Advertisement
Browse Subject Areas
?

Click through the PLOS taxonomy to find articles in your field.

For more information about PLOS Subject Areas, click here.

  • Loading metrics

Diagnostic Accuracy of Age and Alarm Symptoms for Upper GI Malignancy in Patients with Dyspepsia in a GI Clinic: A 7-Year Cross-Sectional Study

  • Hooman Khademi,

    Affiliations Digestive Disease Research Institute, Shariati Hospital, Tehran University of Medical Sciences, Tehran, Iran, International Agency for Research on Cancer, Lyon, France

  • Amir-Reza Radmard,

    Affiliation Digestive Disease Research Institute, Shariati Hospital, Tehran University of Medical Sciences, Tehran, Iran

  • Fatemeh Malekzadeh,

    Affiliation Digestive Disease Research Institute, Shariati Hospital, Tehran University of Medical Sciences, Tehran, Iran

  • Farin Kamangar,

    Affiliations Digestive Disease Research Institute, Shariati Hospital, Tehran University of Medical Sciences, Tehran, Iran, Department of Public Health Analysis, School of Community Health and Policy, Morgan State University, Baltimore, Maryland, United States of America

  • Siavosh Nasseri-Moghaddam,

    Affiliation Digestive Disease Research Institute, Shariati Hospital, Tehran University of Medical Sciences, Tehran, Iran

  • Mattias Johansson,

    Affiliation International Agency for Research on Cancer, Lyon, France

  • Graham Byrnes,

    Affiliation International Agency for Research on Cancer, Lyon, France

  • Paul Brennan,

    Affiliation International Agency for Research on Cancer, Lyon, France

  • Reza Malekzadeh

    malek@ams.ac.ir

    Affiliation Digestive Disease Research Institute, Shariati Hospital, Tehran University of Medical Sciences, Tehran, Iran

Diagnostic Accuracy of Age and Alarm Symptoms for Upper GI Malignancy in Patients with Dyspepsia in a GI Clinic: A 7-Year Cross-Sectional Study

  • Hooman Khademi, 
  • Amir-Reza Radmard, 
  • Fatemeh Malekzadeh, 
  • Farin Kamangar, 
  • Siavosh Nasseri-Moghaddam, 
  • Mattias Johansson, 
  • Graham Byrnes, 
  • Paul Brennan, 
  • Reza Malekzadeh
PLOS
x

Abstract

Objectives

We investigated whether using demographic characteristics and alarm symptoms can accurately predict cancer in patients with dyspepsia in Iran, where upper GI cancers and H. pylori infection are common.

Methods

All consecutive patients referred to a tertiary gastroenterology clinic in Tehran, Iran, from 2002 to 2009 were invited to participate in this study. Each patient completed a standard questionnaire and underwent upper gastrointestinal endoscopy. Alarm symptoms included in the questionnaire were weight loss, dysphagia, GI bleeding, and persistent vomiting. We used logistic regression models to estimate the diagnostic value of each variable in combination with other ones, and to develop a risk-prediction model.

Results

A total of 2,847 patients with dyspepsia participated in this study, of whom 87 (3.1%) had upper GI malignancy. Patients reporting at least one of the alarm symptoms constituted 66.7% of cancer patients compared to 38.9% in patients without cancer (p<0.001). Esophageal or gastric cancers in patients with dyspepsia was associated with older age, being male, and symptoms of weight loss and vomiting. Each single predictor had low sensitivity and specificity. Using a combination of age, alarm symptoms, and smoking, we built a risk-prediction model that distinguished between high-risk and low-risk individuals with an area under the ROC curve of 0.85 and acceptable calibration.

Conclusions

None of the predictors demonstrated high diagnostic accuracy. While our risk-prediction model had reasonable accuracy, some cancer cases would have remained undiagnosed. Therefore, where available, low cost endoscopy may be preferable for dyspeptic older patient or those with history of weight loss.

Introduction

Dyspepsia, a condition defined as recurrent or persistent pain or discomfort centered in the upper abdomen, [1] affects 25%–40% of adults in the general population of the United States, incurring over $12 billion per year in direct annual costs in the United States and nearly £1 billion per year in the United Kingdom. [2][6] Several benign or malignant disorders may underlie dyspepsia, including esophagitis, gastroesophageal reflux disease (GERD), peptic ulcer disease (PUD), erosive duodenitis, [7] and most importantly upper gastrointestinal (UGI) malignancies, which are estimated to be responsible for 1%–3% of all cases of dyspepsia. [7][10] However, in over half of the dyspeptic patients no obvious structural abnormality can be found, a condition called “functional” or “non-ulcer” dyspepsia. [1], [11][13] Recently some experts have argued that GERD should be excluded from the etiologies of dyspepsia and treated as a different entity, [2], [14] but this is still in dispute. [15], [16]

There are several alternative strategies for initial management of dyspepsia including empirical acid suppressive therapy, H. pylori test and treat, and prompt endoscopy, [17], [18] and several studies have tried to find the best strategy. [11][13], [18][20] It has been suggested that the most cost-effective initial approach in primary care, particularly in countries with low rates of H. pylori infection is test and treat strategy. [17], [21][23] However, it may delay early diagnosis of malignant underlying disease beyond the point where it is still curable and also might not be practical in countries with very high rates of H. pylori infection, such as Iran. In addition, endoscopy is an accurate but costly method of early diagnosis of UGI malignancies, which are considered as the most important causes of global cancer deaths. [24] It may be cost-effective to stratify dyspeptic patients as high-risk and low-risk, and then perform immediate endoscopy on the high-risk group while applying other alternatives for the low-risk group. Thus some experts have recommended prompt endoscopy in newly diagnosed dyspeptic patients having any alarm symptoms including unintentional weight loss (>10% of body weight), dysphagia, GI bleeding, persistent vomiting, abdominal palpable mass and anemia, as well as in patients who are over age 50. [12], [19], [25][27] In contrast, several studies have shown limited predictive value for either alarm features or age to be able to differentiate low- and high-risk dyspeptic patients for underlying malignancies. [28][33] Prompt endoscopy in patients over 50 years regardless of alarm symptom status has been shown to increase the proportion of curable cases of UGI malignancies by as much as 30%, [34][36], but the cost-effectiveness of initial endoscopy in this age group for improving survival of cancer patients is uncertain. [36], [37] Distinct UGI malignancy incidence rates and various distributions of its topographical types in different populations [7][10] as well as differences in H. pylori infection rates [38], [39] could partly explain the variable results.

Gastric cancer, followed by esophageal cancer, is reported as the most common cancer in Iranian men. As well, H. pylori infection is highly prevalent (>80%) in the Iranian adult population. [39][45] Although acid peptic disease is also still common in Iran, [44], [46] the major indication for UGI endoscopy in Iran is ruling out upper GI malignancy as underlying cause. We have conducted a relatively large-scale study to assess the role of alarm symptoms and their diagnostic accuracy in predicting UGI malignancy in patients with dyspepsia in a country with high prevalence of H. pylori infection and upper GI malignancy. Through developing a risk-prediction model, we also tried to find a way to maximally use all information from age and alarm symptoms, altogether, to find high-risk individuals for UGI malignancy. To the best of our knowledge, no previous study investigated alarm symptoms in Western Asia and Middle East region.

Methods

Study population

All consecutive patients referred to Behrooz Clinic, a tertiary referral gastroenterology clinic in Tehran, and diagnosed with dyspepsia from 2002 to 2009 were invited to participate in this study. Patients with UGI malignancy previously diagnosed through other imaging tools, such as CT scan or barium swallow, and patients who had already a diagnosis of UGI cancer or undergone gastrectomy or esophagectomy (4 cases) were not included in this study. All study participants signed a written informed consent and Institutional Review Board of Digestive Disease Research Institute of Tehran University of Medical Sciences (TUMS) approved the study design and methods.

Exposure assessment

Demographic and anthropometric characteristics, history of any alarm symptoms, family history of UGI malignancies in first degree relatives, and also data on cigarette smoking status were collected by interviewing the patients. Body mass (kg) and height (cm) were measured; body mass index (BMI) was calculated and categorized based on WHO recommendations. Pack-years (pys) of cigarette smoking were calculated by multiplying duration of smoking (in years) and daily use amount (in cigarettes per day divided by 20). Accordingly, we categorized all patients into four smoking groups: never smokers, ex-smokers (quit smoking more than a year before interview), current light smokers (less than 20 pys) and current heavy smokers (20 pys or more). Rapid urease test (RUT) was performed on all patients during endoscopy to detect H. pylori infection. Alarm symptoms in this study were unintentional weight loss (≥10% of body weight in recent 6 months), dysphagia (perception of an impediment to the normal passage of swallowed material), GI bleeding (any evidence of hematemesis and/or melena), and persistent vomiting (at least 7 to 10 days of protracted vomiting). [47], [48]

Outcome measurement

All patients underwent prompt endoscopy using Olympus video-endoscopes (GIF type-160), while they were asked not to use proton pump inhibitors (PPIs) or H2 blockers for at least 2 weeks prior to endoscopy to avoid their masking effect on visibility of malignancy during endoscopy. [49] In case of any suspected malignancy, multiple biopsy specimens were taken from the suspected lesion and were sent to two separate pathology centers. All cancer diagnoses were histologically confirmed. UGI malignancy was defined as any histologically confirmed esophageal, gastric or duodenal cancer detected during endoscopy.

Statistical Analysis

Using histology as the gold standard for diagnosis of UGI malignancies, we calculated sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). We also calculated and present functions of sensitivity and specificity, including positive diagnostic likelihood ratio (PDLR), negative diagnostic likelihood ratio (NDLR), and diagnostic odds ratio (OR), and their related 95% confidence intervals, as measures of diagnostic accuracy for each individual alarm symptom. [50]

We estimated odds ratios (ORs) and 95% confidence intervals for age, demographics, and each of alarm symptoms using univariable and multivariable adjusted logistic regression models. Adjusted model included age, gender, level of education, cigarette smoking, history of weight loss, GI bleeding, persistent vomiting and dysphagia. Based on regression model findings, we decided to report diagnostic accuracy measures for each alarm symptom, in four age categories: less than 36 years of age, between 36 and 49, between 50 and 65 and finally 65 or older, as the patients in each group were similar within the group and different from the other groups in terms of OR.

We developed a risk-prediction model, through a backward stepwise selection in a multivariable logistic regression analysis. We used all variables that showed significant association with UGI malignancy in univariable regression analysis, except for BMI and H. pylori. BMI was not included because of its collinearity with weight loss. H. pylori was not included because it was measured during endoscopy (and not prior to that), so it was not helpful to assess the need for endoscopy. We also included all possible two-way interaction terms between four alarm symptoms in the initial model. Subsequently, using an automated backward-stepwise multivariable method, we removed the predictors with highest p values on the basis of Wald test, so that achieving the final model that only comprised by the predictors with a multivariable p value of less than 0.05. Although forward selection gives a more parsimonious model, backward selection is generally preferable if stepwise selection is applied. [51] Since there were 46 missing values in two variables; family history of UGI cancer (17 patients) and level of education (29 patients) we performed the analyses on 2,801 remaining patients; 82 UGI malignancies. The complete case analysis approach was applied, because of the small percentage of missing values in only two variables that was considered to be completely at random.

As the rule for prediction model, we used the linear predictor, which is the sum of the products of regression coefficients with the corresponding variable values from the final logistic model, and for convenience added a constant value of 1.5 to ensure it was positive; Risk Score = 1.5+(Regression Coefficient×Variable Value)

The corresponding risk probability was calculated using the equation; P = 1/(1+e−ß), where ß = Constant+(Regression Coefficient×Variable Value); for example, probability (P) thresholds of 1%, 10% and 25% correspond to risk score (RS) cutoff levels of 3.1, 5.5 and 6.6, respectively (Figure 1).

thumbnail
Figure 1. The risk scores and their corresponding risk probabilities derived from the suggested risk-prediction model.

http://dx.doi.org/10.1371/journal.pone.0039173.g001

Three percent of all patients had UGI malignancies, hence we chose thresholds of 1% and 10%, round numbers that were approximately three times lower and higher than the overall population average, to denote low- or high-risk, respectively. Subsequently, we defined four risk groups for UGI malignancy: low-risk group with a probability of less than 1%, intermediate-risk group with a probability between 1% to 10%, high-risk group with a probability of 10% to 25%, and excessive-risk group with a probability of higher than 25%.

Once the variables to be included in the model were defined, we examined the calibration of the model by performing Hosmer-Lemeshow goodness-of-fit test. We also assessed the overall performance of the model using Nagelkerke's R2-as a measure of explained variation in log-likelihood scale, and Brier score (or average prediction error). [52] Finally, the discrimination capability of the model was estimated using the area under the receiver operating characteristic (ROC) curve or AUC and its 95% confidence intervals. To further evaluating the discrimination ability of our risk-prediction model, we compared it with two other models: 1) age only; 2) age plus alarm symptoms; in terms of AUC (Figure 2). We also plotted reclassification table and calculated net reclassification index (NRI) and integrated discrimination index (IDI), investigating added discriminatory performance of our suggested risk-prediction model compared to model 2. [53][58] In evaluating model performance, over-fitting is a well-known statistical phenomenon where a model will always perform better on the data used to construct it than when predicting from independent but similar data. While we had no data to externally validate our model, we used repeated 10% cross-validation to guard against over-fitting. [51], [52] In this procedure the model was fitted to a randomly selected 90% of the data then tested on the remaining 10%; the procedure then being repeated 10 times and the resulting statistics averaged.

thumbnail
Figure 2. ROC curves based on three predictive model for upper GI malignancy in dyspeptic patients, including model1; only age included, model2; age plus alarm symptoms, and model3; the risk-prediction model.

http://dx.doi.org/10.1371/journal.pone.0039173.g002

All statistical analyses were conducted using STATA statistical software, version 11 (STATA Corp, College Station, TX) and all reported p-values are 2-sided.

Results

A total of 2,847 patients with dyspepsia who were referred to Behrooz clinic, a tertiary GI clinic in Tehran, participated in this study. Table 1 shows the demographic characteristics, as well as habits and distribution of alarm symptoms in study participants. The mean (±SD) age of the participants was 42.5 (±15.0), and approximately half (50.5%) were females. Of these, 1,131 patients (39.7%) had at least one alarm symptom; the most common reported alarm symptom was dysphagia (n = 547; 19.2%).

UGI malignancies were histologically confirmed in 87 (3.1%) cases. Compared to all patients with dyspepsia, patients with UGI malignancies were significantly older (mean age of 58.0 years); more likely to be male (60.9%); to have an education of less than high school diploma (52.4%); to have a family history of UGI malignancies (14.9%); to have ever smoked cigarettes (35.6%); and more likely to be positive for H. pylori (70.6%) (Table 1). Patients reporting at least one of the alarm symptoms constituted 66.7% of patients with UGI malignancies compared to 38.9% in patients without cancer (p value<0.001).

thumbnail
Table 1. Demographic characteristics, habits and distribution of alarm symptoms in all dyspeptic patients and who with upper GI malignancy.

http://dx.doi.org/10.1371/journal.pone.0039173.t001

Table 2 shows the endoscopic and histological findings in the study participants. The most common endoscopic findings were GERD (72.6%), followed by PUD (15.0%), and UGI malignancies (3.1%). The Los Angeles (LA) classification was used for the endoscopic diagnosis of GERD, which classifies it into 4 subgroups; A to D, according to number and length of observed mucosal breaks and involvement of one or more mucosal folds. [59] Of the 87 patients with cancer, 68 (78.2%) were diagnosed with gastric cancer, 16 (18.4%) with esophageal cancer, and 3 (3.4%) with duodenal cancer. The majority of all malignancies (54.0%) were well-differentiated. Esophageal cancers were located more in the middle-third of the esophagus (68.7%) and were more of squamous cell type (62.5%). The majority of gastric cancers were adenocarcinomas (88.2%) and were located in the antrum (35.3%). Of the 3 duodenal cancers, 2 (66.7%) were seen in D1. (Table 2)

Table S1 shows diagnostic values for alarm symptoms for all participants and by age category. The prevalence of UGI cancers by age category, from youngest to oldest, was 0.71%, 1.89%, 3.65%, and 14.3%, respectively. Due to increasing prevalence of cancer with age, PPV increased as a function of age. For example PPV for dysphagia was nearly 16-fold higher in the oldest versus the youngest group. Among alarm symptoms, weight loss was the strongest predictor of UGI cancers.

We calculated diagnostic values for experiencing only one, two, and more than two alarm symptoms, and at least one symptom, compared to patients who had never experienced any alarm symptoms, as reference group (Table S2). Positive predictive value increased with increasing number of reported alarm symptoms and older age, such that PPV increased from 0.70% in patients younger than 35 years of age with only one alarm symptom to 58.3% in patients older than 65 years of age with more than two alarm symptoms. Having at least one alarm symptom had the highest sensitivity (66.7%) but the lowest specificity (61.1%).

We used multivariable adjusted logistic regression models to study the independent diagnostic ORs for each alarm symptom (Table 3). In these models, age showed very high ORs, in both unadjusted and adjusted models, with OR (95%CI) of 22.8 (8.86–58.5), for the oldest (≥65 years old) compared to the youngest age group; a near 2-fold increase in cancer odds was observed for each 10 years increase in age. Men did not have significantly higher odds of UGI cancers than women. Heavy smokers were in a significantly higher risk of developing UGI malignancies compared to never-smokers (OR (95%CI): 5.07 (2.33–11.0)). Among alarm symptoms, weight loss was the leading predictive factor for UGI malignancy in adjusted models (OR (95%CI) = 4.89 (2.91–8.23)), while persistent vomiting with OR (95%CI) of 2.26 (1.27–4.03) was the second most important alarm symptom. Patients with UGI cancer were approximately twice as likely to have positive RUT results compared to patients without malignancy, mainly due to the association between H. pylori and gastric cancer (OR (95%CI) = 3.05 (1.65–5.64)).

As fully explained in the methods part, we developed a risk-prediction model to predict UGI malignancies in dyspeptic patients (Table 4). Figure 1 shows the risk score (RS) thresholds and their corresponding risk probabilities, as described in the methods section; low-risk group with a RS<3.1; intermediate-risk group with a 3.1≤RS<5.5; high-risk group with a 5.5≤RS<6.5; and excessive-risk group with a RS of ≥6.6. To provide some examples, a 35-year-old never-smoker who did not have any of the alarm symptoms had a risk score of 1.5 and was categorized as low-risk (<1% chance of UGI cancer), whereas a 65-year-old current heavy smoker with weight loss but no other alarm symptoms had a score of 7.9 ( = 1.5+3.3+1.3+1.8) and was therefore categorized as excessive-risk (>25% chance of cancer). Using RS = 2.2(or 5.5, or 6.6) as the cutoff levels, the estimated sensitivity and specificity was equal to 100% (or 42.0%, or 29.6%) and 24.5% (or 95.0%, or 98.9%), respectively (Table 5).

thumbnail
Table 3. Estimated odds ratios of demographic characteristics and alarm symptoms for upper GI malignancies, based on unadjusted and multivariable adjusted regression models.

http://dx.doi.org/10.1371/journal.pone.0039173.t003

thumbnail
Table 4. Risk-prediction model for predicting risk of upper GI malignancy in dyspeptic patients.

http://dx.doi.org/10.1371/journal.pone.0039173.t004

thumbnail
Table 5. The diagnostic characteristics of choosing different risk score cut-off levels, derived from the risk-prediction model.

http://dx.doi.org/10.1371/journal.pone.0039173.t005

The suggested risk-prediction model, including 8 predictors, had the number of events per variable (EPV) of about 10, which indicates an acceptable sample size to provide an adequate risk-prediction model. [52] As shown in Table 6, a nonsignificant Hosmer-lemeshow test (p = 0.71) showed that this model adequately predicts, for each level of risk, the percentage of patients with the outcome (good calibration). The comparison of Brier score and Nagelkerke's R2 measures in three models, suggested slightly better overall performance for the risk-prediction model. The Akaike information criterion (AIC) was in favor of the proposed risk-prediction model (model3); however, the Bayesian information criterion (BIC) didn't show any superiority for model3 versus model2. The net reclassification index (NRI) of 23% compared to model2, again advocated for the third model; however, the integrated discrimination index (IDI), which was calculated by subtracting discrimination slopes of compared models, indicated a minor improvement in discrimination ability of model3 versus model2; 4.3% (Table 6). The estimated AUC comparison showed a statistically significant higher discriminatory capacity of model3, though not substantially; the AUC (95% CI) of 0.852 (0.812–0.893) for risk-prediction model was significantly higher than both model2 (p = 0.022) with AUC (95% CI) of 0.822 (0.774–0.870) and model1 (p<0.001) with AUC (95% CI) of 0.760 (0.709–0.812) (Figure 2).

thumbnail
Table 6. Performance of three different models for predicting UGI malignancy in dyspeptic patients.

http://dx.doi.org/10.1371/journal.pone.0039173.t006

Using the marginal numbers of a reclassification table (Table 7), we evaluated the calibration of the risk-prediction model compared to model2, which demonstrated comparable predicted probabilities with observed proportions, except in the third group, with risk probabilities between 10% and 25%.

thumbnail
Table 7. Reclassification table for age and alarm symptom model (Model2) and risk-prediction model (Model3).

http://dx.doi.org/10.1371/journal.pone.0039173.t007

The results of repeated 10% cross-validation indicated that the estimated average of AUC from risk-prediction model (0.820; 95%CI: 0.764–0.876) was not largely different from the average AUC in validation set (0.796; 95%CI: 0.744–0.848); 0.024.

Discussion

We studied age and several alarm symptoms to learn whether they can provide useful diagnostic information to classify dyspeptic patients, referred to a tertiary GI clinic, as high-risk and low-risk for UGI cancers.

In the adjusted models, older age, history of weight loss, history of GI bleeding, persistent vomiting, being current cigarette smoker, family history of UGI cancer, and H. pylori positivity were all positively associated with risk of UGI cancers. Of these, age and weight loss were the most important predictors. Other predictors, such as male sex, lower education, and history of dysphagia were also associated with higher risk in unadjusted models, but lost statistical significance in the adjusted models. Since we measured H. pylori infection by rapid-urease-test (RUT) during endoscopy, this variable wasn't included in risk-prediction model. Moreover, the majority of H. Pylori infected gastric cancer patients develop severe gastric atrophy before gastric cancer, making stomach environment unfavorable for H. pylori survival, and thus would become H. pylori negative by RUT.

Some previous studies have also assessed the value of age and alarm symptoms in predicting risk of cancer in dyspeptic patients. [28], [60], [61] Bai and colleagues studied the predictive value of alarm symptoms and age for UGI malignancy in China and found limited value for either age or any alarm symptoms. [61] In their study, alarm symptoms were highly specific but had low sensitivity. However, they based most of their discussion on PDLR of each symptom and did not build models using all predictor variables to predict risk of UGI malignancies. Performing a meta-analysis, Fransen and colleagues found limited diagnostic values including sensitivity, specificity and predictive values, for each individual alarm symptom, i.e., dysphagia, weight loss, bleeding, and vomiting. [33] They suggested using alarm symptoms in combination with other factors – such as age, gender, or smoking – might be a better tool for selection of high-risk patients; however they were unable to test their hypothesis. Kapoor et al., [28] built a model using a number of alarm symptoms and age, and validated their model in another group of patients. Using a combination of symptoms, they were able to generate a model with high sensitivity and high NPV, but low specificity and low PPV, to predict risk of UGI malignancies. However, they did not use the weight of the symptoms based on their odds ratio and perhaps did not make use of the full extent of information in their dataset. Finally, Numans and colleagues [60] developed a risk-prediction model using calculated total scores and showed that classical alarm symptoms, via a risk-prediction model, are useful predictors of UGI malignancy. However, their model is somewhat complex and, with inclusion of several variables, somewhat unstable. Like the results of our study, nearly all of these studies showed relatively low value for each alarm symptom, but perhaps a number of unnecessary endoscopies could be avoided using a combination of symptoms.

Although we found several variables that were each associated with higher risk of having cancer, our results showed that no single predictor could perfectly differentiate between high- and low-risk groups; sensitivities and specificities for each of the predictors were far from one. In principle, simply adding the number of risk factors is not the most efficient use of data, as different risk factors predicted cancer with substantially different odds ratios. Therefore, the most appropriate way of predicting risk would be using the risk-prediction model. However, our results show that the proposed risk-prediction model was unable to provide any important improvement in prediction compare to a model based on including only age and the generally accepted alarm symptoms.

Our proposed risk-prediction model was not perfect, despite acceptable overall model fit and calibration. However, such model could somewhat adequately discriminate patients in our setting into a wide range, with risks less than 1% to risks over 25%, with acceptable calibration. Given that all of the predictors used in this risk-prediction model could easily be obtained from a simple questionnaire, this might provide useful information for the physician in deciding whether to perform immediate endoscopy or to first try empiric forms of treatment.

Some predictors of cancer, such as dysphagia, were not statistically significantly associated with odds of cancer and were excluded from our model. Recent studies showed that dysphagia could be, more often, considered as a GERD symptom, rather than esophageal cancer. [62] Owing the fact that the majority (72.6%) of study participants were GERD patients, finding no significant associations shouldn't be surprising. Furthermore, most esophageal cancer patients with dyspepsia are relatively elderly patients and dysphagia indicate that the cancer is beyond the point of curability, [63] therefore, use of this risk-prediction model would perhaps not be a significant risk to them.

Making a decision as to perform endoscopy versus provide other treatments first requires a careful cost-benefit analysis. Such analysis depends partly on the risk-prediction model but it needs to take into consideration other factors such as probability of missing a potentially curable cancer if the treatment is delayed by a few weeks; additional benefits of endoscopy such as diagnosis of conditions other than cancer, as well as its harms and cost; prevalence of cancer and other underlying diseases causing dyspepsia; availability of endoscopic facilities; and its cost in any specific health setting.

In the United States and most European countries where H. pylori and UGI malignancy prevalence is low, while the cost of upper GI endoscopy is very high, cost-effectiveness analysis usually reveal that initial endoscopy is not beneficial and a test and treat approach is the most cost-efficient strategy. [64] However, applying a validated risk-prediction model to find high-risk patients for UGI malignancies and targeting them for performing endoscopy might be an alternative strategy to better compare the cost-benefit of two approaches. Furthermore, for Asian countries such as China and Iran, this recommendation, probably would not be applicable. [61] Unless, non-invasive and cheaper tests become available in countries like Iran, where endoscopy is widely available with a relatively low cost, prompt endoscopy may be recommended in all dyspeptic patients older than 50, with weight loss, or with any additional alarm symptoms.

The strengths of our study are relatively large sample size, availability of data on at least 10 predictors, and constructing risk-prediction model. A limitation of the study is that the risk-prediction model was based on a development (training) set and there was no external validation set.

In summary, none of the predictors that we studied demonstrated high diagnostic accuracy. Using age, alarm symptoms, family history of UGI cancer and smoking, we were able to construct a useful risk-prediction model that distinguished between high-risk and low-risk individuals with a ROC curve AUC of 0.85 and adequate overall calibration and model fit measures. However, the decision on how to use this model will depend on cost-benefit analytic models that depend on several other factors.

Supporting Information

Table S1.

Diagnostic accuracy measures of alarm symptoms on UGI malignancies.

doi:10.1371/journal.pone.0039173.s001

(DOC)

Table S2.

Diagnostic accuracy measures for having only one, two, more than two and at least one alarm symptom in all ages, as well as according to age categories.

doi:10.1371/journal.pone.0039173.s002

(DOC)

Author Contributions

Conceived and designed the experiments: HK ARR RM. Performed the experiments: HK ARR FM SNM RM. Analyzed the data: HK FK MJ GB PB. Wrote the paper: HK ARR FM FK SNM MJ GB PB RM.

References

  1. 1. Barbara L, Camilleri M, Corinaldesi R, Crean GP, Heading RC, et al. (1989) Definition and investigation of dyspepsia. Consensus of an international ad hoc working party. Dig Dis Sci 34: 1272–1276.
  2. 2. Talley NJ, Vakil NB, Moayyedi P (2005) American gastroenterological association technical review on the evaluation of dyspepsia. Gastroenterology 129: 1756–1780.
  3. 3. Sandler RS, Everhart JE, Donowitz M, Adams E, Cronin K, et al. (2002) The burden of selected digestive diseases in the United States. Gastroenterology 122: 1500–1511.
  4. 4. Talley NJ, Zinsmeister AR, Schleck CD, Melton LJ 3rd (1992) Dyspepsia and dyspepsia subgroups: a population-based study. Gastroenterology 102: 1259–1268.
  5. 5. El-Serag HB, Talley NJ (2004) Systemic review: the prevalence and clinical course of functional dyspepsia. Aliment Pharmacol Ther 19: 643–654.
  6. 6. Moayyedi P, Mason J (2002) Clinical and economic consequences of dyspepsia in the community. Gut 50: iv10–12.
  7. 7. Heikkinen M, Pikkarainen P, Takala J, Rasanen H, Julkunen R (1995) Etiology of dyspepsia: four hundred unselected consecutive patients in general practice. Scand J Gastroenterol 30: 519–523.
  8. 8. Ayoola EA, al-Rashed RS, al-Mofleh IA, al-Faleh FZ, Laajam M (1996) Diagnostic yield of upper gastrointestinal endoscopy in relation to age and gender: a study of 10112 Saudi patients. Hepatogastroenterology 43: 409–415.
  9. 9. Mansi C, Mela GS, Pasini D, Grosso M, Corti L, et al. (1990) Patterns of dyspepsia in patients with no clinical evidence of organic diseases. Dig Dis Sci 35: 1452–1458.
  10. 10. Talley NJ (1993) Nonulcer dyspepsia: current approaches to diagnosis and management. Am Fam Physician 47: 1407–1416.
  11. 11. Bazaldua OV, Schneider FD (1999) Evaluation and management of dyspepsia. Am Fam Physician 60: 1773–1784, 1787–1778.
  12. 12. Talley NJ, Silverstein MD, Agreus L, Nyren O, Sonnenberg A, et al. (1998) AGA technical review: evaluation of dyspepsia. American Gastroenterological Association. Gastroenterology 114: 582–595.
  13. 13. Fisher RS, Parkman HP (1998) Management of nonulcer dyspepsia. N Engl J Med 339: 1376–1381.
  14. 14. Bytzer P, Talley NJ (2001) Dyspepsia. Ann Intern Med 134: 815–822.
  15. 15. Thomson AB, Barkun AN, Armstrong D, Chiba N, White RJ, et al. (2003) The prevalence of clinically significant endoscopic findings in primary care patients with uninvestigated dyspepsia: the Canadian Adult Dyspepsia Empiric Treatment – Prompt Endoscopy (CADET-PE) study. Aliment Pharmacol Ther 17: 1481–1491.
  16. 16. Moayyedi P, Axon AT (1999) The usefulness of the likelihood ratio in the diagnosis of dyspepsia and gastroesophageal reflux disease. Am J Gastroenterol 94: 3122–3125.
  17. 17. Delaney B, Ford AC, Forman D, Moayyedi P, Qume M (2005) Initial management strategies for dyspepsia. Cochrane Database Syst Rev. CD001961 p.
  18. 18. Ford AC, Moayyedi P (2008) Current guidelines for dyspepsia management. Dig Dis 26: 225–230.
  19. 19. Talley NJ, Vakil N (2005) Guidelines for the management of dyspepsia. Am J Gastroenterol 100: 2324–2337.
  20. 20. (2004) National Institute for Health and Clinical Excellence. Dyspepsia: Managing dyspepsia in adults in primary care.
  21. 21. Malfertheiner P, Megraud F, O'Morain C, Hungin AP, Jones R, et al. (2002) Current concepts in the management of Helicobacter pylori infection–the Maastricht 2–2000 Consensus Report. Aliment Pharmacol Ther 16: 167–180.
  22. 22. Barkun A, Crott R, Fallone C, Kennedy W, Lachaine J, et al. (2010) A one-year economic evaluation of six alternative strategies in the management of uninvestigated upper gastrointestinal symptoms in Canadian primary care. Can J Gastroenterol 24: 489–498.
  23. 23. Vakil N, Talley N, van Zanten SV, Flook N, Persson T, et al. (2009) Cost of detecting malignant lesions by endoscopy in 2741 primary care dyspeptic patients without alarm symptoms. Clin Gastroenterol Hepatol 7: 756–761.
  24. 24. Parkin DM, Bray FI, Devesa SS (2001) Cancer burden in the year 2000. The global picture. Eur J Cancer 37: S4–66.
  25. 25. Eisen GM, Dominitz JA, Faigel DO, Goldstein JA, Kalloo AN, et al. (2001) The role of endoscopy in dyspepsia. Gastrointest Endosc 54: 815–817.
  26. 26. van Zanten SJOV, Flook N, Chiba N, Armstrong D, Barkun A, et al. (2000) An evidence-based approach to the management of uninvestigated dyspepsia in the era of Helicobacter pylori. Canadian Medical Association Journal 162: S3–S23.
  27. 27. (1998) American Gastroenterological Association medical position statement: Evaluation of dyspepsia. Gastroenterology 114: 579–581.
  28. 28. Kapoor N, Bassi A, Sturgess R, Bodger K (2005) Predictive value of alarm features in a rapid access upper gastrointestinal cancer service. Gut 54: 40–45.
  29. 29. Wallace MB, Durkalski VL, Vaughan J, Palesch YY, Libby ED, et al. (2001) Age and alarm symptoms do not predict endoscopic findings among patients with dyspepsia: a multicentre database study. Gut 49: 29–34.
  30. 30. Vakil N, Moayyedi P, Fennerty MB, Talley NJ (2006) Limited value of alarm features in the diagnosis of upper gastrointestinal malignancy: systematic review and meta-analysis. Gastroenterology 131: 390–401; quiz 659–360.
  31. 31. Hammer J, Eslick GD, Howell SC, Altiparmak E, Talley NJ (2004) Diagnostic yield of alarm features in irritable bowel syndrome and functional dyspepsia. Gut 53: 666–672.
  32. 32. Voutilainen M, Mantynen T, Mauranen K, Kunnamo I, Juhola M (2005) Is it possible to reduce endoscopy workload using age, alarm symptoms and H. pylori as predictors of peptic ulcer and oesophagogastric cancers? Dig Liver Dis 37: 526–532.
  33. 33. Fransen GA, Janssen MJ, Muris JW, Laheij RJ, Jansen JB (2004) Meta-analysis: the diagnostic value of alarm symptoms for upper gastrointestinal malignancy. Aliment Pharmacol Ther 20: 1045–1052.
  34. 34. Dickerson LM, King DE (2004) Evaluation and management of nonulcer dyspepsia. Am Fam Physician 70: 107–114.
  35. 35. Axon AT, Bell GD, Jones RH, Quine MA, McCloy RF (1995) Guidelines on appropriate indications for upper gastrointestinal endoscopy. Working Party of the Joint Committee of the Royal College of Physicians of London, Royal College of Surgeons of England, Royal College of Anaesthetists, Association of Surgeons, the British Society of Gastroenterology, and the Thoracic Society of Great Britain. BMJ 310: 853–856.
  36. 36. Hallissey MT, Allum WH, Jewkes AJ, Ellis DJ, Fielding JW (1990) Early detection of gastric cancer. BMJ 301: 513–515.
  37. 37. Delaney BC, Wilson S, Roalfe A, Roberts L, Redman V, et al. (2000) Cost effectiveness of initial endoscopy for dyspepsia in patients over age 50 years: a randomised controlled trial in primary care. Lancet 356: 1965–1969.
  38. 38. Ford AC, Axon AT (2010) Epidemiology of Helicobacter pylori infection and public health implications. Helicobacter 15: 1–6.
  39. 39. Babaei M, Pourfarzi F, Yazdanbod A, Chiniforush MM, Derakhshan MH, et al. (2010) Gastric cancer in Ardabil, Iran–a review and update on cancer registry data. Asian Pac J Cancer Prev 11: 595–599.
  40. 40. Malekzadeh R, Derakhshan MH, Malekzadeh Z (2009) Gastric cancer in Iran: epidemiology and risk factors. Arch Iran Med 12: 576–583.
  41. 41. Kolahdoozan S, Sadjadi A, Radmard AR, Khademi H (2010) Five common cancers in Iran. Arch Iran Med 13: 143–146.
  42. 42. Sadjadi A, Nouraie M, Mohagheghi MA, Mousavi-Jarrahi A, Malekezadeh R, et al. (2005) Cancer occurrence in Iran in 2002, an international perspective. Asian Pac J Cancer Prev 6: 359–363.
  43. 43. Mohagheghi MA, Mosavi-Jarrahi A, Malekzadeh R, Parkin M (2009) Cancer incidence in Tehran metropolis: the first report from the Tehran Population-based Cancer Registry, 1998–2001. Arch Iran Med 12: 15–23.
  44. 44. Malekzadeh R, Sotoudeh M, Derakhshan MH, Mikaeli J, Yazdanbod A, et al. (2004) Prevalence of gastric precancerous lesions in Ardabil, a high incidence province for gastric adenocarcinoma in the northwest of Iran. J Clin Pathol 57: 37–42.
  45. 45. Derakhshan MH, Yazdanbod A, Sadjadi AR, Shokoohi B, McColl KE, et al. (2004) High incidence of adenocarcinoma arising from the right side of the gastric cardia in NW Iran. Gut 53: 1262–1266.
  46. 46. Barazandeh F, Yazdanbod A, Pourfarzi F, Ghajarieh Sepanlou S, Derakhshan MH, et al. (2012) Epidemiology of Peptic Ulcer Disease: Endoscopic Results of a Systematic Investigation in Iran. Middle East Journal of Digestive Disease 4: 90–96.
  47. 47. Spechler SJ (1999) AGA technical review on treatment of patients with dysphagia caused by benign disorders of the distal esophagus. Gastroenterology 117: 233–254.
  48. 48. Schroeder BM (2003) Evaluation of epigastric discomfort and management of dyspepsia and GERD. Am Fam Physician 68: 1215–1216, 1219–1220.
  49. 49. Bramble MG, Suvakovic Z, Hungin AP (2000) Detection of upper gastrointestinal cancer in patients taking antisecretory therapy prior to gastroscopy. Gut 46: 464–467.
  50. 50. Deeks JJ, Altman DG (2004) Diagnostic tests 4: likelihood ratios. BMJ 329: 168–169.
  51. 51. Steyerberg EW, Bleeker SE, Moll HA, Grobbee DE, Moons KG (2003) Internal and external validation of predictive models: a simulation study of bias and precision in small samples. J Clin Epidemiol 56: 441–447.
  52. 52. Steyerberg EW, Harrell FE Jr, Borsboom GJ, Eijkemans MJ, Vergouwe Y, et al. (2001) Internal validation of predictive models: efficiency of some procedures for logistic regression analysis. J Clin Epidemiol 54: 774–781.
  53. 53. Steyerberg EW, Vickers AJ, Cook NR, Gerds T, Gonen M, et al. (2010) Assessing the performance of prediction models: a framework for traditional and novel measures. Epidemiology 21: 128–138.
  54. 54. Grund B, Sabin C (2010) Analysis of biomarker data: logs, odds ratios, and receiver operating characteristic curves. Curr Opin HIV AIDS 5: 473–479.
  55. 55. Vickers AJ, Cronin AM, Begg CB (2011) One statistical test is sufficient for assessing new predictive markers. BMC Med Res Methodol 11: 13.
  56. 56. Cook NR (2007) Use and misuse of the receiver operating characteristic curve in risk prediction. Circulation 115: 928–935.
  57. 57. Cook NR (2008) Statistical evaluation of prognostic versus diagnostic models: beyond the ROC curve. Clin Chem 54: 17–23.
  58. 58. Pencina MJ, D'Agostino RB, Vasan RS (2010) Statistical methods for assessment of added usefulness of new biomarkers. Clin Chem Lab Med 48: 1703–1711.
  59. 59. Armstrong D, Bennett JR, Blum AL, Dent J, De Dombal FT, et al. (1996) The endoscopic assessment of esophagitis: a progress report on observer agreement. Gastroenterology 111: 85–92.
  60. 60. Numans ME, van der Graaf Y, de Wit NJ, de Melker RA (2001) How useful is selection based on alarm symptoms in requesting gastroscopy? An evaluation of diagnostic determinants for gastro-oesophageal malignancy. Scand J Gastroenterol 36: 437–443.
  61. 61. Bai Y, Li ZS, Zou DW, Wu RP, Yao YZ, et al. (2010) Alarm features and age for predicting upper gastrointestinal malignancy in Chinese patients with dyspepsia with high background prevalence of Helicobacter pylori infection and upper gastrointestinal malignancy: an endoscopic database review of 102,665 patients from 1996 to 2006. Gut 59: 722–728.
  62. 62. Nasseri-Moghaddam S, Razjouyan H, Alimohamadi SM, Mamarabadi M, Ghotbi MH, et al. (2007) Prospective Acid Reflux Study of Iran (PARSI): methodology and study design. BMC Gastroenterol 7: 42.
  63. 63. Chen LQ, Hu CY, Ghadirian P, Duranceau A (1999) Early detection of esophageal squamous cell carcinoma and its effects on therapy: an overview. Dis Esophagus 12: 161–167.
  64. 64. Ford AC, Qume M, Moayyedi P, Arents NL, Lassen AT, et al. (2005) Helicobacter pylori “test and treat” or endoscopy for managing dyspepsia: an individual patient data meta-analysis. Gastroenterology 128: 1838–1844.