Machine Learning Could Aid Diagnosis of Barrett’s Esophagus, Avoid Invasive Testing – Medical Bag

A risk prediction model consisting of 8 independent diagnostic variables, including age, sex, waist circumference, stomach pain frequency, cigarette smoking, duration of heartburn and acidic taste, and current history of antireflux medication use, can provide potential insight into a patients risk for Barretts esophagus before endoscopy, according to a study in published Lancet Digital Health.

The study assessed data from 2 prior case-control studies: BEST2 (ISRCTN Registry identifier: 12730505) and BOOST (ISRCTN Registry identifier: 58235785). Questionnaire data were assessed from the BEST2 study, which included responses from 1299 patients, of whom 67.7% (n=880) had Barretts esophagus, which was defined as endoscopically visible columnar-lined oesophagus (Prague classification C1 or M3), with histopathological evidence of intestinal metaplasia on at least one biopsy sample. An algorithm was used to randomly divide (6:4) the cohort into a training data set (n=776) and a testing data set (n=523). A total of 398 patients from the BOOST study, including 198 with Barretts esophagus, were included in this analysis as an external validation cohort. Another 200 control individuals were also included from the BOOST study.

Researchers used a univariate approach called information gain, as well as a correlation-based feature selection. These 2 machine learning filter techniques were used to identify independent diagnostic features of Barretts esophagus. Multiple classification tools were assessed to create a multivariable risk prediction model. The BEST2 testing data set was used for internal validation of the model, whereas the BOOST external validation data set was used for external validation.

In the BEST2 study, the investigators identified a total of 40 diagnostic features of Barretts esophagus. Although 19 of these features added information gain, only 8 features demonstrated independent diagnostic value after correlation-based feature selection. The 8 diagnostic features associated with an increased risk for Barretts esophagus were age, sex, cigarette smoking, waist circumference, frequency of stomach pain, duration of heartburn and acidic taste, and receiving antireflux medication.

The upper estimate of the predictive value of the model, which included these 8 features, had an area under the curve (AUC) of 0.87 (95% CI, 0.84-0.90; sensitivity set, 90%; specificity, 68%). In addition, the testing data set demonstrated an AUC of 0.86 (95% CI, 0.83-0.89; sensitivity set, 90%; specificity, 65%), and the external validation data set featured an AUC of 0.81 (95% CI, 0.74-0.84; sensitivity set, 90%; specificity, 58%).

The study was limited by the fact that it collected data solely from at-risk patients, which enriched the overall cohorts for patients with Barrets esophagus.

The researchers concluded that the risk prediction panels generated from this study would be easy to implement into medical practice, allowing patients to enter their symptoms into a smartphone app and receive an immediate risk factor analysis. After receiving results, the authors suggest, these data could then be uploaded to a central database (eg, in the cloud) that would be updated after that person sees their medical professional.

Reference

Rosenfeld A, Graham DG, Jevons S, et al; BEST2 study group. Development and validation of a risk prediction model to diagnose Barretts oesophagus (MARK-BE): a case-control machine learning approach [published online December 5, 2019]. Lancet Digit Health. doi:10.1016/S2589-7500(19)30216-X.

Original post:
Machine Learning Could Aid Diagnosis of Barrett's Esophagus, Avoid Invasive Testing - Medical Bag

Related Posts

Comments are closed.