E0195 - Algorithm Training and Independent Test Set Performance for a Molecular Non-Endoscopic Test for Detection of Esophageal Adenocarcinoma and Barrett’s Esophagus in Multicenter Cohorts

Introduction: Sedated endoscopy for Barrett’s esophagus (BE) and esophageal adenocarcinoma (EAC) detection is invasive and expensive. Non-endoscopic BE/EAC detection tools have been guideline-endorsed to facilitate higher patient participation at lower cost. We previously described a promising panel of 5 methylated DNA markers (MDMs) assayed on esophageal specimens obtained by a sponge-on-a-string (SOS) cell collection device in phase II studies. We aimed to train an algorithm (establishing marker cut offs, to adjudicate samples as positive/negative) using a final MDM panel followed by testing in an independent sample set.

Methods: Algorithm training samples (N=352) were prospectively collected from patients seen at 6 US medical centers. Test samples (N=125) were obtained from an independent, NIH-funded study conducted at 3 US medical centers. Cases had columnar metaplasia with intestinal metaplasia; controls had no endoscopic evidence of BE. Histology was reviewed by expert GI pathologists. The SOS device (25 mm, 10 ppi) was swallowed and withdrawn after 6-8 minutes followed by criterion standard endoscopy within 24 hours. DNA was extracted and bisulfite treated. Five MDMs were blindly assayed using the long probe quantitative amplified signal method. The algorithm was set using cross-validated logistic regression. The locked algorithm was applied to assay results from the test set.

Results: Baseline characteristics of patients in training and test sets were comparable (Table). The final panel included 3 MDMs (NDRG4, VAV3, and ZNF682) and a reference marker (B3GALT6). Overall sensitivity for BE/EAC detection in the training set was 81% (95% CI 76-88%) at 90% (85-94%) specificity. Overall BE/EAC sensitivity in the test set was 88% (78-94%) at 84% (70-93%) specificity. Sensitivity for HGD and EAC was 100% in the training and test sets. Sensitivity for short segment NDBE in the test set was 63% (38-84%). Areas under the receiver operating characteristic (AUROC) curve for BE/EAC detection were 0.92 (95% CI 0.89-0.95) and 0.94 (0.90-0.98) in the training and test sets, respectively (Figure). The algorithm was not influenced by age, sex, or smoking history. 97% of participants in the training set and 85% in the test set successfully swallowed the SOS device, which was well tolerated and safe.

Discussion: A 3-MDM panel for BE/EAC detection demonstrated excellent sensitivity for high risk BE cases in multi-center case control training and test sets.