = 16) from the validation analysis. Gastroesophageal reflux disease controls were a random sample of patients (∼50%) who underwent endoscopy for reflux symptoms, but who were biopsy-proven negative for Barrett esophagus. Of the 463 patients selected to be gastroesophageal reflux disease controls, 418 (90.8%) were successfully interviewed and were included in the validation analysis.

We used basic descriptive statistics to characterize the study populations. For comparisons between Barrett esophagus cases and inflammation controls, we used the χ² test for categorical variables and the Student t test for continuous variables. Statistical computations were carried out using SAS software (version 9.2; SAS Institute), and all tests for statistical significance were 2-sided at α = 0.05.

We developed separate risk models to predict nondysplastic Barrett esophagus and dysplastic Barrett esophagus in patients with gastroesophageal reflux symptoms. However, we were unable to externally validate the results for dysplastic Barrett esophagus and report only the nondysplastic Barrett esophagus model here (see Supplementary Material for the dysplastic Barrett esophagus risk model). As there is a sex difference in the incidence of Barrett esophagus and as some risk factors appear to have different associations with Barrett esophagus in men and women, we identified the important sex-specific predictors for Barrett esophagus and then pooled these in an overall risk model which included a term for sex. The predictors were identified using 2 phases of stepwise backwards logistic regression. In the first phase, we included in the multivariate model those variables that were statistically significantly associated with Barrett esophagus at the 5% level in univariate analyses and we conducted a backwards stepwise regression procedure, whereby those factors losing their significance in the multivariate analysis were dropped. In the second phase, those factors not significant in the univariate analyses were subsequently fitted to the multivariate model to identify those effects detectable only after adjusting for the major risk factors. There was no evidence of multicollinearity between the final list of predictor variables (all variance inflation factors << 10), and no interaction terms were statistically significant (all P values for the type III analysis of effects for interaction terms were >0.10).

The accuracy of the model was assessed using tests for discrimination and calibration (32). We evaluated predictive discrimination using the area under the receiver operator characteristic curve (AUC; also known as the c-statistic) and its 95% confidence interval (95% CI). The AUC can be interpreted as the probability that the model will assign a higher probability of actually having Barrett esophagus to a randomly chosen patient with Barrett esophagus than to a randomly chosen patient without Barrett esophagus, or simply the ability of the model to separate cases and inflammation controls. An AUC of 0.5 indicates that the model has a predictive discrimination no better than chance, whereas an AUC of 1.0 indicates a perfectly discriminating model. The second measure calculated was calibration, which compares the predicted probability with the observed risk. When the average predicted risk within decile categories matches the proportion observed, the model is well calibrated. We evaluated calibration using the Hosmer–Lemeshow goodness-of-fit statistic (33), where a high P value indicates excellent calibration. Calibration curves were also plotted to illustrate the model's fit across the range of predicted risk for Barrett esophagus compared with the observed outcome.

Epidemiologic data were available for the derivation analysis from 285 patients with nondysplastic Barrett esophagus (cases) and 313 inflammation controls (Table 1). Barrett esophagus cases were more likely to be men (64% vs. 47%, P < 0.001) and were, on average, 5 years older than inflammation controls. The majority (96%–97%) of cases and inflammation controls reported being Caucasian (P = 0.50).

All the potential predictive covariates with their univariate analyses are presented in Table 2. In the univariate analyses among men (181 cases, 147 inflammation controls), highest level of education, BMI, tobacco smoking, and frequency of use of acid suppressant medications were all statistically significantly associated with Barrett esophagus risk. In women (104 cases, 166 inflammation controls), only highest level of education and frequency of use of acid suppressant medications were significantly associated with Barrett esophagus risk in the univariate analyses. Alcohol consumption, frequency of use of NSAIDs, physical activity, fruit and vegetable consumption, and the presence of comorbidities were not statistically significantly associated with Barrett esophagus for men or women in the univariate analyses.

The variables retained in the final multivariate risk model included age, sex, smoking status, BMI, highest level of education, and frequency of use of acid suppressant medications (Table 3). There were no statistically significant interactions between sex and the other variables in the final model.

The risk prediction model for nondysplastic Barrett esophagus had good discrimination, with an AUC of 0.70 (95% CI, 0.66–0.74) in the development dataset. The discriminatory performance of the model in the validation dataset was more moderate with an AUC of 0.61 (95% CI, 0.56–0.66). Performance of the model was statistically good by the goodness-of-fit test (Hosmer–Lemeshow test, P = 0.75), and the calibration curve (Fig. 1) shows good agreement between predicted probabilities and actual Barrett esophagus risk across the observed range of risk.

To assess the potential effects of using the prediction model to guide referral for endoscopic screening for Barrett esophagus, we calculated the proportion of patients that would be referred for endoscopy at different probability thresholds for Barrett esophagus (Table 4). The first row gives the scenario of referring every patient with gastroesophageal reflux symptoms for endoscopy and therefore identifying all patients with gastroesophageal reflux symptoms who have Barrett esophagus (i.e., sensitivity is 100%). If patients are referred for endoscopy only if their predicted probability is, for example, 50% or more, the proportion of patients referred for endoscopy will be reduced to 46%. At that threshold, however, about 41% of Barrett esophagus cases (sensitivity, 59%) will not be referred for endoscopy. As the threshold increases, the number of referrals is reduced; as a consequence, however, the number of patients with Barrett esophagus who will not be referred for endoscopy increases.

In this study, we developed and externally validated a clinical risk prediction model for Barrett esophagus based on existing data from a large, population-based study. We used a rigorous statistical approach to determine the most important panel of risk factors to predict the presence of Barrett esophagus in patients with gastroesophageal reflux symptoms. As recent epidemiologic studies have shown that some risk factors, notably obesity, appear to have different associations with Barrett esophagus in men and women, we identified sex-specific predictors and then pooled these in an overall model. The final risk model included terms for age, sex, smoking status, BMI, highest level of education, and frequency of use of acid suppressant medications. External validation of the model showed that it performed moderately well in discriminating between patients with nondysplastic Barrett esophagus complicating their gastroesophageal reflux (cases) and those with no Barrett esophagus (inflammation controls), and the predicted risk correlated well with the observed risk.

As Barrett esophagus is assumed to be an intermediate step in the development of esophageal adenocarcinoma, screening to identify people with Barrett esophagus may be an effective strategy to prevent progression to cancer, at least in theory. Given the high prevalence of gastroesophageal reflux symptoms in the population and low prevalence of Barrett esophagus in these patients, endoscopic screening for Barrett esophagus in all patients with gastroesophageal reflux symptoms is not recommended in current international guidelines (13, 14, 34). The American Gastroenterological Association guideline recommends screening in selective populations with multiple risk factors for esophageal adenocarcinoma (age 50 years or older, male sex, white race, chronic gastroesophageal reflux, hiatal hernia, and elevated BMI; ref. 14). The evidence base underpinning these guidelines is not strong however, and adherence to the guidelines is likely to be incomplete. Data for Australia are limited; however, a New Zealand study showed that approximately 50% of indications for endoscopy in 2003 were for heartburn or dyspepsia (i.e., to exclude Barrett esophagus/esophageal adenocarcinoma), significantly higher than in 1997 (35). In healthcare systems throughout the world, there is an increasing need for evidence-based strategies including the need to establish an effective means of risk stratification for endoscopic Barrett esophagus screening among patients with gastroesophageal reflux symptoms.

Risk prediction can be used in clinical settings to stratify individuals into homogeneous risk groups. Risk prediction models are used in public health to quantify the probability of disease based on a combination of risk factors (36). So far, risk prediction approaches have been used extensively for cardiovascular diseases (16), and more recently, there has been a focus on deriving cancer risk prediction models (17–26). These models can complement clinical assessment, but they also ensure that the decision making process is more uniform across different centers by moving away from using any individual clinician's personal experience (37).

Previous efforts to develop a risk model for Barrett esophagus focused on identifying Barrett esophagus among patients with gastroesophageal reflux using gastrointestinal symptoms (27, 28). While these models performed well (AUCs > 0.70), they have not been externally validated. Our study used similar methods to derive a risk prediction model for Barrett esophagus using phenotypic and environmental risk factors, and tested its performance in an external population. The variables included in the model are all important and not necessarily causal correlates of Barrett esophagus and are supported by published findings from our own and other case–control and cohort studies of Barrett esophagus (10, 12, 38–41). Furthermore, to encourage generalizability, we emphasized the use of information on risk factors that can be obtained by practitioners in the office setting during routine healthcare. Importantly, the discriminatory accuracy for our model (AUC = 0.61 in the external validation dataset) compares favorably with cancer risk prediction models, such as the Gail (42) model for Breast cancer risk (0.58; ref. 43) and the LLP (22), Spitz (23), and Bach (21) lung cancer risk models (0.69, 0.69, and 0.62, respectively; ref. 44).

Applying this model to all patients with gastroesophageal reflux symptoms currently being considered for endoscopy, and using a threshold for making a decision has the potential to reduce the number of unnecessary endoscopies conducted to exclude Barrett esophagus. This model makes explicit the proportion of Barrett esophagus cases who would be missed because their predicted risk lies below the threshold. Because of the case–control study design, we were unable to determine precisely the positive and negative predictive values for the model. However, if we assume 5% prevalence of Barrett esophagus among persons with gastroesophageal reflux symptoms, our best estimate is that 5% to 17% of persons referred for endoscopy will have Barrett esophagus (depending upon the cut point chosen for referral) and 95% to 100% of persons not referred will not have Barrett esophagus. Assuming 15% prevalence, 15% to 41% of persons referred for endoscopy will have Barrett esophagus and 85% to 100% of people not referred will not have Barrett esophagus. In general, determining an acceptable threshold involves a trade-off between sensitivity and specificity. In screening for a lethal cancer, for example, high sensitivity is desirable, whereas for diseases with lower severity, a lower sensitivity can be tolerated. For our Barrett esophagus model, as the absolute risk of progression to cancer in patients with Barrett esophagus is low (8), a threshold whereby fewer investigations are conducted at the risk of missing more Barrett esophagus cases (i.e., increased specificity and decreased sensitivity) may be desirable.

Our study had strengths and limitations. The large samples of patients newly diagnosed with Barrett esophagus in the 2 settings were recruited prospectively, and comparable, consistent, and standardized criteria were used for histologic and endoscopic definitions of Barrett esophagus. Both the derivation sample and the validation sample only included patients who were selected for endoscopy; the large (but unknown) proportion of patients with gastroesophageal reflux symptoms not referred for screening had already been triaged away from endoscopy by clinicians using their own internal algorithms. Presumably, the clinicians had decided that those patients were at such low risk of significant pathology that there was no net benefit from undergoing endoscopic investigation. As such, it is likely that had those low risk patients been included in the 2 samples, our prediction models would have performed even better. While our modeling assumes that endoscopy is conducted in the setting of gastroesophageal reflux symptoms solely to exclude Barrett esophagus, endoscopy may be undertaken for other indications in this clinical setting. If so, then this would tend to attenuate the predictive value of the models that we have derived, as those patients being referred for other indications would presumably be at a lower risk of Barrett esophagus than those being referred to confirm the clinical diagnosis.

A limitation of the Australian study was the relatively low rate of participation, raising concerns about possible biased selection of cases and controls. Because Barrett esophagus cases and inflammation controls were sampled from the same population, navigated the same clinical pathways, and were recruited using identical methods, it is unlikely that systematically biased selection of one or other group explains our findings. Moreover, Barrett esophagus cases and inflammation controls were not informed about the hypotheses being tested, and while biased recall of nonreflux exposures is possible, we consider the likelihood that this accounts for our observations as low. Although there were 108 dysplastic Barrett esophagus cases in our development dataset, we were unable to obtain a validation dataset with dysplastic Barrett esophagus cases. The best estimate for the performance of our model for predicting Barrett esophagus with dysplasia was an AUC estimate of 0.87 (95% CI, 0.83–0.91) in the development dataset. Recently, central obesity has been found to be more strongly associated with Barrett esophagus than BMI; however, measures of central obesity (e.g., waist-to-hip ratio) were not collected for this study. It is likely that adding such measures to the model, rather than BMI, may improve predictive accuracy. Finally, our sample was predominantly white (∼97%) and thus the models may not be applicable to other ethnic groups.

This parsimonious model, however, could be considered as a starting point for further development, as a number of risk factors were not included and genetic information may also be important in predicting risk of Barrett esophagus. The inclusion of other environmental risk factors and the extension of the model to include biomarkers may go further to improving performance. However, a recent study has shown that breast cancer risk prediction does not improve significantly when genetic information is included in the risk model (45).

In summary, we have derived and externally validated a risk prediction model, which estimates the likelihood of undiagnosed Barrett esophagus in patients with gastroesophageal reflux symptoms being considered for upper gastrointestinal endoscopy. The prediction model has the potential to be a useful tool in the clinical setting for decisions regarding investigation and treatment of patients with gastroesophageal reflux.

The contents of this publication are solely the responsibility of the authors and do not necessarily represent the official views of the National Cancer Institute. No potential conflicts of interest were disclosed.

Conception and design: A.P. Thrift, B.J. Kendall, T.L. Vaughan, D.C. Whiteman

Development of methodology: A.P. Thrift, D.C. Whiteman

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): T.L. Vaughan, D.C. Whiteman

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): A.P. Thrift, B.J. Kendall, N. Pandeya, T.L. Vaughan, D.C. Whiteman

Writing, review, and/or revision of the manuscript: A.P. Thrift, B.J. Kendall, T.L. Vaughan, D.C. Whiteman, N. Pandeya

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): N. Pandeya, D.C. Whiteman

Study supervision: N. Pandeya, D.C. Whiteman

The authors thank the following institutions: Sullivan and Nicolaides Pathology (Brisbane); Queensland Medical Laboratory (Brisbane); and Queensland Health Pathology Services (Brisbane). The authors also thank the contribution of the study nurses and research assistants and all of the people who participated in the study.

Study of Digestive Health Investigators: Queensland Institute of Medical Research, Brisbane, Australia: David C Whiteman MBBS, PhD; Adele C Green MBBS, PhD; Nicholas K Hayward PhD; Peter G Parsons PhD; Sandra J Pavey PhD, David M Purdie PhD; Penelope M Webb DPhil.

University of Queensland, Brisbane, Australia: David Gotley FRACS; Mark Smithers, FRACS.

The University of Adelaide, Adelaide, Australia: Glyn G Jamieson FRACS.

Flinders University, Adelaide, Australia: Paul Drew, PhD; David I Watson FRACS.

Envoi Pathology, Brisbane, Australia: Andrew Clouston, PhD, FRCPA.

Study of Digestive Health Research Staff:

Project Manager: Suzanne O'Brien (QIMR); Data Manager: Troy Sadkowsky (QIMR); Research Nurses: Andrea McMurtrie, Linda Terry, Michael Connard, Lea Jackman, Susan Perry, Marcia Davis.

Histopathology Collaborators: Ian Brown (Envoi Pathology), Neal Walker (Envoi Pathology).

Funding for the Study of Digestive Health (5 RO1 CA001833-02) and the validation study (R01 CA072866, K05 CA124911, R01 CA136725) was received from the National Cancer Institute.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.