Search

Summary Background Screening for Barrett's oesophagus relies on endoscopy, which is invasive and few who undergo the procedure are found to have the condition. We aimed to use machine learning techniques to develop and externally validate a simple risk prediction panel to screen individuals for Barrett's oesophagus. Methods In this prospective study, machine learning risk prediction in Barrett's oesophagus (MARK-BE), we used data from two case-control studies, BEST2 and BOOST, to compile training and validation datasets. From the BEST2 study, we analysed questionnaires from 1299 patients, of whom 880 (67·7%) had Barrett's oesophagus, including 40 with invasive oesophageal adenocarcinoma, and 419 (32·3%) were controls. We randomly split (6:4) the cohort using a computer algorithm into a training dataset of 776 patients and a testing dataset of 523 patients. We compiled an external validation cohort from the BOOST study, which included 398 patients, comprising 198 patients with Barrett's oesophagus (23 with oesophageal adenocarcinoma) and 200 controls. We identified independently important diagnostic features of Barrett's oesophagus using the machine learning techniques information gain and correlation-based feature selection. We assessed multiple classification tools to create a multivariable risk prediction model. Internal validation of the model using the BEST2 testing dataset was followed by external validation using the BOOST external validation dataset. From these data we created a prediction panel to identify at-risk individuals. Findings The BEST2 study included 40 diagnostic features. Of these, 19 added information gain but after correlation-based feature selection only eight showed independent diagnostic value including age, sex, cigarette smoking, waist circumference, frequency of stomach pain, duration of heartburn and acidic taste, and taking antireflux medication, of which all were associated with increased risk of Barrett's oesophagus, except frequency of stomach pain, with was inversely associated in a case-control population. Logistic regression offered the highest prediction quality with an area under the receiver-operator curve (AUC) of 0·87 (95% CI 0·84–0·90; sensitivity set at 90%; specificity of 68%). In the testing dataset, AUC was 0·86 (0·83–0·89; sensitivity set at 90%; specificity of 65%). In the external validation dataset, the AUC was 0·81 (0·74–0·84; sensitivity set at 90%; specificity of 58%). Interpretation Our diagnostic model offers valid predictions of diagnosis of Barrett's oesophagus in patients with symptomatic gastro-oesophageal reflux disease, assisting in identifying who should go forward to invasive confirmatory testing. Our predictive panel suggests that overweight men who have been taking antireflux medication for a long time might merit particular consideration for further testing. Our risk prediction panel is quick and simple to administer but will need further calibration and validation in a prospective study in primary care. Funding Charles Wolfson Charitable Trust and Guts UK.

Introduction Barrett’s oesophagus (BO) is the only known precursor of esophageal adenocarcinoma. The current gold-standard test for diagnosing BO is endoscopy which is expensive and impractical as a population screening tool. We aimed to develop a robust questionnaire that could be used in routine clinical practice to identify patients with different likelihoods of having BO. Methods We retrospectively analyzed data from two independent prospective datasets: BEST2 and BOOST case-control studies with 1299 and 398 patients, respectively. The BEST2 dataset was split into testing (n=776) and independent validation (n=523) cohorts. Using machine learning techniques (information gain and correlation-based feature selection followed by logistic regression); we identified panels of markers which predicted the presence of BO and then cross validated them between the datasets in order to create a robust set of markers that could be used for future studies. Results Panels comprising the same thirteen demographic and patient symptom features predicted the occurrence of BO in both datasets. These included three demographic features (age, gender and ethnicity), three general patient characteristics (weight, waist circumference, quantity of cigarettes smoked), and six symptoms (years since heartburn and acid taste developed, frequency of heartburn, acid taste symptoms, chest pain and taking stomach medicines). Although there were minor variations in logistic weights for each feature between the panels, they yielded accuracies with areas under the curve (AUC) of 0.91 in the BEST2 and 0.84 in the BOOST cohorts. Furthermore, despite differences in population composition, the panels validated across datasets. Validating the BEST2 model with BOOST yielded a ROC curve of 0.83 and validating the BOOST model against BEST2 yielded a ROC curve of 0.91. The final panel was also validated with the external BEST2 validation cohort (n=523) with AUC = 0.91 (table 1). Conclusions This study identifies thirteen markers as a potential method for non-invasive screening for BO. These markers are consistent across different patient study groups. The panel needs to be validated in a prospective study.