Using Data Mining to Help Detect Dysplasia: Extended Abstract

In this paper we explore how data mining can be applied to gastroenterology, and specifically to aid in the diagnosis of patients with high-risk lesions within Barrett´s oesophagus (BE). BE is the only identifiable premalignant lesion for oesophageal adenocarcinoma (OA), a tumor whose incidence has been rising rapidly in the Western World.This paper makes two key contributions. First, as patient information is open to interpretation, we demonstrate that composite rules learned from multiple experts can be more accurate than that of one expert alone. Even expert doctors interpret endoscopy scans differently, potentially making it important to aggregate multiple opinions. Second, we demonstrate that decision trees can generate simple rules for dysplasia diagnosis. These rules can either be used to encapsulate the rules of the most accurate expert for training purposes or to help identify diagnostic errors.