Interactive decision support for esophageal adenocarcinoma screening and surveillance

Incidence of esophageal adenocarcinoma (EAC) has risen markedly in many western countries. Most cases can be attributed to known risk factors, such as symptomatic gastroesophageal reflux (sGERD), central obesity, cigarette smoking and family history [1]. Nevertheless, the relative rarity of the cancer, combined with the cost and invasiveness of upper endoscopy for identifying early cancers and high-risk pre-cancers (e.g., Barrett’s esophagus (BE) with dysplasia or genomic abnormalities) [2, 3] make it challenging to define effective screening and surveillance strategies [4].

A key barrier has been identifying those most likely to benefit from endoscopy or newer non-endoscopic tissue sampling methods [5, 6]. Clinical practice guidelines vary by country and professional society regarding criteria for initial screening for BE or EAC, as well as the definition of BE [7, 8]. Furthermore, none consider the strong effect of age on EAC incidence except in defining a fixed age threshold, and all tend to treat the remaining risk factors as equally important. For example, 2016 American College of Gastroenterology (ACG) guidelines [7] suggest that screening may be considered among men with sGERD plus two or more other specified risk factors for BE or EAC. However, this qualitative approach excludes the approximately 47% of all EAC that present in persons without significant sGERD, who may be at increased risk due to other factors, and does not take advantage of known quantitative relationships (i.e., strength of association and dose-response) between EAC incidence and sGERD, smoking and obesity, for example [9].

To address this barrier, an online Interactive and Contextual Risk Calculator (IC-RISC™; https://​ic-risc.​esocan.​org) was developed to take advantage of existing knowledge from observational studies and clinical trials to estimate more precisely an individual’s absolute risk of developing EAC over a ten-year period, and to convey this estimate in the context of risk of dying from other cancers or from common causes such as injury, stroke or heart disease. Using this calculator, we compared the risk distribution between EAC cases and population-based controls from six studies in the Barrett’s and Esophageal Adenocarcinoma Consortium (BEACON; https://​beacon.​esocan.​org) to contrast discrimination metrics of current guidelines vs. more stringent thresholds that might be used.

The application interface consists of six tabs, including the calculator itself (Fig. 1.) Values for demographics, risk and preventive factors are entered on the left side of the page. Estimated 10-year probability of developing EAC is displayed on the right side in two ways: as a “thermometer” with color coded risk (on a log scale), along with an estimate of its uncertainty; and as a set of 1000 circles depicting the expected number who will (red fill) and will not (unfilled) develop the cancer in the next ten years. Any change in the predictors on the left is immediately reflected in the risk estimates on the right. The risk factors available for input and used in calculations depend on whether BE has been diagnosed. Figure 1 illustrates inputs and 10-year risk for an individual who has not been screened (BE status unknown) (Fig. 1a), and one who has been diagnosed with BE (Fig. 1b) (See legend for details.) Mortality rates for selected causes of death are displayed in the center of the thermometer; these are specific for age, sex and race (black/white), but do not consider non-demographic risk factors.

The landing page is the “About” tab describing the purpose and background. Tabs are available to calculate BMI and categorize usual physical activity, which are not usually known. The “Risk factors” tab includes three options for displaying the input data used: a graph (Additional file 1: Figure S1, of the RR estimates and 95% CIs for each factor, stratified by the presence of diagnosed BE (unknown/negative vs. positive); and detailed tables (Additional file 1: Tables S1 and S2, containing all the RR data with supporting references. Reliable risk factor data specific for black females and other demographic groups of either sex are not currently available. The “More information” tab contains an option for viewing the age-, race- and sex-specific incidence and mortality rates used in the calculator (stratified by BE status) and all-cause mortality (Additional file 1: Figure S2) Additional menu options include contacts, version history, and licensing information.

Figure 2 illustrates the distribution of 10-year risk estimates for EAC cases and controls using data from population-based BEACON studies, along with estimates of specificity, sensitivity and positive predictive value (ppv) according to increasingly more stringent thresholds for further action. The area under the associated ROC curve (not shown) is 0.81 (95% CI = 0.79–0.83.) When applied to the general population, as in Fig. 2, the calculator provides insights into the performance of current practice guidelines (two examples of which are represented by vertical solid lines, see legend) [7], which sacrifice a significant amount of specificity in order to capture 79–86% of eventual cases. A small increase in action threshold, for example to 2 per 1000, would potentially eliminate almost half of procedures or tests (reducing 40% false positives to 20%), while only reducing sensitivity to 66%; whereas an increase to 4 per 1000, would further reduce false positives to 12%, with a sensitivity of 49%. Unfortunately, given the rarity of EAC, only with higher thresholds of 8 per 100,000 or more does the ppv surpass 1%.

It is notable that absolute risks of greater than 2.0 are readily reached by persons not included in current guidelines. For example, a 65-year-old white female ex-smoker with frequent (“> daily”) reflux symptoms, no family history, and a BMI of 31 has an estimated risk of 4.4 per 1000. Similarly, a 60-year-old non-smoking white male without reflux symptoms (“rarely”), a BMI of 31 (obese category I) and one first degree relative with BE/EAC has an estimated risk of 4.8 per 1000. Finally, a 65-year-old ex-smoking black male with occasional reflux symptoms (“< weekly”), no family history, and a BMI of 36 (obese category II) has an estimated risk of 2.1 per 1000.

Given that almost half of EAC cases in the general population do not report significant symptoms of reflux [9], we investigated the performance of the risk calculator in the subset of the cases (46%) and controls (70%) from the BEACON studies who reported sGERD only occasionally (“< weekly”) or rarely (Additional file 1: Figure S3) The overall discriminatory ability in this subset was similar to that in the entire dataset, though slightly more modest, with an area under the ROC curve of 0.78 (95% CI = 0.75–0.81.) While a specificity of 95% was reached at a lower risk threshold than in the entire dataset (about 4 vs. 8 per 1000 over ten years), the sensitivity at the lower threshold in this low-sGERD subset was only half (25%) that of the entire dataset (50%), and the ppv did not rise above 0.5%.

This educational tool was developed to facilitate shared decision-making between a health provider and patient regarding: i) how personal risk of EAC fits into the “bigger picture” of health and disease, ii) whether preventive actions are indicated to possibly reduce risk of EAC and other conditions, and iii) whether additional tests and procedures might be warranted to identify and manage those with higher risk profiles [54].

A strength of IC-RISC™ is the robust data on which it is based. Incidence rates of EAC were based on the SEER 18 registries, which cover more than a quarter of the US population, and RR and prevalence estimates for the general population calculations came from large individual pooling efforts (including between 900 and 1500 cases). This contrasts with previous efforts in the general population which have both estimated associations and created risk models using relatively small individual studies, often with varying availability and comparability of potential predictors [42, 55, 56]. For example, in a prospective cohort study with over 350,000 participants, 220 developed EAC over up to eight years of follow up; BMI, smoking and a prior esophageal condition or treatment were found to be predictive and included in a risk score [56]. In another study of 189 EAC cases, final predictive variables included sGERD and/or use of anti-reflux medication, BMI, tobacco smoking, duration of living with a partner, previous diagnoses of esophagitis and diaphragmatic hernia, and previous surgery for esophagitis, diaphragmatic hernia or severe reflux or gastric or duodenal ulcer [55].

In persons with BE, the evidence base for predicting progression to EAC is more limited, and results often conflicting. Although many of the studies informing RR estimates are cohort studies, which have many strengths including internal consistency and lack of recall bias, the small number of outcomes for a rare disease such as EAC, varying designs, limited risk factor information and questions about generalizability remain important issues. For example, in a report on a non-HGD cohort with 154 EAC/HGD outcomes, three factors were observed to be significantly predictive: segment length, smoking and confirmed low-grade dysplasia [38]. Unexpectedly, this study also observed that progressors had significantly lower BMI (p = 0.012) and obesity (p = 0.049, RR = 0.68) than non-progressors, which is inconsistent with most observational studies [18, 55, 57]. In a smaller single institution study, proton-pump inhibitor use, segment length greater than 3 cm and history of esophageal candidiasis were the key predictors beyond demographics, but information on smoking and use of NSAIDs or statins was not available [57]. Finally, an analysis involving 103 EAC cases from the U.K. General Practice Research Database found overweight and statins as a significant predictors, but not smoking [18].

The present calculator has limitations. While confounding factors were adjusted for in the underlying studies, the possibility of interactions among the factors has not been adequately examined. Thus, the calculations assume the risk factors act multiplicatively. This is a particular concern at more extreme risk estimates, where supporting data may be quite sparse. Although survival from EAC is poor (about 20% at five years), not all diagnosed with the cancer will die from it. This should be kept in mind when comparing estimated incidence rates of EAC in the thermometer with comparison rates which are mortality based. In addition, while mortality rates are specific for persons in the same age, sex and race group, within this group they represent population averages as they are not adjusted for other risk factors. Thus, risk of death from lung cancer or heart disease, for example, would be higher than shown for a smoker and lower than shown for a non-smoker. Use of a separate disease-specific risk calculator with additional disease-specific risk factors would be needed for more precise estimates. The less abundant and reliable data on progression to EAC in persons with BE should be considered in interpreting the absolute risk estimates in this setting. Since the BEACON studies used in the Fig. 2 analyses also contributed between 32 and 44% of cases to meta-analyses regarding four risk factors, the discriminative metrics may have some level of optimism. The relatively low ppv indicates that identification of additional predictive biomarkers, likely blood- and cytology-based in the general population, and biopsy-based among those diagnosed with BE, is greatly needed to improve discrimination accuracy in both groups [12]. Finally, it is applicable only to the US population.

The accuracy of all risk calculators depends upon the quality of the underlying data and generalizability to the targeted population. We drew upon multiple studies of various designs in arriving at a set of RR estimates. For some factors, such as smoking and BMI in the general population, this is straightforward given the published meta-analyses. For other factors, especially in the setting of BE, this was a more subjective process with ample opportunity for disagreement among experts. We addressed this by designing an open-source self-documenting application, with input parameters contained in few simple spreadsheets, which can be downloaded and modified to meet the needs of an institution or clinician. As a corollary, as technology, clinical practice and epidemiologic knowledge evolve, IC-RISC™ can be improved and kept relevant. For example, predictors based on germline mutations can be added as panels become larger, more predictive and widely available [58]. As non-endoscopic sampling methods are introduced into clinical practice (e.g. Cytosponge, encapsulated balloon), together with tailored assays incorporating somatic genetic and epigenetic abnormalities, they can also be included. [58‐60] Its generic architecture also lends itself to being extended to other cancers with distinct pathways and/or intermediate stages, such as hepatocellular cancer, with discrete causes and rapidly developing screening and treatment modalities [61].

Decisions regarding EAC prevention faced by patients and providers are plentiful. For example, they might involve preventative actions centered on lifestyle (e.g., weight loss, exercise, dietary change), chemopreventative actions (e.g., aspirin), non-invasive screening, endoscopic screening, surveillance, and treatment of high-risk Barrett’s. Informed judgment plays an important role in most of these decisions since compelling evidence does not currently exist to support hard decision thresholds. If decision thresholds did exist, they would have to account for vastly different ratios of risk to likely benefit for each option, and large inter-individual differences (in the general population as well as providers) in risk tolerance and willingness to undergo medical interventions, among many other factors. In this context, IC-RISC™ provides a simple and rapid means to graphically communicate risk of EAC in the context of other health risks, allows “what-if” scenarios regarding potential preventative actions, and can directly inform discussions regarding screening, surveillance and treatment options. It extends current qualitative clinical practice guidelines for endoscopic screening into a quantitative assessment, which brings the possibility of screening being offered to higher-risk persons not currently targeted (e.g., males without sGERD and females), while also potentially identifying lower-risk persons who might otherwise be targeted for screening as unlikely to benefit. It also may be useful in defining higher-risk persons for intervention trials. The calculator would benefit from prospective testing and application to existing well-characterized cohort studies to refine the estimates and facilitate harmonizing existing risk calculators so that a uniform message can be developed for clinical use. Until such time, IC-RISC™ can be used to inform joint decision-making rather than to indicate specific preventative actions or thresholds.

Project name: IC-RISC.

Project home page: Software - https://​github.​com/​FredHutch/​IC-RISC-Working

Application – https://​ic-risc.​esocan.​org

Operating system: Platform independent.

Programming language: R/Shiny.

License: Academic use - 2-Clause BSD; Non-academic - License needed.