Risk Prediction Models for Barrett's Esophagus Discriminate Well and Are Generalizable in an External Validation Study.

backgroundBarrett's esophagus is the precursor to the highly lethal esophageal adenocarcinoma. Risk prediction models have been developed to assist in its detection, potentially improving early identification and treatment of esophageal adenocarcinoma. Six models have been developed.

aimsTo externally validate three models (Rubenstein, Thrift, and Baldwin-Hunter models) and compare them to a fourth risk prediction model (Ireland model) for Barrett's esophagus.

methodsData from 120 Barrett's cases and 235 population controls were available to externally validate the three models. Discriminatory ability of these models was assessed by the area under the receiver operating characteristic curve. Calibration was assessed with the calibration slope, Hosmer-Lemeshow test, and Lowess smoother calibration plot. Following external validation, diagnostic accuracy of the three models was compared to that of the Ireland model.

resultsOn external validation, the Rubenstein model had an area under the receiver operating characteristic curve of 0.71 and was well calibrated (Hosmer-Lemeshow test, p = 0.67). Likewise, the Thrift and Baldwin-Hunter models had similar discrimination (0.71 and 0.70, respectively) and were also well calibrated (p = 0.69 and p = 0.28). Our previous external validation of the Ireland model provided an area under the receiver operating characteristic curve of 0.83 and was well calibrated (p = 0.14). The Ireland model demonstrated a statistically significantly greater area under the receiver operating characteristic curve than the Rubenstein (p = 0.02), Thrift (p = 0.001), and Baldwin-Hunter (p = 0.002) models.

conclusionWe externally validated the Rubenstein, Thrift, and Baldwin-Hunter risk prediction models and compared them to the Ireland model. The Ireland model demonstrated improved accuracy, albeit with slightly poorer calibration.