Predicting Progression in Barrett’s Esophagus: Is the Holy Grail Within Reach?

Abstract

Risk stratification in BE may allow aggressive management of those at “high risk” and reduction in surveillance in at “low” risk. Davidson et al report results of the independent validation of a multi-biomarker panel (Tissue Cypher assay) performed on biopsy tissue, in a case control study. “High risk” patients progressed 5 times more than those at “low risk.” Sensitivity and specificity for “high risk” patients were 29% and 86% with a PPV of 23%. NPV of a “low risk” score was 96%. These findings may allow more intensive surveillance in those at “high risk”. Despite some limitations this assay is a potentially major advance in the management of BE patients without dysplasia.

Patients with Barrett’s esophagus (BE) are placed on endoscopic surveillance to detect dysplasia or adenocarcinoma (EAC), which can be treated by endoscopic eradication therapy (EET). However, in those with no dysplasia (NDBE), the annual rate of progression to EAC is low (0.33%) (1). Although guidelines continue to recommend surveillance (1, 2), identification of those at higher risk may allow either intensive endoscopic surveillance (to detect incident dysplasia/EAC) or proactive EET, while surveillance intervals could be lengthened in those at low-risk. This risk prediction has been referred to as the “Holy Grail” in BE given the challenges in the development and validation of such scores, primarily due to the low rate of progression.

In this issue, Davison et al. validate the performance of a biomarker panel to predict progression to high grade dysplasia (HGD) and EAC in BE (3). Formalin-fixed biopsy tissue was obtained from patients in two medical centers. The risk score was derived from the TissueCypher BE assay, developed and validated in a previous study (4). Nine protein-based biomarkers (p16, p53, AMACR, HER2, cytokeratin 20, CD68, COX-2, HIF-1alpha, and CD45Ro), along with tissue architecture and nuclear morphology, are used in this assay to provide a numeric score (0–10) to predict the risk of progression. This assay has also been found to detect prevalent HGD and EAC (5). Cut-off thresholds for low, intermediate, and high-risk groups validated by Davison et al. were determined and locked down from these studies.

Fifty-eight progressors (BE cases with no dysplasia, indeterminate, or low grade dysplasia [LGD] diagnosed with HGD/EAC > 1 year after diagnosis), and 210 non-progressors (BE cases without progression for at least 5 years) were included in the study. Of note, most (75%) progressors developed HGD, not EAC. The sensitivity and specificity of a “high-risk” score for predicting progression was 29% and 86%, respectively, with a positive predictive value (PPV) of 23%. Those with a “high-risk” score were 4.7x more likely to progress compared to those at “low-risk”. In comparison, sensitivity and specificity for predicting progression using expert diagnosis of LGD were 19% and 88%, respectively, only 3.8x more likely to progress compared to those with NDBE. Risk scores were not separately calculated for those with LGD (N=18) or indeterminate dysplasia (N= 23). The authors proposed that patients scoring “high-risk” may benefit from proactive EET or intensive surveillance.

This was a well conducted study with many strengths. First, the test performance was assessed with a “locked-down” algorithm from prior studies, making this a true validation study in an independent sample. The study was adequately powered. Assays were run without clinical information (single-blinded), and followed an a priori analysis plan (4, 5). Multivariate analysis showed that the risk-score added independent predictive information beyond clinical variables (eg. male gender). Additionally, to account for spatial heterogeneity in dysplasia detection, the assay was done on multiple levels of biopsies. This revealed that scoring all levels of biopsies was more accurate than one level, as upstaging of the risk score occurred in almost 40% of progressors and 20% of non-progressors. This reveals that the assay remains susceptible to varying expression of markers across the BE segment. It is also unclear from the manuscript as to when these markers turn positive (i.e. how long before progression).

Although this was a rigorous validation study, the predictive performance in terms of progression appears to be modest. Thirty-one out of 51 progressors scored “low-risk” by the assay, rendering a sensitivity of 29%. While specificity was better (86%), PPV was only 23%. Hence any intervention may be unnecessary in almost 80% of those deemed at high risk. The prevalence adjusted NPV on the other hand was high at 96%. Additionally, the case-control nature of the study makes assessment of true incidence of progression challenging. While a prospective cohort study assessing the predictive accuracy of this test would be ideal, given the low rate of progression, it would be hard to perform.

To put these results in perspective, it is relevant to review the performance of other “BE progression predictor panels” in the literature. Using a panel of molecular markers detected by DNA fluorescence in situ hybridization (FISH) on esophageal brush cytology specimens collected in a prospective cohort of 428 patients with 22 progressors, Timmer et al. reported a relatively high sensitivity of 86% but a modest specificity of 54% in predicting progression (6). In another study, Jin et al. explored an 8 methylated-DNA-marker panel for BE progression prediction, and reported a sensitivity of 44% at a specificity of 90% (7). This study has not been externally validated. Finally, in a multi-institutional study with 2697 BE patients, Parasa et al. developed a progression in Barrett’s esophagus (PIB) Score, based only on clinical characteristics (8). The performance of this risk score was also modest, with a c-statistic of 0.76, Those categorized as high-risk, had a 12% cumulative risk of progression, which is lower than reported in the current study.

These results raise the broad question of what would be the acceptable performance of any BE prediction panel for clinical use? Do we emphasize sensitivity or specificity? What level of sensitivity and specificity would be acceptable before these tests are implemented? This would likely depend on the intervention which follows the test: intensive surveillance or proactive EET? EET is more expensive, associated with potential complications, and hence would merit higher sensitivity and specificity. In the current range of accuracy from multiple predictive tests, intensive surveillance is likely the more acceptable follow up option.

How should we use the findings from this important study in the management of patients with BE? With a PPV of only 23%, it would be hard to justify EET in a population where close to 80% may not benefit. Intensive surveillance may be more reasonable to detect missed dysplasia or EAC in those with a “high-risk” score. This could include not only (electronic) chromoendoscopy but also other technologies such wide area transepithelial sampling (WATS) or volumetric endomicroendoscopy. Equally important would be the identification of those at low-risk. If the NPV of such a test is high (> 90%) one could make an argument to reduce the frequency of surveillance in those with a “low-risk” score. Indeed, the prevalence adjusted NPV of the test was 96%. This may be another application of this test.

Therefore, while the current study is exciting in that the cut-offs established in prior studies for the Tissue Cypher test are validated, the performance characteristics likely need to be improved to justify proactive EET in those with a “high-risk” score. Intensive surveillance using adjunctive imaging technology is potentially a more justifiable intervention. Additionally, performance can likely be improved by using a more comprehensive sample of the BE segment (obtained via an endoscopic brush or a capsule sponge) given discordant results in biopsies from multiple levels, given the results from this study. While the test uses multiple biomarkers, combinations with either additional biomarker classes or a clinical score may improve performance. The multitarget stool DNA test is one example of a complementary combination of marker classes (9). MDMs for BE and EAC have been described (10, 11), and a similar discovery approach may yield markers to predict BE progression. An optimized BE progression prediction test has the potential to truly transform the landscape of BE management.