A new model that predicts Barrett’s esophagus (BE) risk relies on data that can be sourced from the electronic health record.
The tool was developed and validated by researchers seeking to improve early diagnosis of esophageal adenocarcinoma (EAC) through early detection of BE. Currently, 5-year survival from EAC is just 20%, and fewer than 30% of patients have curative options at diagnosis because of late-stage disease.
Several clinical guidelines recommend screening for BE in high-risk individuals. The trouble is that adherence is low: A meta-analysis of more than 33,000 patients found that 57% of newly diagnosed EAC patients had a simultaneous, first-time diagnosis of BE, which suggests missed screening opportunities. These patients also had higher probability of late-stage disease and worse mortality outcomes than patients who had a previous BE diagnosis. Low adherence to screening guidelines can be attributed to difficulties in implementation, as well as unsatisfactory outcomes in real-world settings. In a previous study, the researchers examined the efficacy of existing guidelines in diagnosing BE in a primary care population, and found all of these screening guidelines had a low area under the receiver operating curve ranging from 0.50 to 0.60.
The researchers sought to address these challenges in the current study published in Clinical Gastroenterology and Hepatology using a model development cohort of 274 patients with BE and 1,350 patients without BE. The patients were seen at Houston Veterans Affairs clinics between 2008 and 2012, and were between ages 40 and 80 years. The researchers included common risk factors identified among existing guidelines. The final model, Houston-BEST, incorporated sex, age, race/ethnicity, smoking status, body mass index, symptoms of gastroesophageal reflux disease (heartburn or reflux at least 1 day/week), and first-degree relative history of esophageal cancer.
The validation cohorts included patients from primary care clinics at the Houston VA who were undergoing screening colonoscopy (44 with BE, 469 without BE), as well as patients at the University of Michigan, Ann Arbor, or Ann Arbor VA clinics who presented for first esophagogastroduodenoscopy (71 with BE, 916 without BE).
In the development cohort, the researchers set an a priori threshold of predicted probability that corresponded to sensitivity of 90%; the threshold predicted probability of BE was 9.3% (AUROC, 0.69; 95% confidence interval, 0.66-0.72). The specificity was 39.9% (95% CI, 37.2%-42.5%). In the Houston area validation cohort, the model had a sensitivity of 84.1% (AUROC, 0.68; 95% CI, 0.60-0.76). The number needed to treat to detect a single BE case was 11.
Among the University of Michigan/Ann Arbor validation cohort, the model had an AUROC of 0.70 (95% CI, 0.64-0.76), but it had a sensitivity of 0%. The researchers also tested the ability of the model to discriminate early neoplasia from no BE, and found an AUROC of 0.72 (95% CI, 0.67-0.77).
The researchers tested other models based on existing guidelines in the Houston area cohort and found that their model performed at the high end of the range when compared with those other models (AUROCs, 0.65-0.70 vs. 0.58-0.70). “While the predictive performance of Houston-BEST model is modest, it has much better discriminative ability compared to current societal clinical practice guidelines. However, the model may need to be further refined for lower risk (nonveteran) populations,” the authors wrote.
The strength of the model is that it relies on data found in the EHR. The researchers suggest that future studies should look employing the model alongside e-Trigger tools that can mine electronic clinical data to identify patients at risk for a missed diagnosis.
The authors reported no personal or financial conflicts of interest.
