Endoscopic Surveillance of Intestinal Metaplasia of the Esophagogastric Junction: A Decision Modeling Analysis.

We developed a state-transition microsimulation model using TreeAge Pro 2023 (©TreeAge, Willamstown, MA). Health states in the natural history model included EGJIM, low-grade dysplasia (LGD), high-grade dysplasia (HGD), pre-clinical cancer, symptomatic cancer (stages I through IV), and death ( Figure 1A ). The surveillance arms included the above health states with the addition of the following: complete eradiation of intestinal metaplasia (CE-IM) after EET, recurrence states (IM, LGD, HGD), and T1a cancer after endoscopic resection (ER). Possible causes of death included age-related mortality, cancer-related mortality, and mortality related to procedures (gastrectomy and endoscopy). We simulated a base-case scenario of EGJIM detected in 50-year-old individuals, who were followed until age 80 years or death. All individuals were assumed to have a correct diagnosis of EGJIM at entry into the model. The cycle length was 1 month; within each cycle, a patient could stay in the same state, progress to a more advanced state, or die from age-related mortality. The following endoscopic surveillance strategies for EGJIM were evaluated ( Figure 1B ): (1) No surveillance (standard of care, NSRV), (2) surveillance every 3 years for EGJIM (Q3SRV), (3) surveillance every 5 years for EGJIM (Q5SRV), (4) one-time surveillance at 3 years (stop surveillance if no dysplasia detected, ONESRV). Calibration targets for transition probabilities for EGJIM progression to cancer based on published literature and expert consensus are provided in table 1 . In the surveillance strategies, individuals underwent endoscopic surveillance of EGJIM at the intervals specified above. If subsequent examinations identified more advanced pathology, the surveillance intervals would be adjusted accordingly based on the American College of Gastroenterology guidelines for Barrett’s Esophagus 5 ( Supplemental Table 2 ). For instance, patients with LGD underwent surveillance EGD at 6 and 12 months from diagnosis, then annually, without endoscopic therapeutic intervention unless more advanced pathology was diagnosed. Patients diagnosed with HGD underwent endoscopic eradication therapy (EET), including endoscopic resection of visible nodules, followed by up to 2 sessions of radiofrequency ablation (RFA). EET could be ineffective, and after EET, patients underwent surveillance at 3-month intervals with “touch up” RFAs for 1 year. Those with remaining IM, low- or high-grade dysplasia after 3 rounds of “touch up” RFAs progressed along the natural history according to their pathologic state, with surveillance according to the post-EET schedule (surveillance at 3, 6, 12 months, then annually thereafter). After CE-IM there was fixed annual recurrence rate; recurrence probabilities included all states from IM to preclinical cancer. Patients underwent repeat EET if recurrent HGD was found outside of the “touch up” RFA period, regardless of prior treatment history. Patients who developed cancer underwent surgery according to Surveillance, Epidemiology and End-Results (SEER) database-derived rates for each stage. For patients with early-stage cancer (T1a), there was a possibility to undergo endoscopic resection or esophagectomy. If ER achieved curative resection, patients underwent surveillance at 3, 6, 12 months from ER procedure then annually. If ER was not curative, patients were considered for surgery. If recurrent cancer was found on endoscopy, patients could undergo repeat endoscopic resection and re-enter the surveillance program or undergo surgery. Patients with cancer stage T1b or above were modeled as surgically resectable or unresectable based on SEER rates. The absorbing state was age-related all-cause mortality as derived from the 2019 Center for Disease Control and Prevention United States Life Tables 27 or cancer-related death by stage as derived from literature. Patients with T1a cancer after ER were subject to all-cause mortality unless developed a recurrence of cancer. Transition probabilities derived from literature are outlined in Supplemental Table 1 . There is considerable uncertainty regarding the natural history of EGJIM, with key challenges including indistinct anatomic landmarks of the EGJ on endoscopy and poor interobserver reliability in rating length of columnar mucosa in the tubular esophagus for lesions <1 cm. 11 We therefore calibrated the natural history model to a range of transition probabilities drawn from robust BE literature and limited EGJIM literature to represent this uncertainty ( Table 1 ). We considered multiple studies for Scenario A and calibrated to Pohl et al. 24 (annual progression 0.01%) as low numbers of patients with EGJIM and relatively short follow-up time was limiting given the low rate of EGJAC reported. 9 , 10 Scenario C (0.12%), representing a pathology registry-based study by Hvid-Jensen et al. is likely to overestimate the progression of EGJIM as it includes all lengths of BE in the estimate. 26 We therefore modeled an additional scenario B (0.05%), representing a lower progression rate between scenario A and scenario C. To model heterogeneity of risk in different subpopulations within the US and determine the upper limit of progression, we calibrated Scenario D (0.22%) to a prospective study by Gu et al. 25 * This study was conducted in Linzhou, China, an area of high EGJAC incidence within China, with high H. pylori infection rate, 28 , 29 which likely represents a different mechanism the scenarios outlined above ( H. pylori -induced carditis vs reflux). Target lifetime incidence of cancer were agreed upon with reference to SEER EAC incidence, considering competing risks of death. We used calibration to identify a set of transition probabilities to ensure the model outputs aligned with our two cancer incidence targets: 10-year and lifetime incidence. We used the simulated annealing algorithm, which is an iterative process that selects random parameters within a set range. 30 To determine the suitability of the selected parameters, the chi-squared goodness of fit was the metric used to compare model outputs with our calibration targets. We generated a set of transition probabilities for each of our four of cancer incidence scenarios (A through D). As an additional check, we compared our outputs to the ranges of transition probabilities and treatment efficacy reported for EAC by the National Cancer Insitute Cancer Intervention and Surveillance Modelling Network (CISNET). 31 Costs and utilities from published literature are indicated in Supplemental Table 1 . Costs of cancer care inputs were based on treatment costs by phase of care and treatment modality for esophageal cancer as reported by Tramontano et al., 32 originally derived from the sum of Medicare, coinsurance, copayments and deductibles from the SEER-Medicare database to estimate direct cost. Phase of care was defined in months (30 days); after diagnosis of cancer, patients would undergo a 1-month surgery phase or staging phase (for those not undergoing surgery), followed by an initial phase (6 months), continuing phase (varying length until death), and a terminal phase (6 months before death). 32 All costs were adjusted to 2020 US dollars, converted using the Medical Care Consumer Price Index. 33 Both costs and utilities were discounted at a standard annual rate of 3%. 34 Incremental cost-effectiveness ratio (ICER) and quality-adjusted life years (QALY) of competing surveillance strategies were the primary outcomes. A wiliness to pay (WTP) threshold of $100,000/QALY was used to determine cost-effectiveness. 35 , 36 Secondary outcomes of this study included the number of EGDs per patient, number of incident cancer and cancer deaths in the cohort. The base-case cohort included 50-year-old males and females with EGJIM simulated over 30 years or until death with a 1-month cycle length, with microsimulation performed to 100,000 iterations. We report the cost effectiveness of the strategies from a healthcare perspective. Given the uncertainty in key transition probabilities from EGJIM to cancer, we performed four base-case analyses corresponding to four different progression rates as outlined in Table 1 , with other transition probabilities as outlined in Supplemental Table 1 . To investigate the effects of parameter uncertainty in the model, we performed both one-way and probabilistic sensitivity analyses. One-way sensitivity analyses vary one parameter at a time across a range of values while keeping all other parameters constant at their base-case value ( Supplemental Table 3 ). We modeled the variability in the availability of ER and also the likelihood that a lesion would be amenable to ER. We used the 95% confidence intervals outlined in source literature or ± 50% of the base-case value when these data were not available. Additional probabilistic sensitivity analyses were performed to test the effect of varying parameters simultanously. To gain additional insight into the optimal strategy within our WTP threshold, model input variables were assigned specific probability distributions ( Supplemental Table 3 ), and 1,000 second-order trials were performed.

Cost-effectiveness results of our outlined strategies are shown in Table 2 . The cost-effectiveness of endoscopic surveillance was dependent on EGJIM progression rate to cancer. In the lowest incidence scenario (scenario A, 0.01%), surveillance resulted in a loss of life expectancy due to the mortality risks of endoscopic procedures; QALYs decreased with intensity of surveillance. No surveillance strategies were cost-effective under this scenario. In intermediate-incidence scenarios (scenario B, 0.05% and scenario C, 0.12%), surveillance resulted in minor gain in life expectancy overall. However, only ONESRV was cost-effective under these scenarios, with ICERs of $30,989 and $16,525 per QALY gained for scenarios B and C, respectively. All other strategies had ICERs which exceeded the WTP threshold. Notably there was a decrease in QALYs for Q3SRV in scenario B, again due to risks of endoscopic surveillance. For Scenario C, Q3SRV had a relative gain of QALY compared to ONESRV (representing gains from prevention of cancer), however, the ICER exceeded the WTP threshold. For the highest incidence rate (scenario D, 0.22%), NSRV was eliminated (absolute dominance) due to costs of cancer care. Both ONESRV and Q5SRV were cost-effective options on the efficiency frontier, with Q5SRV demonstrating QALY gains of 0.05 and ICER of $7,695 per QALY, as compared to ONESRV. Q3SRV exceeded the WTP threshold. Across all strategies, the number of EGDs increased with intensity of surveillance, but the lifetime incidence of cancer and number of cancer deaths depended on the progression rate. For Scenario A (0.01%), the overall lifetime incidence of cancer and cancer deaths were low at 2–3 per 1,000 individuals. There was no notable difference in these outcomes according to intensity of surveillance. For Scenario B (0.05%), lifetime incident cancer and cancer deaths decreased with intensity of surveillance. However, these differences were small even when comparing maximal surveillance (Q3SRV, which was dominated) and NSRV, with a difference of 2 incident cancers and 2 cancer deaths per 1,000 individuals, achieved with 1.8 EGDs per individual. For Scenario C, the effect of surveillance in cancer incidence was greater overall. For Q3SRV vs NSRV, 8 fewer incident cancers and 8 fewer cancer deaths per 1,000 were observed, though Q3SRV exceeded the WTP threshold. ONESRV, which was under the WTP threshold, resulted in 4 fewer incident cancers and 5 fewer cancer deaths per 1,000, compared to NSRV. This was achieved by 2 EGDs per individual. The effect of surveillance on cancer incidence was most pronounced for Scenario D (0.22%), with maximal difference of 25 incident cancers and 23 cancer deaths per 1,000 for Q3SRV vs NSRV (dominated). For ONESRV, which was within the WTP threshold, 11 fewer incident cancer and 12 fewer cancer deaths were observed in comparison to NSRV, with 2.3 EGDs per individual. Q5SRV was also within the WTP threshold, and resulted in 20 fewer incident cancers and 8 fewer cancer deaths per 1,000 individuals in comparsion with ONESRV, with 8.3 EGDs per individual. Univariate sensitivity analysis demonstrated that the results were generally most sensitive to the risk and cost of EGD, disutility of EGD, and probability of recurrence of IM after endoscopic eradication therapy (RFA +/− EMR) ( Supplemental Figures 1 – 4 ). For Scenarios A-C, the results of the analysis at WTP $ 100,000/QALY did not change within the bounds of the sensitivity analysis. For Scenario D (0.22%), NSRV remained dominated in comparison to ONESRV within the range of the sensitivity analysis. Q5SRV under Scenario D was generally cost-effective, but the strategy exceeded the WTP threshold at higher rates of recurrence of preneoplasia after complete eradication of IM, and higher rates of EGD complication. For Q3SRV under Scenario D, the strategy was generally not cost effective, but could be under the WTP threshold at the lower bounds of cost of EGD ( Supplemental Figure 4 ). Cost-effectiveness acceptability curves that illustrate results from our probabilistic sensitivity analyses are shown in Supplemental Figure 5 . In scenario A (0.01%), NSRV was the preferred strategy in all simulations at WTP $100,000 per QALY gained. For scenario B (0.05%), and scenario C (0.05%), ONESRV was the preferred strategy across all iterations for WTP of $100,000 per QALY gained. NSRV was the preferred strategy only when the WTP threshold was under $25,000 for scenario B and under $15,000 for scenario C. In scenario D (0.22%), Q5SRV was the preferred strategy in the majority of iterations (67%) at our WTP threshold of $100,000 per QALY gained; Q3SRV was the preferred strategy in 26% and ONESRV was preferred in 11% of iterations.

Our study demonstrates that endoscopic surveillance is more cost effective as the incidence rate increases. In the range of estimates for the low-incidence US population (Scenarios A-C), our study suggests that no surveillance (Scenario A) or one-time endoscopic surveillance of EGJIM at 3 years with no further surveillance unless dysplasia is found (Scenarios B-C), was cost-effective. Our analysis found that longitudinal surveillance every 3 or every 5 years, as recommended for Barrett’s esophagus (which may be applied in practice to the clinical scenario of the ‘irregular Z-line’) is generally not cost-effective for EGJIM. In lower incidence scenarios, increased intensity of surveillance resulted in a progressive decrease of QALYs compared to one-time surveillance, representing the harms of endoscopic surveillance. As cancer incidence increased, as seen in the East Asian population (Scenario D), longitudinal but less frequent surveillance (every 5 years) became the preferred strategy. Similarly, the effect of surveillance intensity on the number of cancers prevented and cancer deaths increased with cancer incidence. In Scenario D, the greatest reduction in incident cancer with the preferred strategy of 5-yearly surveillance resulted in 21 fewer incident cancers and 20 fewer cancer deaths per 1,000 individuals with EGJIM as compared to our current standard of care, no surveillance. To contextualize these findings, we first address the limitations of our analysis. The limited data on the natural history of EGJIM is a chief concern, and for this reason, we chose to model four different incidence scenarios as outlined above. Although a model’s validity increases as the estimates for model inputs, such as the progression rate to cancer, are more certain, we performed this analysis as the clinical decision regarding surveillance is being made in thousands of patients in real time with limited quantitative analyses to inform the decisions and to underscore the pressing need for clinical studies. Our scenarios extend from ultra-short (<1 cm) Barrett’s esophagus in a Western population 24 to cardia IM in a Chinese population 25 . We also note that the modeled strategies are extended from guidelines of management of Barrett’s esophagus and expert consensus, and are not derived from directly observed data of an EGJIM cohort. However, these guidelines form the best available framework for surveillance available to clinicans. There is likely heterogeneity in the underlying biology and carcinogenesis of IM in EGJ/cardia by global region, and H. pylori infection may be playing a large role. For scenario D, which forms our upper bound of incidence, our model is calibrated to a data from Gu et al, 25 a large population-based surveillance study of gastric cardia preneoplasia in an area of China with high H. pylori prevalence. 28 , 29 The causal relationship between H. pylori and EGJAC is unclear, and prevalence in the U.S. population is heterogenous. 37 , 38 The highest incidence of esophageal and cardia gastric cancer in the U.S. is in the non-Hispanic white population, suggesting that a different disease mechanism, such as reflux, may be in play. Thus, Scenario D is unlikely to apply to the U.S. population at large but may be relevant to certain populations within the U.S. such as East Asians, perhaps extending to those who may harbor higher rates of H. pylori infection, such as race, ethnic minority and non-U.S. born individuals. The progression rate for patients with EGJIM in the U.S. is likely reflected in Scenario A to B. Given the diversity of race, ethnicity and nativity within the U.S., we acknowledge significant heterogeneity in EGJAC incidence in subpopulations, and these groups may be represented by the other scenarios. The advanced stage at diagnosis, high case fatality, and rising incidence of EGJAC suggest a need for greater prevention efforts. In contrast to EAC, patients diagnosed with EGJAC are less likely to have concurrent BE in contrast to patients with EAC (25% vs 57%); thus current prevention efforts which focus on BE are likely to fall short for EGJAC. 39 Our model shows that given the high costs of cancer care, high-cost endoscopic surveillance can be cost-effective even in lower-incidence scenarios. Based on our model, we suggest that for patients from a high-risk (East Asian) population, longitudinal surveillance (every 5 years) be considered ( Figure 2 ). For the US population at large, either no surveillance or one-time endoscopy can be considered, however, more data is required on progression rates to identify the optimal strategy. Endoscopists may wish to consider the quality of the initial exam and EGJAC risk factors in offering one-time surveillance. In Barrett’s esophagus, the risk of progression is highest in the first follow up endoscopy, which likely represents prevalent cancers that were not noted on the prior endoscopy, thus a shorter-interval for the first surveillance endoscopy (3 over 5 years) may be suggested for EGJIM. 40 This represents the first published model of EGJIM, informing the practice for a prevalent population that is generally excluded from Barrett’s esophagus studies. Our extensive sensitivity analyses demonstrate that our results are robust. As the costs of cancer care and survivorship rise, preventative efforts may become relatively more cost-effective strategies in the future. 41 , 42 Further studies to determine the natural history and progression rate of EGJIM are necessary. Differences in progression by clinical or demographic variables are not known. As more data become available for incidence rate and other key variables, reassessment of this model will be required, and these data will allow more precise modeling of sub-groups of interest. Of particular interest is extension of the model to refine recommendations by sex; in the US, incidence of EGJAC is 4-fold greater in men compared to women, and the most effective and cost-effective strategies may differ. 43 Given high estimated prevalence of EGJIM, careful consideration and further study of impacts on service provision and costs from a societal perspective are required. The diagnosis of EGJIM and the characterization as a precancerous state (regardless of the overall low risk of cancer) can be a source of anxiety for patients. In patients with non-dysplastic Barrett’s, perceived risk of EAC decreased patients’ health-related quality of life (utility scores 0.91 – 0.95). 44 , 45 We included the disutility of EGD surveillance in our model, but the disutility of the perceived risk of EGJIM is not known and requires further study. In conclusion, the cost-effectiveness of endoscopic surveillance increases with increasing incidence of EGJAC. Our analysis demonstrates that in populations with higher EGJAC incidence (East Asian), low-intensity longitudinal endoscopic surveillance is cost-effective. In low-incidence populations (such as the US population at large), endoscopic surveillance is a costly practice and can cause harm. Both no surveillance or one-time endoscopic surveillance was cost-effective within the range of our analysis, and additional data is required. Future studies are needed to better understand the natural history of EGJIM, identify risk factors for progression, and inform appropriate surveillance strategies in clinical practice.

The incidence of adenocarcinoma of the esophagogastric junction (EGJ) and gastric cardia has increased dramatically in past decades, with poor survival attributed to advanced disease at presentation. 1 – 3 EGJ intestinal metaplasia (EGJIM) can be defined as intestinal metaplasia on biopsies of the EGJ, at or just distal to the squamocolumnar junction, without endoscopic finding of Barrett’s esophagus (BE). 4 The description of EGJIM in literature is not uniform. Heterogeneity stems from indistinct anatomic landmarks defining the EGJ on endoscopy, which can change with the degree of insufflation, respiratory motion, hiatal hernia, gastric and esophageal motility. The anatomic EGJ may not overlie the histologic squamocolumnar junction, even as society guidelines define BE as the finding of intestinal metaplasia (IM) in the tubular esophagus in relation to endoscopic landmarks (visible columnar mucosa extending at least 1 cm from the EGJ). 5 – 8 The anatomic definition of the gastric cardia (2 cm distal to the EGJ) is similarly arbitrary, difficult to distinguish endoscopically, and may not correlate with histologic borders. In this context, the “irregular Z-line” is an imprecise term commonly used to describe endoscopic findings which fail to meet the threshold of BE while conveying the uncertainty of the diagnosis. An “irregular Z-line” is prevalent in 5–20% of patients undergoing endoscopy, 9 and biopsies are not recommended in this setting given poor interobserver reliability of Prague Criteria at <1 cm and data that demonstrate lower risk of adenocarcinoma in these patients compared to their BE counterparts. 9 – 11 Nevertheless, biopsies are common in practice. The true prevalence of EGJIM in the general population is not known, however, in patients referred for EGD, the prevalence of EGJIM in the absence of endoscopically visible BE is estimated at 6–23%. 12 – 19 Prevalence may be higher (15–48%) among patients who have squamocolumnar junction abnormalities that do not meet the definition of BE (serrated Z-line, mucosal islands, cephalad displacement <1cm). 18 , 20 There is increasing recognition that adenocarcinomas arising around the EGJ have a common origin, arising on a background of chronic inflammation and intestinal metaplasia, and share histologic, immunophenotypic, and epidemiologic features which suggest a continuum rather than distinct tumor types. 21 – 23 In this context, when EGJIM is found there is uncertainty in its management. The diagnosis of EGJIM commonly leads to endoscopic surveillance and patient anxiety. Given these key uncertainties, the aim of our study was to develop and analyze a decision-analytic model to evaluate the cost-effectiveness of endoscopic surveillance in patients with EGJIM. We modeled the natural history of EGJIM using defined ‘scenarios’, referencing literature ranging from ultrashort-segment BE in the West to cardia IM in a Chinese population with high incidence of esophagogastric junction adenocarcinoma (EGJAC). 24 – 26 We hypothesized that given the potential for early detection of EGJAC with endoscopic surveillance, certain intervals for surveillance could be cost effective, even at low progression rates.

Supplemental Table 1: Model input parameters Supplemental Table 2: Summary of endoscopic surveillance intervals Supplemental Table 3: Sensitivity analysis parameters and distributions. Supplemental Figure 1: Tornado diagram on one-way sensitivity analysis using select model parameters: Scenario A (0.01% 10-year incidence) Supplemental Figure 2: Tornado diagram on one-way sensitivity analysis using select model parameters: Scenario B (0.5% 10-year incidence) Supplemental Figure 3: Tornado diagram on one-way sensitivity analysis using select model parameters: Scenario C (1% 10-year incidence) Supplemental Figure 4: Tornado diagram on one-way sensitivity analysis using select model parameters: Scenario D (2.2% 10-year incidence) Supplemental Figure 5: Probabilistic sensitivity analysis by incidence scenario